Score changes

21 October 2026 5 moves

1. letme

   36.4 F →to 71.8 BB +35.4 Terms Product New evidence Checklist

   • Agent ergonomics 54 to 86 · New evidence. The MCP tool definitions, their read-only annotations and the 4xx error bodies were read live on 20 October 2026. On 2 October they could not be read, and missing items were scored strictly. source
   • Maintenance & community 35 to 61 · Product. The changelog has dated letme entries on 5, 14 and 15 October 2026, and the live build is 2026.10.19-2121. Responsiveness and package health were read the normal way rather than strictly. source
   • Payments & pricing 50 to 60 · Checklist. letme moved from the stricter rule of 2 October to the normal checklist on 20 October 2026. Under the normal reading, a published free price counts as a public price. source
   • Reliability 10 to 80 · Product. Since 2 October 2026 letme has a section on the Anchor Terminal status page, a published limit of 20 requests a second per IP and documented 429 handling. Picking is now read as generally available. source
   • Schema & documentation 68 to 87 · New evidence. The MCP tools' typed schemas were read live on 20 October 2026, 429 is now in the OpenAPI file, and there are more dated changelog entries. The normal readings replaced the strict ones of 2 October. source
   • Security & auth 23 to 58 · Checklist. Under the normal checklist, needing no credential counts as it does for other keyless services. letme also added signed answers on 14 October 2026, a security.txt on letme.dev and safe harbour in the terms. source
   • Transparency & trust 37 to 77 · Terms. letme now has terms dated 14 October 2026, a privacy policy of 17 October and a trust page of 15 October naming sub-processors and UK hosting. A status page and security.txt also count. source
2. Kolibri-1

   75.8 BB →to 61 C −14.8 Checklist

   • Agent ergonomics 85 to unscored · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Agent interface scored 90 · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Maintenance & community 75 to unscored · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Payments & pricing 60 to unscored · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Reach scored 60 · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Reliability 82 to unscored · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Risk scored 65 · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Schema & documentation 95 to unscored · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Security & auth 55 to unscored · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Transparency & trust 74 to unscored · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Documentation and upkeep scored 70 · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
3. Qwen3.8-27B

   63.7 B →to 68.6 B +4.9 Checklist

   • Capability scored 70 · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Agent ergonomics 85 to unscored · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Agent interface scored 70 · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Maintenance & community 25 to unscored · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Payments & pricing 60 to unscored · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Reach scored 90 · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Reliability 90 to unscored · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Risk scored 55 · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Schema & documentation 63 to unscored · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Security & auth 45 to unscored · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Transparency & trust 47 to unscored · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Documentation and upkeep scored 36 · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
4. DeepSeek V4.1 Flash

   62.8 B →to 59 C −3.8 Checklist

   • Capability scored 69 · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Agent ergonomics 70 to unscored · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Agent interface scored 75 · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Maintenance & community 70 to unscored · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Payments & pricing 60 to unscored · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Reach scored 50 · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Reliability 72 to unscored · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Risk scored 50 · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Schema & documentation 69 to unscored · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Security & auth 35 to unscored · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Transparency & trust 69 to unscored · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
   • Documentation and upkeep scored 48 · Checklist. Open-weight models moved to a checklist of their own on 21 October 2026, Capability, Reach, Risk, Agent interface and Documentation and upkeep, in place of the nine categories for tools and APIs, and this model was re-read against it.
5. Pinecone API + MCP

   76.9 BB →to 77.1 BB +0.2 Time

   • Transparency & trust 81 to 83 · Time. pinecone.io, registered on 20 October 2016, is now 10 years old, a step of the domain-age check. That moved the provenance score, half of Transparency, from 83 to 87.

20 October 2026 1 move

1. Chatwoot API

   55.8 C →to 56 C +0.2 Time

   • Transparency & trust 75 to 77 · Time. chatwoot.com, registered on 19 October 2016, is now 10 years old, a step of the domain-age check. That moved the provenance score, half of Transparency, from 82 to 86.

19 October 2026 1 move

1. Activepieces API + MCP

   57.5 C →to 57.7 C +0.2 Time

   • Transparency & trust 73 to 75 · Time. activepieces.com, registered on 18 October 2021, is now 5 years old, a step of the domain-age check. That moved the provenance score, half of Transparency, from 76 to 80.

18 October 2026 1 move

1. Codat

   64.3 B →to 64.5 B +0.2 Time

   • Transparency & trust 71 to 73 · Time. codat.io, registered on 17 October 2016, is now 10 years old, a step of the domain-age check. That moved the provenance score, half of Transparency, from 82 to 86.

17 October 2026 92 moves

1. ScribeUp

   7.9 F →to 35.5 F +27.6 New evidence

   • Agent ergonomics 0 to 39 · New evidence. A public partner API reference at docs.scribeup.io was found on 17 October, missed earlier because the Developer docs link on scribeup.io is disabled. It documents paging, error messages on three operations and few required fields. source
   • Maintenance & community 2 to 74 · New evidence. Public iOS, Android and React Native SDKs were found on 17 October, the first two with release notes. iOS SDK 0.10.1 and Android SDK 0.14.1 shipped on 16 October 2026, after Android 0.13.0 on 19 August. source
   • Schema & documentation 0 to 58 · New evidence. docs.scribeup.io serves a public OpenAPI 3.1 file, last modified 13 May 2026, with seven operations, 19 enums and 66 example values. It was missed on 14 October because the Developer docs link on scribeup.io is disabled. source
   • Security & auth 6 to 33 · New evidence. The partner API reference found on 17 October documents a client token with X-Client-ID, 60-minute end-user tokens, SAML and HMAC-SHA256 signatures on requests and webhooks. Before it was found, those checklist items scored nothing. source
   • Transparency & trust 42 to 51 · New evidence. The Android SDK release notes, read on 17 October, mark breaking changes and deprecate OAuthCallbackActivity without a date. They also gave the listing a public changelog. source
2. Overclock

   7.5 F →to 33.1 F +25.6 Terms Product

   • Agent ergonomics 5 to 54 · Product. Overclock now documents its MCP server at mcp.overclock.tech, with eight read-only tools, result limits and truncation flags. Only the 401 error is documented, and the definitions carry no tool annotations. source
   • Reliability 10 to 15 · Product. The MCP server is now documented with its endpoint, transport and auth and no beta label, and the endpoint answers 401 with the documented header. There is still no status page, rate limits or SLA. source
   • Schema & documentation 0 to 62 · Product. Overclock now publishes tools.json with JSON Schema inputs and outputs for all eight tools, descriptions that say when to use each, and an example per tool. There is no llms.txt, version number or changelog. source
   • Security & auth 10 to 46 · Product. The newly documented MCP server uses OAuth 2.1 with PKCE, one read scope, header-only bearer tokens, one-hour access tokens and rotating refresh tokens, with revocation. There is still no injection guidance, security.txt or disclosure policy. source
   • Transparency & trust 34 to 36 · Terms. The privacy policy, now dated 15 October 2026, names OpenAI, Fireworks AI and Anthropic as AI providers and what each receives, plus Cloudflare. There are still no terms of service, retention periods or DPA. source
3. Trigger.dev

   74.7 BB →to 63.7 B −11 New evidence Negative event

   • Negative events 0 to −12 · Negative event. Deductions for six advisories published on 2 October 2026 and one from 13 August, including a cross-tenant SQL injection, cross-tenant run replay and default secrets in self-hosted setups. All were fixed in July 2026, and four Moderate advisories weren't deducted. source
   • Security & auth 73 to 75 · New evidence. The 6 October check confirmed a valid security.txt expiring 1 September 2027, SOC 2 Type II and advisories published with CVE IDs. source
   • Transparency & trust 73 to 81 · New evidence. The sub-processor page, not read on 1 October, names each sub-processor with purpose and country (26 on 6 October, 27 after its 13 October update). A valid security.txt was also recorded. source
4. Modal Sandboxes

   75.5 BB →to 65.9 B −9.6 Pricing New evidence Correction

   • Payments & pricing 40 to 20 · Pricing. On 6 October the pricing page no longer carried a no-card statement, and Modal's billing guide says a payment method must be on file to use Modal, so the $30 monthly credit no longer counts as a no-card free tier. source
   • Reliability 95 to 55 · Correction. Sandboxes run on Modal's infrastructure, so reliability is now scored on the hosted-service checklist from status.modal.com, as for E2B, Daytona and Vercel Sandbox, not on the SDK checklist. No sandbox rate limits or SLA figure are published. source
   • Security & auth 76 to 79 · New evidence. The 6 October re-check read the Sidecars docs (beta), which show a credential proxy in a separate trusted container so a secret stays out of the sandbox. The 1 October check found no such option. source
   • Transparency & trust 72 to 77 · New evidence. Data locations weren't checked on 1 October. The region guide routes inputs through Virginia by default and names US, EU and Asia-Pacific regions for sandboxes. No subprocessor list was readable. source
5. Resy

   10.3 F →to 19.5 F +9.2 New evidence

   • Maintenance & community 25 to 45 · New evidence. Resy's newsroom dates reservations in ChatGPT to 10 August 2026. With Tock's venues on 11 August and the Resy Credit change on 15 September, that makes three dated product changes in 90 days. source
   • Payments & pricing 0 to 10 · New evidence. The venue plans page was read on 17 October. It prices Platform and Essential at $289 a month and Platform 360 and Premium at $459, with no per-cover fees. No API price is published. source
   • Reliability 25 to 50 · New evidence. The status page's data file was read on 17 October, showing 26 components and four scheduled maintenances. The page dates from December 2025 and has never logged an incident, so a clean 90 days counts only in part. source
   • Transparency & trust 30 to 43 · New evidence. The Resy Technology Service Terms for venues, last updated 11 June 2024 and naming Resy Network, Inc., were found readable through the plans page. They incorporate a data protection addendum listing data subjects, categories and purposes. source
6. monday.com

