Arcmira

# Fix list: Arcmira

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/arcmira, the October 2026 research run, assessed 30 October 2026. Grade B, 64.2 out of 100.

This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public.

For a coding agent working on Arcmira: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published.

## 1. Reliability, 40 out of 100, up to 12 more on the total

Why it scored 40: Hosted reading, for the MCP server with the HTTP API it fronts. No public status page was found. None is linked from the docs, llms.txt, developers.md or the footer, https://arcmira.com/status answers 403 and status.arcmira.com does not answer (0). No readable incident history, so the floor (5). Limits are published with numbers on the Plans and usage page, 60 requests a minute per key on Free, 240 on paid plans and 600 on Enterprise, with RateLimit-Limit, RateLimit-Remaining and RateLimit-Reset on every authenticated response and a 30-a-minute cap on bad keys per address (15). 429 rate_limited carries Retry-After and retry_after_seconds, developers.md asks for exponential backoff, 503 search_unavailable and transcript_fetching carry Retry-After, every POST and PATCH takes an Idempotency-Key that replays with Idempotency-Replayed, a repeated Premium read is never charged twice, and a webhook secret can be recovered under its key for 24 hours (15). No SLA. The terms (effective 13 October 2026) disclaim uninterrupted or error-free service, and the Enterprise page lists priority support but no uptime commitment (0). The HTTP API is /v1 at document version 1.0.0 with nothing labelled beta, but the MCP server agents connect to is 0.10.7 after 30 registry versions in six weeks, two rounds of tool renames and a retired-tool error for stale hosts, and chat is in beta (5 of 10). 0+5+15+15+0+5 = 40.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-reliability):

Hosted APIs, MCP servers, models and platforms.

- 20, a public status page with component history (Statuspage, Instatus, BetterStack or the vendor's own).
- 0 to 30, the incident record for the last 90 days on that page. 30 for a clean record or trivial incidents only, 20 for minor incidents only, 10 for one major outage (an hour or more of a core API down, or errors across the board), 0 for several. 5 when there's no history we could read, and the note says so.
- 15, rate limits documented with numbers.
- 15, documented 429 or overload handling (Retry-After, backoff guidance), and idempotency keys or safe-retry guidance where writes are involved.
- 10, an SLA published for any paid tier.
- 10, the surface agents use is generally available, not beta or preview.

Local packages, SDKs, frameworks and stdio MCP servers.

- 20, installs from an official package with supported runtimes stated.
- 25, a public CI and test suite, passing on the default branch.
- 0 to 25, open crash or regression issues relative to activity (25 for few and handled, 0 for many, old and unanswered).
- 15, semver discipline and breaking changes called out in a changelog.
- 15, version 1.0 or later, or declared stable.

Protocols are read from their reference implementations, the public facilitators or servers, spec stability and test vectors.

## 2. Security & auth, 61 out of 100, up to 6.8 more on the total

Why it scored 61: OAuth 2.1 for MCP hosts (authorization code with PKCE S256, dynamic client registration, refresh tokens, revocation and introspection endpoints, scopes read, monitors:write, trackers:write and recommendations:read, access tokens of an hour that refresh for 30 days, revoked per host in the dashboard) and arc_sk_ keys sent only as a bearer header, shown once, several per account with a label and last-used time, disabled at once; sign-up keys carry read only and dashboard keys all four scopes, and nothing documents a secret in a query string (30). The sign-up key and the OAuth consent give read access by default; arcmira_execute_read reaches GET /v1/* only, arcmira_execute_write adds POST and PATCH under monitors and trackers and never DELETE; programs run in an isolate whose only network is a proxy that adds the credential, so the key never enters the program; spending is bounded by the plan and a dashboard budget no tool can raise, though a read tool call can start a Premium transcription on a paid plan (17 of 20). Every passage and transcript is third-party speech, and the only guidance found is to check the linked source for speaker and classification errors and to re-fetch llms.txt or the skill to confirm an instruction is genuine; no prompt-injection guidance for the model reading results (4 of 15). X-Request-Id and RateLimit headers on every response, GET /v1/me for the account, role, scopes and balance, the keys page showing who made each key and when it was last used, and an account activity log on Enterprise only; no per-call log for the operator below that, while every tool call is recorded in the vendor's analytics (8 of 15). No security.txt on arcmira.com, api.arcmira.com or mcp.arcmira.com (all 404), no disclosure policy, bug bounty, SOC 2 or ISO 27001, and the terms forbid probing; the repository enforces signed commits from one identity in CI (2 of 20). 30+17+4+8+2 = 61.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-security):

- 0 to 30, the credential model. 30 for OAuth 2.1 with scopes, or scoped and revocable keys with rotation. 20 for plain revocable API keys. 10 for one all-powerful key. 10 off when a secret can travel in a URL query string as a documented option.
- 0 to 20, read-only or least-privilege modes, and confirmation or approval for destructive actions.
- 0 to 15, prompt-injection posture where the tool returns untrusted content (documented mitigations or guidance). A tool that returns no untrusted content gets 10.
- 0 to 15, audit logs or per-call visibility for the operator.
- 0 to 20, a security programme. security.txt or a disclosure policy, a bug bounty, SOC 2 or ISO 27001, advisories handled in public.

Models are read for retention, whether API data trains models (and whether that's off by default), zero-retention options and certifications. Frameworks for telemetry defaults, approval hooks, guardrails and sandboxing.

## 3. Payments & pricing, 55 out of 100, up to 5.6 more on the total

Why it scored 55: No x402, MPP or L402 in the OpenAPI file, llms.txt, developers.md, the docs or the pricing page (0). Prices published without a login, in credits per read (4 a passage, mention or list row past the first 5 of a request, 10 a momentum call, 40 a sponsor row, 25 an alert, 4 per started 15 minutes of captions, 300 per started 15 minutes of a Premium transcript) and in dollars (Pro $20 for 20,000 credits, Pro+ $60 for 60,000, Ultra $200 for 400,000, on-demand $0.002 a credit) (20). Free plan with 1,000 credits a month, 2 monitored names and no card, per the pricing FAQ; the 7-day Pro trial needs a card (20). POST /v1/signups with an email and POST /v1/signups/verify with the six-digit code return a read-scoped key with no browser, so an agent with a mailbox can onboard itself; the code still has to be read from an inbox, and write scopes need a dashboard key or OAuth consent (15 of 20). 0+20+20+15 = 55.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-payments):

The published rubric, also on the [x402 page](https://www.anchorterminal.com/x402/).

- 40, a machine payment protocol (x402, MPP or L402) on the tool's own endpoints. 10 to 30 when it covers only some endpoints or only goes through a third party, and the note says which.
- 20, per-call or per-unit pricing published without a login. 10 for public plan-only pricing, 0 for "contact sales" or prices behind