   76.4 BB →to 70.7 BB −5.7 Product Negative event

   • Negative events 0 to −3 · Negative event. On 13 October 2026 monday.com removed two API mutations from every public version, released ones included, with no advance notice. The change is documented with migration notes, so it takes the minimum deduction for a breaking change, as AssemblyAI's same-day change did. source
   • Payments & pricing 50 to 30 · Product. After 8 October 2026 monday.com withdrew its agent signup API. Version 2.0.0 of its signup skill says a person must create the account in a browser and no API provisions one. source
   • Transparency & trust 80 to 78 · Product. On 13 October 2026 monday.com removed the install_app and uninstall_app mutations from every public API version, including released ones, with effect the same day. Its versioning page promises six months of stability. source
7. Twilio SendGrid

   63.6 B →to 68.8 B +5.2 Product New evidence

   • Maintenance & community 47 to 67 · Product. Twilio's changelog announced Unified Login for SendGrid Marketplace customers as GA on 16 October 2026. Read back to 14 July, it shows three SendGrid entries in 90 days, on 3 August, 27 August and 16 October. source
   • Reliability 66 to 76 · New evidence. The status.twilio.com history pages, unreadable on 1 October, were read on 17 October. SendGrid incidents from 19 July were minor or marked no impact, so the history now counts as a clean record of minor incidents. source
   • Schema & documentation 72 to 79 · Product. Twilio's docs-wide llms.txt now has a Twilio SendGrid section, and SendGrid pages answer at .md URLs as Markdown. On 1 October no SendGrid pages were found in it. source
   • Transparency & trust 79 to 82 · New evidence. A dated status notice of 19 August 2026 was found, announcing the removal of legacy SendGrid API and SMTP IP addresses from 24 August, with permanent removal on 31 August. The notice period was short. source
8. pplx-decider

   62.9 B →to 68.1 B +5.2 Product New evidence

   • Agent ergonomics 76 to 83 · Product. Perplexity's Python SDK 0.44.0 and TypeScript SDK 0.39.0 added a typed client.decisions.create method on 15 October 2026. Neither SDK had a Decisions method on 9 October. source
   • Maintenance & community 62 to 65 · Product. Both official SDKs gained Decisions methods on 15 October 2026. That outweighed a deduction for model churn, since the v1.1 and v1 names were moved to v1.2 on the day it shipped with no notice. source
   • Reliability 61 to 66 · Product. The OpenAPI description moved from a 0.1.0 preview file into the main openapi.json at version 1.0.0, and both official SDKs gained a Decisions method on 15 October 2026, so the endpoint no longer reads as a preview. source
   • Schema & documentation 90 to 89 · Product. Since v1.2 shipped on 16 October 2026, the v1.1 and v1 model names select v1.2 and the response echoes the name sent, so a versioned name no longer pins a model. The guide gained advice on choosing a model. source
   • Security & auth 61 to 68 · New evidence. security.txt and the API terms on www.perplexity.ai, unreadable on 9 October, were read on 17 October. They name a disclosure programme, a private Bugcrowd bounty, annual penetration tests and a commitment not to train on customer content. source
   • Transparency & trust 52 to 71 · New evidence. The API terms, DPA and security.txt, unreadable on 9 October, were read on 17 October, adding the legal entity Perplexity AI, Inc. and a no-training commitment. The v1.2 inference code is also now licensed Apache-2.0. source
9. Zep

   69.3 B →to 73.9 BB +4.6 Product

   • Agent ergonomics 70 to 87 · Product. Zep's v4 API, released on 14 October 2026, adds opaque cursors on list methods, errors with a stable code, the field at fault and a request ID, and Idempotency-Key support. source
   • Reliability 65 to 70 · Product. Since v4 on 14 October 2026, any state-changing call takes an Idempotency-Key, and the SDKs send one when they retry a create call. A 16 October error incident was read as minor. source
   • Schema & documentation 89 to 93 · Product. The v4 OpenAPI lists 400, 401, 403 and 404 on almost every operation with one error object, and the Context MCP Server page says when to use its navigation tools. source
   • Transparency & trust 57 to 59 · Product. With v4 on 14 October 2026, Zep dated the v3 API's deprecation for 13 April 2027 and published a method-by-method migration guide. There is still no standing deprecation policy. source
10. Tock

    12.4 F →to 16.6 F +4.2 New evidence

    • Payments & pricing 0 to 10 · New evidence. Resy's plans page, read on 17 October, prices the Tock-powered plans, Essential at $289 a month with 3% on prepayments and Premium at $459 with 2%. source
    • Reliability 25 to 40 · New evidence. The status page's data file was read on 17 October, showing 12 components and incidents back to February 2026. It records a major outage of SMS and the dashboard for 1 hour 58 minutes on 7 October. source
11. Google Drive API + MCP

    79.6 A →to 75.5 BB −4.1 New evidence Correction

    • Agent ergonomics 85 to 87 · New evidence. On 17 October the MCP server's tools/list answered without a token. It showed read-only, destructive and idempotent annotations on all eight tools, which the reference page doesn't show, and about 27 KB of definitions. source
    • Reliability 90 to 70 · Correction. The 4 October check read the Workspace feed as having no incidents after 30 June. The feed has a 12 hour 42 minute Drive disruption in Australia, the Philippines and New Zealand on 18 to 19 August 2026, counted as one regional major. source
    • Schema & documentation 83 to 89 · Correction. Markdown at a .md.txt suffix now scores as it did for Sheets and Forms. The MCP tool descriptions read on 17 October say which tool to use instead and not to guess file IDs, read as Gmail's were.
    • Security & auth 91 to 83 · Correction. Drive now loses 10 points on credentials because Google's OAuth guide documents a query-string token and the Drive API reads one, as for Gmail and Sheets. Two points come back because the MCP setup page, updated 12 October 2026, now requires prompt screening. source
12. Browserbase

    76.2 BB →to 72.2 BB −4 Negative event

    • Negative events 0 to −4 · Negative event. On 24 November 2025 malicious versions of five Browserbase npm packages were published in the Sha1-Hulud worm, and we found no Browserbase advisory. The 1 and 6 October checks missed it, so the deduction is added now. source
13. ZenRows

    74.8 BB →to 78.8 A +4 Product New evidence

    • Maintenance & community 87 to 90 · Product. ZenRows moved its changelog from Intercom, last updated 14 July, to a dated changelog on zenrows.com with entries into October 2026. MCP v2.3.0 shipped on 2 October. source
    • Schema & documentation 83 to 93 · Product. An OpenAPI 3.1 file now covers Fetch, Extract, Batch, agent signup and key rotation, with error responses on every operation. The dated changelog and path versioning also count. source
    • Security & auth 41 to 50 · New evidence. The 6 October check read per-key credit caps for up to 10 keys an account, and an Activity Log of requests kept for up to 60 days, added on 31 July 2026. Fetch still takes the key only in the query string. source
    • Transparency & trust 72 to 78 · New evidence. A versioning policy promising at least twelve months' notice with Deprecation and Sunset headers was read on 6 October. A DPA in effect since 1 June 2022 was read on 17 October. source
14. Vercel Sandbox

    69.6 B →to 73.5 BB +3.9 New evidence Correction Negative event

    • Agent ergonomics 65 to 73 · New evidence. Vercel's public OpenAPI, missed on 1 October, documents cursor pagination with limit, sort, name, status and tag filters on sandbox lists, and pagination on sessions, snapshots, drives and commands. source
    • Maintenance & community 80 to 82 · New evidence. A workflow that runs the tests on every push to main and every pull request was found on 17 October. The 1 October check had not looked at CI, and passing status is still unconfirmed. source
    • Negative events 0 to −3 · Negative event. Three points for Vercel's April 2026 security incident, in which an attacker decrypted some customers' non-sensitive environment variables. It reached the accounts Sandbox runs under and was closed on 16 October 2026, matching Vercel Connect. source
    • Payments & pricing 40 to 50 · New evidence. The Hobby plan page, read on 17 October, calls the plan free and asks for card details only when upgrading to Pro. The 1 October check had not confirmed a no-card signup. source
    • Reliability 70 to 75 · New evidence. The OpenAPI documents 429 as the concurrency limit, and the SDK changelog says the client honours Retry-After up to 20 seconds. Neither was found on 1 October. source
    • Schema & documentation 77 to 88 · New evidence. Vercel's public OpenAPI 3.0.3 document has 27 Sandbox paths and 33 operations, covering sandboxes, sessions, commands, files, snapshots, drives and network policy. The 1 October check did not find them. source
    • Security & auth 80 to 86 · Correction. The security programme was read the same way as Vercel Connect, with SOC 2 Type 2, ISO 27001 and a public incident bulletin. The API's per-session command list was also counted as an audit trail. source
    • Transparency & trust 75 to 78 · New evidence. The DPA, effective 31 March 2026, was read on 17 October. It uses the 2021 SCCs and UK addendum, states primary processing in the United States and points to a subprocessor list. source
15. Backblaze B2

    75.3 BB →to 79 A +3.7 New evidence

    • Reliability 60 to 75 · New evidence. The status page's RSS feed, unread on 1 October, was read on 6 October. It shows scheduled maintenance and two minor EU Central failure-rate entries around 1 October, replacing the lower award for unreadable history. source
    • Transparency & trust 73 to 81 · New evidence. The 6 October re-check read the public EEA DPA and the sub-processor list, which names each region's data centre and location. The privacy notice still gives no retention periods. source
16. Composio (API + MCP)

    75.1 BB →to 78.2 A +3.1 Terms Product New evidence

    • Payments & pricing 40 to 60 · New evidence. The 6 October re-check read the docs for composio login --agent, which creates an agent account and project key from the CLI with no browser or email sign-up. Connecting a person's apps still needs their consent. source
    • Security & auth 70 to 73 · Product. Since Python SDK 0.28.0 and TypeScript SDK 0.25.0 (16 October 2026) a session can require user approval before chosen tools run. Approvals are experimental and off by default. source
    • Transparency & trust 76 to 77 · Terms. The DPA updated on 17 October 2026 agrees with the retention page on deletion within 30 days and on payloads never being used for training, and states the use of de-identified metadata. source
17. Twilio API + MCP

    80.4 A →to 77.4 BB −3 Reliability

    • Reliability 90 to 75 · Reliability. Two major outages counted. Verify codes and SMS delivery failed across several regions for 6 hours 27 minutes on 7 and 8 October 2026, and the REST API returned 502 errors for 76 minutes on 1 September. source
18. Twilio Programmable Voice API + MCP

    80.4 A →to 77.4 BB −3 New evidence

    • Reliability 90 to 75 · New evidence. The status history, unreadable on 4 October, was read on 17 October. It shows two major outages, ConversationRelay down for Deepgram speech models for 2 hours 43 minutes on 3 August 2026 and REST API 502 errors for 76 minutes on 1 September. source
19. Nutrient DWS Processor API

    68.7 B →to 71.5 BB +2.8 Product Negative event Time

    • Agent ergonomics 84 to 79 · Product. On 14 October 2026 Nutrient replaced its local MCP server with a hosted one. Smithery's copy of the hosted tool list shows 14 tools and about 60,000 to 68,000 characters of definitions, a heavier load on an agent's context than the eight local tools. source
    • Maintenance & community 78 to 83 · Product. The hosted MCP server was published to the official registry as io.nutrient/mcp 2.0.0 on 14 October 2026. The local server's repository is now reference-only and the hosted service has no public changelog, so responsiveness scores lower. source
    • Negative events −2 to −1 · Negative event. The deduction was halved. By 17 October 2026 the product page FAQ gave the same 100 requests a minute limit as the pricing guide, and the MCP page no longer offered test keys. Both were corrected within about nine days. source
    • Reliability 68 to 78 · Time. The 61-minute DWS Platform outage of 13 July 2026 has left the 90-day window. Since then the longest Processor API downtime was about 12 minutes, on 23 July, and nothing after 7 September, so minor incidents only. source
    • Schema & documentation 84 to 82 · Product. The only changelog was the local MCP server's, which stopped when that server lost support on 14 October 2026. The hosted 2.0.0 server has none, and no public API changelog was found. source
    • Security & auth 66 to 68 · Product. The supported MCP route is now the hosted server, which reads no local files, so the local server's unrestricted file access without SANDBOX_PATH no longer applies. There is still no confirmation step. source
    • Transparency & trust 73 to 75 · Product. Nutrient dated the local MCP server's end of support (14 October 2026) and named the hosted server as its replacement, though it took effect the day it was announced and there is still no deprecation policy. source
20. Speechify API Voice Cloning

    74.1 BB →to 76.8 BB +2.7 Terms New evidence Negative event

    • Agent ergonomics 82 to 97 · New evidence. The OpenAPI file, read on 6 October, shows GET /v1/voices takes type=personal to list only the workspace's clones, with cursor pagination and a limit up to 200. The 1 October check had not confirmed this. source
    • Negative events 0 to −3 · Negative event. Three points for a breaking change without notice. Since 7 October 2026 voice creation returns 403 voice_cloning_unavailable_in_region for unnamed jurisdictions, announced the same day with no API version change. Existing voices keep working. source
    • Security & auth 73 to 87 · New evidence. The 6 October check read a 90-day project audit endpoint and a trust page claiming SOC 2 Type II. A biometric data policy effective 6 October 2026 also sets destruction periods for voice samples, models and verification data. source
    • Transparency & trust 65 to 74 · Terms. A biometric data policy effective 6 October 2026 sets destruction periods for voice data and consent records. A regions guide of 17 October states that account data stays in the United States and synthesis runs in US East or the Netherlands. source
21. Firecrawl MCP

    75.3 BB →to 77.9 BB +2.6 New evidence

    • Reliability 70 to 78 · New evidence. The 6 October re-check found Firecrawl's Errors reference, linked from the rate-limit page. It says 429 responses carry Retry-After when available and gives backoff advice per error. Two points are held back for no idempotency keys on job-starting calls. source
    • Schema & documentation 88 to 91 · New evidence. The same Errors reference, which the 1 October check didn't find, lists each HTTP status and error string with cause, remedy and whether to retry. The examples and errors item now gets full marks. source
    • Security & auth 67 to 70 · New evidence. The 6 October re-check read the API's opt-in checkPromptInjection classifier (4 extra credits, a 403 when it blocks). It earns part of the prompt-injection item, since the MCP tool schemas don't expose it to an agent. source
22. Robot Framework

    76.4 BB →to 73.8 BB −2.6 Time

    • Maintenance & community 90 to 60 · Time. No release since 7.5 on 14 September 2026, which is now more than 30 days old. 7.5b1 of 17 July has left the 90-day window, leaving two releases in it, short of three. source
23. Amazon Location Service

    74.1 BB →to 76.5 BB +2.4 Security New evidence Correction

    • Payments & pricing 20 to 30 · Correction. The 8 October check kept an older reading that a card is required. The Free Tier FAQ says most new customers need none, and the $200 credits cover Amazon Location, so the free-tier item earns half, as on S3 and Polly. source
    • Security & auth 73 to 76 · Security. Amazon Location is listed in AWS's SOC 1, 2 and 3 scope, and the security.txt on aws.amazon.com, expired since 24 September 2026, was renewed to 16 October 2027. 2 points stay held back for the lack of a paid bug bounty. source
    • Transparency & trust 70 to 77 · New evidence. AWS's sub-processor page, unread on 8 October, names Esri, Grabtaxi Holdings and HERE for Amazon Location with their processing locations and promises 30 days' notice. The renewed security.txt also raised the provenance score, half of Transparency. source
24. Kilo Code CLI

    75.4 BB →to 73 BB −2.4 New evidence Negative event

    • Negative events 0 to −2 · Negative event. Version 7.8.10 (14 October 2026) added the conda client ID to default-on telemetry, sent with every PostHog event, without mention in the changelog or docs. It respects the opt-out, so it takes the minimum deduction for undisclosed telemetry. source
    • Security & auth 66 to 64 · New evidence. The 8 October check looked only at repository advisories. NVD lists CVE-2026-8766 (low, published 17 May 2026) and says Kilo didn't respond to the report, which counts against the security programme line. source
25. Browser Use

    77.9 BB →to 80.2 A +2.3 Product Correction

    • Maintenance & community 77 to 82 · Product. The cloud SDK was tagged 3.12.0 and 3.12.1 on 16 October 2026 and the library 0.13.12 and 0.13.13 on 15 October. The lower award rested on the SDK's last release being 4 September. source
    • Schema & documentation 85 to 86 · Correction. The 8 October note said most v4 operations documented only a 422 error. 36 of 50 operations document other errors, and the vendor's own copy of the spec from 8 October doesn't support the earlier reading. source
    • Security & auth 59 to 69 · Product. The troubleshooting guide now documents restricted API keys with read or write scopes per V4 resource and a 403 naming the missing scope. No rotation or revocation guide was found. source
26. Azure MCP Server

    67.7 B →to 69.9 B +2.2 Product

    • Agent ergonomics 72 to 74 · Product. Azure MCP Server 3.0.0 (12 October 2026) returns 404 for missing resources and 400 for invalid input where it returned 500, rejects unknown parameters and gives clearer authentication errors. No reference documents the error shapes. source
    • Reliability 69 to 77 · Product. On 1 October npm latest pointed at a 3.0.0 beta, so @latest installed a prerelease. 3.0.0 shipped as the stable release on 12 October 2026 and latest now points at it. source
    • Transparency & trust 76 to 79 · Product. Since 8 October 2026 SUPPORT.md puts Azure MCP under the Azure SDK lifecycle policy, at least 12 months per major version. Tools were still renamed between prereleases without notice and no 2.x end date is published. source
27. Mapbox APIs + MCP

    75 BB →to 77.2 BB +2.2 Product New evidence

    • Maintenance & community 80 to 90 · Product. Mapbox shipped MCP server v0.15.0 and v0.15.1 on 8 and 9 October 2026 and v0.16.0 on 15 October, on npm and in the official MCP registry, after no release since 30 July. source
    • Payments & pricing 30 to 40 · New evidence. Earlier checks read only the pricing page and billing guide. Mapbox's sign-up guide has a 'Start without a credit card' trial with monthly limits, so the free-tier item now counts a start with no card. source
28. SOUNDRAW API

    12 F →to 14.2 F +2.2 New evidence

    • Agent ergonomics 0 to 7 · New evidence. A SOUNDRAW guide of 7 August 2026 says length in seconds is the only required field. It is a blog post rather than the API reference, which is still sent only after sign-up. source
    • Schema & documentation 5 to 8 · New evidence. The same 7 August 2026 guide shows one compose request body and one result body for soundraw.io/api/v3, with no error responses. These are the first API examples read for this listing. source
    • Transparency & trust 36 to 43 · New evidence. The privacy policy and GDPR policy were read for the first time on 17 October. They list what is collected and users' rights, without fixed periods. The guide also confirmed the API endpoint sits on soundraw.io. source
29. Amazon Polly

    75.6 BB →to 73.5 BB −2.1 Pricing Security New evidence Correction

    • Payments & pricing 20 to 30 · Pricing. AWS's Free Tier FAQ says that since a new sign-up in October 2026 most new customers need no payment method, and newer accounts get up to $200 in credits usable on Polly. The roll-out is limited, so the free-tier item earns half. source
    • Reliability 100 to 80 · New evidence. The Health Dashboard history file, first read in October, lists Polly in the me-south-1 (Bahrain) disruption open since 2 March 2026. On 15 September AWS said data held only in that Region can't be restored. Counted as one major event. source
    • Security & auth 80 to 83 · Correction. Polly is in AWS's SOC 1, 2 and 3 scope, and on 6 October the security programme item was brought in line with S3 and Secrets Manager. The security.txt was renewed by 17 October, and 2 points stay held back for no paid bounty. source
    • Transparency & trust 77 to 79 · Security. The renewed security.txt on aws.amazon.com, now valid to 16 October 2027, raised the provenance score, half of Transparency. The editorial assessment did not change. source
30. Langfuse API + MCP

    72.7 BB →to 74.8 BB +2.1 Product

    • Security & auth 65 to 77 · Product. Since 16 October 2026 new project API keys take one of four permission sets, including read-only Viewer and write-only Ingestion. Keys can expire and show when they were last used. The MCP docs still leave read-only access to a client allowlist. source
31. Apideck Accounting API + MCP

    72.9 BB →to 70.9 BB −2 Product Negative event

    • Agent ergonomics 88 to 90 · Product. Since spec v10.64.0 on 7 October 2026 the REST API accepts an Idempotency-Key header on five write operations for six connectors. The MCP tools in 0.25.1 don't send it, so partial credit. source
    • Negative events 0 to −3 · Negative event. Spec v10.69.4 on 15 October 2026 changed NetSuite Bill Payments updates to replace allocations rather than merge them, announced the day it shipped with no prior BREAKING.md entry. Documented and fixing a real bug, so the lowest deduction. source
    • Reliability 75 to 78 · Product. From 7 October 2026 an optional Idempotency-Key makes retried journal entry, invoice, credit note and bill writes apply at most once on six connectors. Other connectors and operations refuse it, so partial credit. source
32. Pydantic AI

    83.7 A →to 81.7 A −2 Negative event

    • Negative events −4 to −6 · Negative event. Two points added for High advisory GHSA-6fqq-452j-qhrp, a streaming flaw that let a client exhaust a shared concurrency limiter. It was fixed in 2.53.0 on 2 October 2026 and published on 8 October, after the 4 October check. source
33. Qdrant API + MCP

    75.3 BB →to 73.3 BB −2 Reliability

    • Reliability 80 to 70 · Reliability. On 8 October 2026 some Qdrant Cloud clusters in Azure Germany West Central were unavailable and others ran with fewer replicas. The status page shows the region down for 1 hour 54 minutes, one major outage of the core service. source
34. Resend API + MCP

    75.1 BB →to 77.1 BB +2 Product New evidence

    • Payments & pricing 40 to 50 · New evidence. The 6 October check found that since 22 September 2026 stripe projects add resend/email creates an account and API key through the operator's Stripe login. The 1 October check had missed it. A person still needs a Stripe account. source
    • Security & auth 65 to 69 · Product. resend-mcp 2.25.0 on 6 October 2026 turned off local file reads for attachments and contact imports over HTTP unless allowed folders are set, and the README tells stdio users to set them against prompt injection. source
35. Shopify API + MCP

    75 BB →to 77 BB +2 New evidence

    • Reliability 73 to 83 · New evidence. The incident history, unread on 1 October, was read on 6 and 17 October. The newest posted incident is a Shop Pay outage on 24 June 2025, so the 90 days are clean, though the page is rarely written to. source
36. Amazon Transcribe

    73.4 BB →to 75.3 BB +1.9 Security New evidence Correction

    • Maintenance & community 35 to 80 · New evidence. The 1 October check read only the document history. The API models show changes on 31 July, 3 September and 18 September 2026, including customer-managed KMS keys, and the SDKs were current on 14 October. source
    • Payments & pricing 20 to 30 · Correction. The 1 October note said a new AWS account needs a card. The Free Tier FAQ says most new customers need no payment method and new accounts get up to $200 in credits, so the free-tier item earns half, as on Polly and S3. source
    • Reliability 95 to 75 · New evidence. The 1 October check read only three regional feeds. AWS's Health Dashboard history lists Transcribe in the me-south-1 (Bahrain) disruption open since 2 March 2026, where AWS said on 15 September some data can't be restored. Counted as one major event. source
    • Security & auth 80 to 83 · Security. Transcribe is listed in AWS's SOC 1, 2 and 3 scope, and the security.txt on aws.amazon.com, expired since 24 September 2026, was renewed to 16 October 2027. 2 points stay held back for the lack of a paid bug bounty. source
    • Transparency & trust 82 to 84 · Security. The renewed security.txt on aws.amazon.com, now valid to 16 October 2027, raised the provenance score, half of Transparency. The editorial assessment did not change. source
37. Amazon Bedrock AgentCore Code Interpreter

    73.1 BB →to 71.3 BB −1.8 Security Negative event

    • Negative events 0 to −2 · Negative event. Deductions were added for two high-severity command-injection advisories in the Python SDK's install_packages(), GHSA-6rfw-mq36-jm8h (19 June 2026, fixed in 1.6.1) and CVE-2026-16796 (23 July 2026, fixed in 1.18.1). Both predate the 8 October check. source
    • Transparency & trust 70 to 72 · Security. AWS renewed the security.txt on aws.amazon.com, now valid to 16 October 2027 after expiring on 24 September 2026. That raises the computed provenance half. source
38. Angi (including HomeAdvisor, now Angi Leads)

    12.7 F →to 14.5 F +1.8 Correction

    • Maintenance & community 34 to 54 · Correction. The 90-day release count included only app releases. It now counts dated product changes as well, as other listings do, which adds the Gemini connected app (12 August) and the Sponsored Agent pilot in ChatGPT (16 September). source
39. Context7

    73 BB →to 71.3 BB −1.7 Product Negative event

    • Agent ergonomics 69 to 71 · Product. Since 4.2.1 on 9 October 2026 query-docs sets isError when a request fails, so clients no longer read error text as documentation. resolve-library-id still returns errors as plain text. source
    • Negative events 0 to −2 · Negative event. Minor release 4.2.0 on 7 October 2026 made self-hosted HTTP servers require a credential by default, so anonymous calls began failing with 401. The changelog names an opt-out, and stdio and the hosted endpoint were unaffected. source
40. Gemini Enterprise

    68.1 B →to 66.4 B −1.7 Correction

    • Security & auth 87 to 77 · Correction. The credential line now loses 10 points, as for the Google Workspace APIs on 17 October. Google's OAuth guide documents sending the token as an access_token query parameter, and the Discovery Engine API read a token sent that way in our test. source
41. Chrome DevTools MCP

    77 BB →to 75.4 BB −1.6 Product

    • Agent ergonomics 83 to 73 · Product. Release 1.11.0 on 15 October 2026 turned the two screencast tools on by default, taking the default tool list from 30 to 32, above the checklist's limit of 30 tools. source
42. Circle Wallets (Agent Wallets, Programmable Wallets)

    73.9 BB →to 72.3 BB −1.6 Reliability

    • Reliability 68 to 60 · Reliability. On 8 October 2026 Modular Wallets were paused on every production network for about five hours, and the paymaster stayed paused until 10 October. Counted as a major outage beside the September webhook incident. source
43. OpenAI Decisions API

    71.5 BB →to 73.1 BB +1.6 Product Time

    • Reliability 53 to 60 · Time. The Decisions endpoint now has 11 days of public record since its 6 October 2026 release, with no incident naming it, where the 8 October check had two days. source
    • Schema & documentation 89 to 87 · Product. OpenAI's OpenAPI document began accepting image URLs on 8 October 2026 and system, assistant and function-call inputs on 15 October, but the guide still says images must be base64 and input is user messages only. The two disagree. source
    • Security & auth 83 to 86 · Product. On 9 October 2026 the OpenAPI document added a Decisions usage endpoint, grouped by project, user, API key or model in minute, hour or day buckets. That gives the per-key usage view the OpenAI API listing is scored on. source
44. Amazon SES

    75.1 BB →to 73.6 BB −1.5 New evidence Correction

    • Payments & pricing 20 to 30 · Correction. We had assumed AWS sign-up takes a card. The Free Tier FAQ says most new customers need no payment method, so the free-tier item earns half, as on the S3, Secrets Manager and Polly listings. source
    • Reliability 95 to 75 · New evidence. The earlier check read only the us-east-1 feed. AWS's Health Dashboard history file lists SES in the me-central-1 and me-south-1 disruptions open since March 2026, where AWS said on 15 September some data can't be restored. Counted as one major event. source
    • Security & auth 87 to 89 · Correction. The security programme item was brought in line with the S3, Secrets Manager and Polly listings on the same evidence (SOC scope, a HackerOne disclosure programme, bulletins). The security.txt has since been renewed, and 2 points stay held back for no paid bounty. source
    • Transparency & trust 77 to 87 · New evidence. AWS's sub-processor page, not read before, ties sub-processors to the Data Processing Addendum, names the SES entities and promises 30 days' notice. The renewed security.txt also raised the provenance score, half of Transparency. source
45. Amazon EventBridge

    73.7 BB →to 75.1 BB +1.4 Security New evidence

    • Reliability 73 to 78 · New evidence. The Health Dashboard history file, unread on 9 October, was read on 17 October. It lists EventBridge in the me-central-1 and me-south-1 disruptions open since March 2026, one major event, which replaces the lower award for no readable history. source
    • Security & auth 88 to 89 · Security. The security.txt on aws.amazon.com was renewed and is valid to 16 October 2027. The disclosure policy, read for the first time, names no paid bounty, so 2 points stay held back. source
    • Transparency & trust 87 to 89 · Security. The renewed security.txt on aws.amazon.com, now valid to 16 October 2027, raised the provenance score, half of Transparency. The editorial assessment did not change. source
46. ElevenLabs Voice Cloning and Voice Design API

    73.6 BB →to 74.8 BB +1.2 New evidence

    • Agent ergonomics 85 to 88 · New evidence. The 6 October re-check read the API errors reference. Errors carry a type, a named code and a param for validation failures, with recovery hints on many codes. source
    • Schema & documentation 87 to 91 · New evidence. The earlier note placed error codes only in help-centre articles. The API docs have an errors reference listing 10 error types and about 70 named codes, though endpoint pages still list only 200 and 422. source
47. Spider

    74.1 BB →to 75.3 BB +1.2 Terms New evidence

    • Reliability 80 to 85 · New evidence. The incident history, closed to our reader on 1 October, was read on 6 and 17 October. It lists no incident in 90 days, but its only entry ever is a 2024 template example, so not full marks. source
    • Transparency & trust 66 to 69 · Terms. The privacy policy was rewritten in October 2026. It now gives retention periods for logs, cached pages and recordings and BAGELMEN LLC's address, though the DPA is Enterprise-only and providers are named by role only. source
48. Agent Development Kit (ADK)

    74.7 BB →to 73.6 BB −1.1 New evidence Negative event

    • Agent ergonomics 70 to 80 · New evidence. The 6 October re-check found the context compression page, which documents token-based and sliding-window compaction through EventsCompactionConfig. With RunConfig's cap on model calls, output-size control now scores as it does for the OpenAI Agents SDK. source
    • Maintenance & community 87 to 90 · New evidence. The 1 October check didn't look at the Go, Java and Kotlin packages. On 6 October the Go module, the Java artefact and the npm package were all current, so package currency now gets full marks.
    • Negative events −4 to −7 · Negative event. Google published two more ADK CVEs in September 2026, which the 1 October check missed. CVE-2026-79696 (critical, adk web code injection, fixed in 2.7.0) costs 2 points and CVE-2026-79707 (file read, fixed in 1.22.0) costs 1. source
49. GroqCloud

    75.6 BB →to 74.5 BB −1.1 Reliability New evidence

    • Reliability 100 to 90 · Reliability. Groq posted three minor incidents after 6 October 2026. The enterprise-only preview MiniMax M2.7 had about six hours of failed requests on 10 October and 18 minutes of partial outage on 12 October, and the console had 21 minutes of errors on 12 October. source
    • Security & auth 77 to 82 · New evidence. The 1 October check didn't find key rotation documented. On 6 October the security onboarding page was read, which advises rotating keys periodically and revoking them at once if compromise is suspected. source
50. Temporal

    77.2 BB →to 78.3 A +1.1 New evidence

    • Reliability 85 to 80 · New evidence. The full status history, partly unread on 1 October, was read on 6 October. It shows a configuration change on 10 September 2026 that blocked workflow progress in us-central1 and asia-southeast2 for over an hour, one major outage on the core path. source
    • Security & auth 86 to 88 · New evidence. A valid security.txt, expiring 15 January 2027, was confirmed on 6 October with a responsible disclosure policy and safe harbour. Temporal also issues CVEs as a CVE Numbering Authority. source
    • Transparency & trust 70 to 90 · New evidence. The published DPA, effective 10 October 2024, and the trust centre's list of seven subprocessors with purpose and location were read on 6 October. The terms and a valid security.txt were also recorded. source
51. Bird API + MCP

    76.5 BB →to 77.5 BB +1 Product New evidence Correction

    • Reliability 63 to 70 · Product. Bird's SMS, WhatsApp and email pricing pages now publish rate limits per plan, such as 60 SMS sends a minute on pay as you go and up to 1,500 on Scale. source
    • Security & auth 85 to 80 · New evidence. On 17 October the trust centre listed no certification covering Bird Platform, with ISO 27001 scoped to other products and no SOC 2 report. Message-safety rules added to the bot templates on 16 October recovered a little. source
    • Transparency & trust 80 to 85 · Correction. The 1 October check said no retention periods were published. The privacy statement of 31 August 2026 already gave them by product, such as six months for SMS, voice and email traffic data. source
52. DeepSeek API

    46.8 D →to 45.8 E −1 Reliability

    • Reliability 65 to 60 · Reliability. On 15 October 2026 the V4 Pro and V4.1 Flash APIs were partly unavailable from 00:11 to 05:04 UTC, a second major outage in 90 days after the three-hour outage of 14 September. source
53. Google Calendar API

    80.6 A →to 79.6 A −1 Correction

    • Schema & documentation 83 to 88 · Correction. Google doc pages return Markdown at a .md.txt suffix, though nothing advertises it. Drive, Sheets and Forms already got partial credit for that, and on 17 October Calendar was read the same way. source
    • Security & auth 87 to 77 · Correction. Calendar now loses 10 points on credentials, as Gmail, Sheets and Forms already did. Google's OAuth guide documents sending a token in the query string, and on 17 October the Calendar API read a token sent in the URL. source
54. Hookdeck

    76.9 BB →to 75.9 BB −1 Reliability

    • Reliability 90 to 85 · Reliability. On 15 October 2026 Event Delivery was degraded for 1 hour 53 minutes with high delivery latency, which Hookdeck put down to a Google Cloud Pub/Sub incident. With no downtime recorded it counts as a minor incident. source
55. Liquid d1

    43.5 E →to 44.5 E +1 Pricing

    • Payments & pricing 27 to 35 · Pricing. Liquid's pricing page now lists d1 at $0.02 and d1-ultra at $0.042 per million input tokens, where on 8 October the price was only in a blog post. New organisations get $2 of free credit, with no word on a card. source
56. OpenAI embeddings

    73.2 BB →to 72.2 BB −1 Correction

    • Reliability 65 to 60 · Correction. The 1 October check counted only the September incidents. Embeddings was also affected by hour-plus error incidents on 23 to 24 July and 25 July 2026, inside the same window, so four majors, as the OpenAI API listing reads the same feed. source
57. Parallel Search and Task APIs

    74 BB →to 75 BB +1 New evidence

    • Agent ergonomics 81 to 87 · New evidence. Earlier checks couldn't see the Search MCP's tool annotations. A live tools/list on 17 October showed both tools marked read-only, non-destructive and idempotent. Task creation still has no idempotency key. source
58. Restate

    73.3 BB →to 74.3 BB +1 Time

    • Reliability 91 to 96 · Time. The only incident in the earlier window, a 36-minute Cloud UI outage on 16 July 2026, has left the 90 days that start on 19 July. The feed lists nothing newer, so the record is clean. source
59. Stripe Identity

    77.7 BB →to 78.7 A +1 New evidence

    • Reliability 62 to 67 · New evidence. Stripe's Statuspage history at www.stripestatus.com, unread before, was read on 17 October. It shows 64 minutes of elevated API errors on 24 July 2026, counted as one major outage on the API the Identity endpoints use. source
60. Google People API

    65.2 B →to 64.3 B −0.9 Correction

    • Schema & documentation 79 to 84 · Correction. On 17 October People's doc pages returned Markdown at a .md.txt suffix with nothing advertising it. The other Google Workspace APIs got partial credit for that, so People was read the same way. source
    • Security & auth 86 to 76 · Correction. People now loses 10 points on credentials, as the other Google Workspace APIs do. Google's OAuth guide documents a query-string token, and on 17 October the People API read a token sent in the URL. source
61. Amazon Bedrock Guardrails

    74.8 BB →to 75.6 BB +0.8 Security New evidence Correction

    • Payments & pricing 20 to 30 · Correction. We had assumed every new AWS account needs a card. The Free Tier FAQ says most new customers don't, and new accounts get up to $200 in credits covering Bedrock, so the free-tier item now earns half, as on other AWS listings. source
    • Reliability 80 to 75 · New evidence. AWS's Health Dashboard history file, read in October, lists Bedrock in the me-central-1 and me-south-1 disruptions open since March 2026, counted as one major event. Per-Region quotas published in the AWS General Reference recovered part of the drop. source
    • Security & auth 94 to 96 · Correction. The security programme line was brought in line with the S3, Secrets Manager and Polly listings on the same evidence (HackerOne VDP, bulletins, SOC scope). The security.txt has since been renewed, and 2 points stay held back for no paid bounty. source
    • Transparency & trust 67 to 69 · Security. The renewed security.txt on aws.amazon.com, now valid to 16 October 2027, raised the provenance score, half of Transparency. The editorial assessment did not change. source
62. Gmail API

    77.8 BB →to 78.6 A +0.8 Correction

    • Schema & documentation 82 to 87 · Correction. Google doc pages return Markdown at a .md.txt suffix, though nothing advertises it. Drive, Sheets and Forms already got partial credit for that, and on 17 October Gmail was read the same way. source
63. Google Cloud Secret Manager

    76.5 BB →to 77.3 BB +0.8 New evidence

    • Security & auth 85 to 90 · New evidence. The 1 October check didn't read certifications. On 6 October Secret Manager was found in Google Cloud's SOC 2 scope and Google's security.txt, with its reward programme link, was re-read, so the security programme item now matches Google Model Armor. source
64. OpenCode

    67.7 B →to 68.5 B +0.8 Product

    • Reliability 68 to 72 · Product. OpenCode 2 shipped on 12 October 2026 as a new major with its breaking changes named in a migration guide, opt-in through opencode upgrade --major, while npm's latest and the install script stay on 1.x. source
65. You.com APIs

    76.5 BB →to 75.7 BB −0.8 Reliability New evidence

    • Agent ergonomics 84 to 81 · New evidence. The 6 October check found the TypeScript SDK @youdotcom-oss/sdk deprecated on npm since 3 March 2026 in favour of the MCP package, leaving Python as the only current SDK. source
    • Reliability 85 to 80 · Reliability. An authentication outage on the AI Search API was resolved on 5 October 2026, ending a record that had shown no incidents. The feed gives no duration, so it counts as minor. source
    • Security & auth 53 to 57 · New evidence. The troubleshooting page says Platform Logs list recent requests with error messages, response details and a request ID. No retention period or audit log of key actions was found. source
66. AgentMail API + MCP

    74.9 BB →to 75.6 BB +0.7 Terms Product

    • Maintenance & community 85 to 87 · Product. The MCP server now pins agentmail 0.5.46 and agentmail-toolkit 0.15.1, the latest on npm on 17 October, and runs security workflows in CI. Changelog entries followed on 7, 8, 12 and 14 October. source
    • Schema & documentation 85 to 88 · Product. agentmail-toolkit 0.14.0 rewrote the MCP tool descriptions on 13 October 2026. 25 of 45 tools now say which other tool to use or when not to call. source
    • Security & auth 62 to 65 · Product. Per-key permission lists (read on 6 October) and Agent Armor, which since 12 October screens inbound mail for prompt injection as an opt-in, raised it. The documented ?apiKey= query option now costs 10 points, as on Gmail and Drive. source
    • Transparency & trust 68 to 63 · Terms. The privacy notice was replaced on 12 October 2026. It drops the stated retention periods for 'as long as necessary' and now allows training models on email content unless agreed otherwise in writing. source
67. GitHub Copilot CLI

    57.6 C →to 58.3 C +0.7 Product

    • Security & auth 60 to 63 · Product. Local sandboxing became available to all users in 1.0.93 on 7 October 2026, and GitHub calls it generally available, with enterprise enforcement. It stays off by default and two organisation MCP policies still aren't enforced, so 3 points are held back. source
    • Transparency & trust 68 to 70 · Product. On 15 October 2026 GitHub gave dated notice, with named alternatives, that three models leave Copilot on 16 November. A dated notice with no written notice period is scored as it is for OpenAI Codex and Gemini CLI. source
68. Infisical

    83.7 A →to 84.4 A +0.7 Pricing

    • Payments & pricing 30 to 35 · Pricing. Infisical now publishes a per-request price for Agent Vault, 5,000 brokered requests a month free and $2 per 1,000 after, beside its per-identity plans. Plans plus a per-unit price score between plan-only and per-unit pricing, as for Freshdesk. source
69. Kiro CLI

    59.9 C →to 60.6 C +0.7 Security Product

    • Security & auth 66 to 69 · Security. The aws.amazon.com security.txt is valid again (to 16 October 2027), read as for every AWS listing, with 2 points still held back for kiro.dev's missing file. One more for 2.31.0 (13 October 2026), which brought the opt-in command sandbox to every V3 user. source
    • Transparency & trust 67 to 69 · Product. Kiro's models page added a model lifecycle section with Experimental, Active and EOL stages and dated end-of-life entries. No minimum notice period is stated, and we found no notice of the 14 October 2026 Sonnet 4 date before that day. source
70. Nevermined API + MCP

    70.8 BB →to 71.5 BB +0.7 Reliability New evidence Correction Negative event

    • Agent ergonomics 74 to 76 · New evidence. The 17 October re-grade read the live tools/list without a key. Refusals return a retryable flag and name the next tool, and repeated paid calls with the same requestId are safe, though none of the 12 tools carries annotations. source
    • Maintenance & community 89 to 81 · Correction. The platform is closed and its issue queues couldn't be read, so responsiveness is now scored on the 0 to 15 scale for closed services, as in the Paid listing. Catalog MCP 1.69.0 was published on 15 October 2026.
    • Negative events 0 to −2 · Negative event. Nevermined's certifications and pricing pages present ISO/IEC 27001 certification as held, while section 8.5 of its Data Policy says certification is in progress. At least one is wrong, read on 17 October 2026, as for ScaleDown. source
    • Payments & pricing 65 to 70 · New evidence. The 17 October re-grade found no card requirement on the $0 Free plan and a sandbox that runs on testnet funds with two free test merchants, so the free-tier item now gets full marks.
    • Reliability 72 to 63 · Reliability. Nevermined's Upptime status page stopped recording. On 17 October 2026 its newest commit was the 6 October summary, and it never monitored the MCP server or Router. Documented 429 handling with Retry-After and requestId offsets part of this. source
    • Schema & documentation 80 to 85 · New evidence. The 1 October check couldn't read the MCP schemas and didn't open the OpenAPI files. On 17 October all 12 tools showed typed JSON Schemas and descriptions saying when to use each, though some documented routes are missing from the OpenAPI files. source
    • Security & auth 63 to 74 · Correction. The 17 October re-grade scored the credential model as OpenRouter's is scored, with role-based bearer keys sent only in the header and OAuth PKCE limited to a Delegation a person approved. nevermined.app also serves a valid security.txt, and audit records were found. source
    • Transparency & trust 51 to 68 · New evidence. The 17 October re-grade read the Data Policy on nevermined.app, which the 1 October check had found only under NDA, with retention by category and seven sub-processors listed with locations. The versioning page was read as a deprecation policy. source
71. Pinecone API + MCP

    76.2 BB →to 76.9 BB +0.7 New evidence

    • Transparency & trust 73 to 81 · New evidence. A public Data Processing Addendum dated 24 May 2024 was read on 6 October. The trust centre's Subprocessor List, which would not load then, was read on 17 October, with nine entries each giving a purpose and a location. source
72. Clearbit

    13.5 F →to 14.1 F +0.6 New evidence

    • Transparency & trust 66 to 72 · New evidence. The sub-processor page, not read in full on 14 October, names nine sub-processors with purpose and country and promises 30 days' notice of a new one. source
73. Grafana MCP Server

    74.5 BB →to 73.9 BB −0.6 Product New evidence Negative event

    • Maintenance & community 95 to 96 · Product. Grafana published v2.0.2 to the official MCP registry on 10 October 2026, so the registry entry is now current with the latest release. source
    • Negative events −3 to −4 · Negative event. The X-Grafana-URL deduction, earlier read as an unpublished fix, now rests on two published advisories at 2 points each, CVE-2026-15583 (high, 15 July 2026) and CVE-2026-19516 (critical SSRF, 11 August 2026, fixed in 1.1.0). source
    • Security & auth 65 to 67 · New evidence. The 8 October check found no advisory for this server. Grafana Labs had published two, on 15 July and 11 August 2026, which shows advisories handled in public. Still no bounty, security.txt or SECURITY.md. source
74. Tempo

    76.6 BB →to 77.2 BB +0.6 Product New evidence

    • Agent ergonomics 78 to 94 · New evidence. Both MCP servers' tool lists were read on 6 October. The API server loads its tools on demand through four small meta-tools, every tool carries read-only and destructive annotations, and several writes take an Idempotency-Key. source
    • Reliability 71 to 66 · Product. On 9 October 2026 Tempo replaced its published rate-limit table with a statement that limits depend on the endpoint, key and billing state. Only the 100 a minute for anonymous MCP calls is still published. source
    • Schema & documentation 88 to 90 · New evidence. Reading both MCP servers on 6 October showed they no longer conflict. mcp.tempo.xyz serves documentation and api.tempo.xyz/mcp the API, whose meta-tools say when to inspect a tool before calling it. source
    • Security & auth 70 to 60 · New evidence. The OpenAPI, read on 6 October, lists a key query parameter as a supported alternative for API keys, so a secret can travel in a URL. The checklist takes points off for that. source
    • Transparency & trust 63 to 67 · New evidence. The privacy policy, last updated 31 July 2026, was re-read on 6 October. It names Tempo Labs Inc., covers US transfers under SCCs and names PostHog, though it sets no retention periods. source
75. Medusa API + MCP

    63.5 B →to 63 B −0.5 Product Negative event

    • Negative events 0 to −2 · Negative event. Two preview builds of @medusajs/medusa from 24 November 2025 are listed as malicious code from the Sha1-Hulud npm wave. Only the preview dist-tag was affected, and we found no Medusa statement. source
    • Security & auth 40 to 49 · Product. Medusa launched a hosted, read-only MCP server in October 2026, with environment choice on the consent page, revocable connections and MFA confirmation before tools read store data. The pricing page lists Audit Logs on Enterprise. source
76. Stripe API + MCP

    82.4 A →to 82.9 A +0.5 Product New evidence

    • Agent ergonomics 94 to 91 · Product. The MCP page listed 15 tools on 17 October against 10 on 4 October, adding Atlas company formation and Treasury tools in public preview. More tools cost more context. source
    • Reliability 63 to 68 · New evidence. Stripe's Statuspage history at www.stripestatus.com, unread on 4 October, was read on 17 October. It shows 64 minutes of elevated API errors on 24 July 2026, counted as one major outage. source
77. OpenAI API

    83.3 A →to 82.9 A −0.4 Product

    • Maintenance & community 91 to 87 · Product. On 14 October 2026 OpenAI gave GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini a shutdown date of 14 January 2027, 92 days, with no safety or compliance reason. Its deprecations page promises GA models at least six months. source
78. OpenAI Codex

    73 BB →to 73.4 BB +0.4 New evidence

    • Transparency & trust 79 to 84 · New evidence. The 1 October check read only the GitHub releases. The changelog at learn.chatgpt.com gives dated model retirements with replacements, such as GPT-5.5 announced on 14 September for 14 October 2026, though no notice period is written down. source
79. ZeroEntropy zerank and zembed

    13.7 F →to 14.1 F +0.4 New evidence

    • Transparency & trust 52 to 57 · New evidence. Our editorial score did not change. The provenance score, half of Transparency, rose because a status page at status.zeroentropy.dev, missing from the listing before, was recorded on 17 October. source
80. Arize Phoenix

    75.4 BB →to 75.7 BB +0.3 Correction

    • Security & auth 56 to 58 · Correction. The remote MCP docs tell agents to treat query results as data, not instructions. The guidance was in the repository by 11 September and was missed on 1 October. It now gets the award used elsewhere for guidance without a mitigation. source
81. Cloudflare R2

    77.1 BB →to 77.4 BB +0.3 New evidence

    • Security & auth 83 to 85 · New evidence. Cloudflare's public ISO certifications page, unread on 4 October, lists ISO 27001, 27018 and 27701, and the disclosure policy names a HackerOne bounty. 2 points stay held back because security.txt has no Expires field. source
82. Nylas Email API

    78.7 A →to 79 A +0.3 Product

    • Security & auth 87 to 89 · Product. Nylas documents a public API for IAM principals and credentials, ten operations with IAM Admin and Auditor roles, so keys can be issued and rotated from code. The guide is dated 9 October 2026, and the 8 October check read IAM as Dashboard-only. source
83. AWS Secrets Manager

    77.7 BB →to 77.5 BB −0.2 Security Time

    • Maintenance & community 65 to 60 · Time. Workload Credentials Provider 3.1.0 of 15 July 2026 has passed out of the 90-day window, leaving one release, 3.1.1 of 21 July. The Secrets Manager API has not changed since 11 December 2025. source
    • Transparency & trust 77 to 79 · Security. The security.txt on aws.amazon.com, expired since 24 September 2026, was renewed to 16 October 2027. That raised the provenance score, half of Transparency, while the editorial assessment did not change. source
84. Amazon Bedrock AgentCore Identity

    74.8 BB →to 75 BB +0.2 Security

    • Transparency & trust 75 to 77 · Security. AWS renewed the security.txt on aws.amazon.com, now valid to 16 October 2027 after expiring on 24 September 2026. That raised the computed provenance score, half of Transparency. source
85. Black Forest Labs FLUX API

    63.9 B →to 64.1 B +0.2 Correction

    • Transparency & trust 62 to 65 · Correction. The release notes carry dated deprecation notices, such as FLUX Pro 1.0 and the Finetuning API announced on 3 October 2025 for 31 October 2025, which the 1 October note didn't credit. There is still no written policy. source
86. Prefactor

    65.6 B →to 65.7 B +0.1 Time

    • Transparency & trust 42 to 44 · Time. prefactor.tech, registered on 16 October 2024, is now 2 years old, a step of the domain-age check. That moved the provenance score, half of Transparency, from 50 to 54.

15 October 2026 2 moves

1. Claude Agent SDK

   72.1 BB →to 70.1 BB −2 Negative event

   • Negative events −6 to −8 · Negative event. A deduction was added for fixes in the bundled Claude Code CLI, 2.1.288 to 2.1.299 (2 to 14 October 2026), disclosed only in its changelog. They include a dangerous rm running without a prompt and sandbox writes reaching denied paths. source
2. Claude Code

   61.9 C →to 59.9 C −2 Negative event

   • Negative events −6 to −8 · Negative event. Claude Code 2.1.288 to 2.1.299 (2 to 14 October 2026) fixed permission and sandbox gaps, among them a dangerous rm inside bash -c running without a prompt. They're disclosed only in the changelog, so the group counts as one deduction. source

12 October 2026 9 moves

1. HeyGen Voice

   71.3 BB →to 70.9 BB −0.4 New evidence

   • Transparency & trust 69 to 64 · New evidence. Its terms of service and privacy policy on heygen.com were read in full on 12 October 2026, so the two provenance checks count what they state, where before they counted only that each was published.
2. Agentzon

   51.4 D →to 51.1 D −0.3 New evidence

   • Transparency & trust 45 to 41 · New evidence. Read on 12 October, agentzon.co's terms (effective 8 October 2026) state no governing law, termination or change terms, and its privacy policy no retention period or user rights. That lowers the provenance score, half of Transparency. source
3. Antigravity CLI

   57.7 C →to 57.4 C −0.3 New evidence

   • Transparency & trust 54 to 51 · New evidence. Its terms of service and privacy policy on antigravity.google were read in full on 12 October 2026, so the two provenance checks count what they state, where before they counted only that each was published.
4. Drex 1.5

   69.7 B →to 69.5 B −0.2 New evidence

   • Transparency & trust 63 to 61 · New evidence. Its terms of service and privacy policy on nace.ai were read in full on 12 October 2026, so the two provenance checks count what they state, where before they counted only that each was published.
5. Microsoft-Decision-1

   39 E →to 38.8 E −0.2 New evidence

   • Transparency & trust 32 to 29 · New evidence. The Azure OpenAI data privacy page it links as its privacy policy was read in full on 12 October 2026. It states few of the things a reader expects in a privacy policy, such as how long data is kept or who to contact.
6. rotki MCP server

   45.6 E →to 45.4 E −0.2 New evidence

   • Transparency & trust 48 to 46 · New evidence. Its terms of service and privacy policy on rotki.com were read in full on 12 October 2026, so the two provenance checks count what they state, where before they counted only that each was published.
7. Hyperbolic

   48 D →to 48.1 D +0.1 Time

   • Transparency & trust 61 to 63 · Time. hyperbolic.ai, registered on 11 October 2021, is now 5 years old, a step of the domain-age check. That moved the provenance score, half of Transparency, from 76 to 80.
8. Kite Agent Passport

   51.5 D →to 51.4 D −0.1 New evidence

   • Transparency & trust 56 to 54 · New evidence. Its terms of service and privacy policy on agentpassport.ai were read in full on 12 October 2026, so the two provenance checks count what they state, where before they counted only that each was published.
9. Qwen-Image 2.1 (Model Studio)

   64.7 B →to 64.6 B −0.1 New evidence

   • Transparency & trust 66 to 65 · New evidence. Alibaba Cloud's Model Studio privacy notice was read on 12 October 2026 and couldn't be read in full, so the privacy check stopped counting it as a published policy that states what a reader expects.

11 October 2026 1 move

1. MiniMax Video API

   60 C →to 61 C +1 Reliability

   • Reliability 70 to 75 · Reliability. MiniMax's rate-limits page now lists 300 requests a minute and 30 tasks in flight for MiniMax-H3 on Video Generation V2, and 20 a minute for Hailuo. On 1 October pay-as-you-go limits weren't published. source

10 October 2026 3 moves

1. Gemini CLI

   72 BB →to 66.7 B −5.3 New evidence Correction Negative event

   • Negative events −2 to −5 · Negative event. A deduction was added because the README (9 October 2026) and the authentication page still advertise the free tier of 1,000 requests a day that ended on 18 June 2026. source
   • Payments & pricing 40 to 20 · Correction. Google ended the free Google sign-in tier of 1,000 requests a day on 18 June 2026, which the 1 October check still credited. Whether an unpaid API key keeps a free allowance is unclear. source
   • Transparency & trust 87 to 89 · New evidence. Google publishes dated notices for the change (the 19 May 2026 announcement, the 18 June shutdown, and end-of-sale and renewal dates from 9 October 2026 to December 2027), which raise the deprecation-notice item. source
2. LocalGhost

   52.6 D →to 55.5 C +2.9 Product

   • Maintenance & community 78 to 79 · Product. wisp 0.0.6, released on 10 October 2026, passes the Go suite at the tag and on main, where CI was failing at 0.0.5. source
   • Reliability 55 to 67 · Product. At wisp 0.0.6 (10 October 2026) the Go suite passes, including TestFetchListAndSpools, which failed at 0.0.5. The release also ships the fix for the Wikipedia import that stopped at an invalid byte. source
   • Security & auth 47 to 50 · Product. Since 0.0.6, DEPLOY, ROLL BACK and LOCK BOX NOW ask for confirmation and deletes take a second tap. Device certificates now last two weeks and renew daily, but carry no scopes. source
   • Transparency & trust 77 to 76 · Product. The README still describes the weather pull of 0.0.5, while the privacy page (updated 10 October 2026) and the code describe 0.0.6's, so data handling loses a point.
3. Clef

   66.1 B →to 65.9 B −0.2 Product

   • Maintenance & community 56 to 54 · Product. On 9 October 2026 Cloudflare cut hosted Clef-flash's context from 64k to 24,576 tokens on an unversioned model ID, with no notice period and no changelog entry. That counts as model churn. source

8 October 2026 422 moves

6 October 2026 6 moves

1. LocalGhost

   45.8 E →to 53 D +7.2 Terms Product Correction Negative event

   • Agent ergonomics 2 to 39 · Correction. The 6 October re-grade scored the phone's device API as the surface graded, where the 3 October grade found no reachable surface. A reading change, not a code change. Every error still reaches the caller as one 503.
   • Maintenance & community 71 to 78 · Product. The release toolchain moved to Go 1.27.1 in 0.0.4, no issues are open, and each tag ships signed builds. CI was failing on main at the time of the check.
   • Negative events −3 to −2 · Negative event. The location claim is fixed. wisp 0.0.4 removed the hand-off to Android's geocoder, and the privacy page and llms.txt list each lookup with the release that ended it, so the deduction drops to the minimum. source
   • Reliability 59 to 55 · Product. At wisp 0.0.5 (5 October 2026) and on main, TestFetchListAndSpools fails after a commit that day added a twelfth news feed, so CI isn't passing. A box now installs from the signed bundle with install.sh. source
   • Schema & documentation 44 to 49 · Correction. The device API's OpenAPI document, in the code since July and covering 46 of 105 paths, is now graded as the contract. Under the same reading, examples and errors score lower because every error becomes one 503.
   • Security & auth 52 to 47 · Correction. Least privilege and audit were read against the device API rather than the operator's terminal. DEPLOY in the app is one tap and there's no per-call audit trail.
   • Transparency & trust 77 to 82 · Terms. The privacy page, updated 5 October 2026, lists every outbound lookup, and the code at v0.0.5 matches it, which lifts the data-handling line. source
2. Gemini Developer API

   62 B →to 66 B +4 New evidence

   • Agent ergonomics 85 to 90 · New evidence. The v1beta Discovery document shows function calling mode ANY with allowedFunctionNames, which forces a named tool, and a VALIDATED mode. The 1 October check had not checked forcing modes. source
   • Schema & documentation 65 to 85 · New evidence. Google serves a keyless v1beta Discovery document (revision 20261004) typing 86 methods and 224 schemas. The 1 October check had not looked for one. It doesn't cover the Interactions API. source
3. Azure AI Speech speech-to-text

   77 BB →to 73.3 BB −3.7 Reliability Security

   • Reliability 90 to 80 · Reliability. Azure's 23 July 2026 post-incident review for West US names Azure Speech, with connectivity failures from 14:44 to 19:41 UTC. The 1 October check missed it, and it now counts as one major outage. source
   • Security & auth 95 to 85 · Security. The Realtime route for MAI-Transcribe-2-Streaming, added on 1 October 2026, documents passing the resource key as an api-key query-string parameter on the WebSocket URI. The other speech-to-text routes take the key in a header only. source
4. ElevenLabs Text to Speech API + MCP

   73.1 BB →to 75.4 BB +2.3 New evidence

   • Agent ergonomics 82 to 92 · New evidence. ElevenLabs' OpenAPI spec shows GET /v2/voices paging with page_size and next_page_token and filtering by search, category, language and more. The 1 October check couldn't confirm paging or filters. source
   • Security & auth 60 to 65 · New evidence. The Eleven v4 launch page states SOC 2 Type II, ISO 27001 and PCI DSS Level 1. These count as vendor claims, because the trust centre didn't render and no report was seen. source
   • Transparency & trust 71 to 69 · New evidence. The Eleven v4 launch page says scripts aren't used for training without consent, while the Terms (31 March 2026) and Privacy Policy (20 May 2026) allow training unless the user opts out. Statements that don't agree score lower on data handling. source
5. Azure AI Speech text-to-speech

   73.7 BB →to 71.7 BB −2 Reliability

   • Reliability 90 to 80 · Reliability. Azure's 23 July 2026 post-incident review for West US (ZJV6-SGG) names Azure Speech, with connectivity failures from 14:44 to 19:41 UTC. The 1 October check missed it, and it now counts as one major outage. source
6. Cohere Embed and Rerank

   72.5 BB →to 72.8 BB +0.3 Reliability New evidence

   • Maintenance & community 87 to 90 · New evidence. PyPI lists cohere 7.2.0 (28 September 2026), matching the GitHub releases. On 1 October our reader saw 6.1.0 on PyPI, so the published version was unclear. source
   • Payments & pricing 35 to 40 · New evidence. The Rerank 4 prices, $2.00 (Fast) and $2.50 (Pro) per 1,000 searches, were read from the pricing page's retrieval tab, which didn't render on 1 October. source
   • Reliability 83 to 73 · Reliability. The 1 October note said no incidents were posted, but Cohere's status page records degraded embed and rerank performance on 1 September 2026, 15:41 to 19:47 UTC, during a Google Cloud outage. Cohere rated it minor. source
   • Security & auth 50 to 55 · New evidence. Cohere's Trust Center has a downloadable Responsible Disclosure Policy for its bug bounty and lists ISO 27001 and ISO 42001 beside SOC 2 Type II. There's still no security.txt. source
   • Transparency & trust 69 to 76 · New evidence. Cohere's Trust Center publishes a subprocessor list with a location for each, all in the USA, and says hosting is on Google Cloud US-Central. The 1 October check found no such list. source

5 October 2026 14 moves

1. Novu

   77.4 BB →to 64.6 B −12.8 New evidence Negative event

   • Negative events 0 to −13 · Negative event. Deductions were added for Novu's published advisories, missed on 1 October. One Critical (GHSA-w2vj-px2p-76fr, 7 July 2026) let any account extract another organisation's secret keys, and seven High advisories date from February to June 2026. source
   • Security & auth 63 to 64 · New evidence. Novu has published 34 GitHub security advisories with reporter credits, which the 1 October check had not counted. There's still no bug bounty, and novu.co returned 404 for security.txt on 4 October 2026. source
2. Pydantic AI

   80 A →to 83.9 A +3.9 New evidence Negative event

   • Agent ergonomics 85 to 95 · New evidence. The MCP docs show a two-line agent and toolsets filtered on tool annotations, prefixed, or deferred behind tool search. These were public on 1 October but not seen by that check. source
   • Negative events −2 to −4 · Negative event. A deduction was added for GHSA-h4xc-3qfq-jf93 (High, 12 August 2026), where a website could trigger agent runs through the local web chat UI. It was fixed in 2.28.0 and missed on 1 October. source
   • Schema & documentation 95 to 100 · New evidence. The docs say when not to reach for a feature. The graph page points readers to other multi-agent approaches first, and the MCP page says when to use MCPToolset. The 1 October check had not re-read this. source
   • Security & auth 80 to 95 · New evidence. Read-only workspaces in core (since 2.52.0 on 30 September 2026), Harness sandboxes and prompt-injection guidance in the docs were public but missed on 1 October. source
   • Transparency & trust 77 to 87 · New evidence. The Logfire docs say nothing is sent unless instrumentation is configured. The provenance record now has Pydantic's terms, privacy policy (updated 22 August 2026) and a valid security.txt, none of which the earlier check had. source
3. Apify MCP Server

   78.6 A →to 76.6 BB −2 Negative event

   • Negative events −3 to −5 · Negative event. A deduction was added for GHSA-6gr2-qh89-hxwm (CVE-2026-50143, High). Versions before 0.10.11 could send the user's Apify token to a malicious Actor's host. It was fixed on 28 May 2026 and missed on 1 October. source
4. Infisical

   81.9 A →to 83.9 A +2 New evidence

   • Reliability 90 to 100 · New evidence. Infisical publishes an SLA for Enterprise Cloud customers, 99.99 per cent monthly availability with service credits of 10, 20 or 30 per cent, linked from its terms hub and pricing page. The 1 October check missed it. source
5. MongoDB MCP Server

   78.6 A →to 80 A +1.4 New evidence

   • Reliability 85 to 88 · New evidence. MongoDB's docs keep dated release notes for the server, including the 3.0.0 breaking changes. The 1 October check found no notes for v3.0.0, and the GitHub release body is still empty. source
   • Schema & documentation 79 to 81 · New evidence. The dated release notes in MongoDB's docs (1.8.1 to 3.0.4) now count towards versioning, less a point because the GitHub v3.0.0 release has no notes and the docs don't list 3.0.5 yet. source
   • Transparency & trust 78 to 84 · New evidence. The provenance check now records MongoDB's status page at status.mongodb.com, which the listing had left blank. The page covers MongoDB Cloud and has no MCP server component. source
6. Google Calendar API

   79.5 A →to 80.8 A +1.3 New evidence

   • Security & auth 82 to 87 · New evidence. Google Workspace SOC 2 Type II reports, issued quarterly through Compliance Reports Manager, list Calendar in scope. The 1 October check had not checked certifications. source
   • Transparency & trust 79 to 84 · New evidence. We read the Workspace sub-processor list (last modified 23 July 2026), which names each sub-processor with the countries it processes in and covers Calendar. The 1 October check had not re-read it. source
7. Google Drive API + MCP

   78.6 A →to 79.9 A +1.3 New evidence

   • Security & auth 86 to 91 · New evidence. Drive log events, read for the first time, record third-party app actions with the OAuth client ID, app name and API method for Workspace admins. Google Workspace SOC 2 Type II reports list Drive in scope. source
   • Transparency & trust 74 to 78 · New evidence. We read the Workspace sub-processor list (updated 23 July 2026), which names Drive and gives processing countries for each sub-processor. It is scored the same way as on the Google Calendar listing. source
8. Bird API + MCP

   77.7 BB →to 76.5 BB −1.2 New evidence Correction

   • Agent ergonomics 90 to 95 · New evidence. Bird's generated API contract lists the Idempotency-Key header on SMS and WhatsApp sends, and the TypeScript SDK reuses one key across retries. The 1 October check had not confirmed this for sends. source
   • Reliability 73 to 63 · Correction. The 1 October check counted four incidents in 90 days, when Bird's status feed listed 12 entries, one of them scheduled maintenance. One now counts as major, 75 minutes of elevated US API errors from 15:31 UTC on 17 September 2026. source
9. Amazon S3

   79.3 A →to 78.2 A −1.1 Pricing Reliability New evidence

   • Payments & pricing 20 to 30 · Pricing. AWS's Free Tier FAQ now says most new customers can sign up without a payment method, under a sign-up launched in October 2026. AWS says it reaches a limited number of customers so far, so the free-tier item earns half. source
   • Reliability 95 to 80 · Reliability. AWS's history lists S3 in a service disruption in ME-CENTRAL-1 (UAE) open since 1 March 2026. On 15 September 2026 AWS said data held only in the mec1-az2 zone can't be restored. source
   • Security & auth 86 to 88 · New evidence. The 4 October re-check confirmed that AWS's SOC reports list S3 in scope, which the 1 October check had not re-read. The security.txt on aws.amazon.com was still expired.
   • Transparency & trust 80 to 84 · New evidence. We read AWS's sub-processor page (updated 28 July 2026), which names third-party sub-processors and says processing happens in the Regions the customer selects. The 1 October check had not read it. source
10. Cloudflare R2

    78.4 A →to 77.4 BB −1 New evidence

    • Payments & pricing 30 to 20 · New evidence. R2's get-started guide says to complete a checkout to add an R2 subscription, and Cloudflare's billing docs say a payment method is required to buy products. The 1 October check had not established whether a card was needed. source
    • Transparency & trust 75 to 77 · New evidence. We read Cloudflare's sub-processor list (last updated 1 October 2025), which gives each sub-processor's purpose and locations. The 1 October check had not read it. source
11. Descope Agentic Identity Hub

    79.2 A →to 78.3 A −0.9 New evidence

    • Maintenance & community 76 to 74 · New evidence. The Agent Auth SDK isn't published. @descope/agent-auth returns 404 on npm and descope-agent-auth returns 404 on PyPI, though its README gives install commands for both. source
    • Schema & documentation 82 to 78 · New evidence. The Fetch Outbound App User Token reference, unreadable on 1 October, was read on 4 October 2026. It marks no field required, appId and userId included, and lists no enums or constraints. source
12. OpenAI API

    82.8 A →to 83.5 A +0.7 New evidence

    • Transparency & trust 85 to 93 · New evidence. We read OpenAI's subprocessor list (updated 9 July 2026), with processing countries for each, and its DPA (effective 1 January 2026) for the first time. Neither conflicts with the data-controls guide. source
13. Stripe API + MCP

    82.4 A →to 82.6 A +0.2 New evidence

    • Transparency & trust 87 to 89 · New evidence. We read Stripe's sub-processor list (last updated 27 September 2026), which names each provider with its data, purpose and entity country and gives 30 days to object. It gives entity countries, not processing locations. source
14. AWS Secrets Manager

    78.1 A →to 78 A −0.1 Pricing Reliability New evidence Correction

    • Payments & pricing 20 to 30 · Pricing. AWS's Free Tier FAQ now says most new customers can sign up without a payment method. AWS says the new sign-up reaches a limited number of customers so far, so the free-tier item earns half. source
    • Reliability 87 to 77 · Reliability. AWS's history lists Secrets Manager in the ME-CENTRAL-1 (UAE) service disruption open since 1 March 2026, the same event that affects S3. The 1 October check had read only the us-east-1 feed.
    • Security & auth 88 to 91 · New evidence. The re-check confirmed Secrets Manager is in AWS's SOC 1, 2 and 3 scope (list updated 11 August 2026), which the 1 October check had not re-read. The expired security.txt still costs points. source
    • Transparency & trust 79 to 80 · Correction. The sub-processor item now gets full marks, to match the Amazon S3 listing, which is scored on the same AWS sub-processor list (updated 28 July 2026).

3 October 2026 1 move

1. LocalGhost

   27.6 F →to 45.8 E +18.2 Checklist Negative event

   • Agent ergonomics 0 to 2 · Checklist. From 3 October 2026 LocalGhost is graded by the normal checklist, at our founder's request, instead of the stricter rule. The single documented 503 now earns a little on the error-handling item, though no surface an agent can call reaches the box.
   • Maintenance & community 39 to 71 · Checklist. Graded by the normal checklist from 3 October 2026 instead of the stricter rule. With wisp 0.0.3 that day, three releases in 16 hours meet the release count, and the signed builds on GitHub and the mirror now count. source
   • Negative events −6 to −3 · Negative event. Under the normal rule the deductions became one, for the README and privacy page saying no position leaves the phone while the app sends location fixes to Android's geocoder. The Open-Meteo position call was removed in 0.0.3 on 3 October 2026. source
   • Payments & pricing 40 to 60 · Checklist. Graded by the normal checklist from 3 October 2026 instead of the stricter rule. The self-hosted rule gives free software with no account the autonomous onboarding points, which the strict reading had withheld.
   • Reliability 44 to 59 · Checklist. Graded by the normal checklist from 3 October 2026 instead of the stricter rule, which raised the package, open-issues and semver items. CI's server tests also passed at the wisp 0.0.3 tag and on main.
   • Schema & documentation 30 to 44 · Checklist. Graded by the normal checklist from 3 October 2026 instead of the stricter rule. Descriptions, worked examples and versioning (semver tags, notes per release, a dated changelog) were read as for other listings.
   • Security & auth 32 to 52 · Checklist. Graded by the normal checklist from 3 October 2026 instead of the stricter rule. The query-string deduction for the enrolment QR no longer applies, and confirmations, the fetch log and SECURITY.md earn what other listings get.
   • Transparency & trust 67 to 77 · Checklist. Graded by the normal checklist from 3 October 2026 instead of the stricter rule. Licence, telemetry and per-release compatibility notes earn more, while data handling still loses points because the privacy page disagrees with the code.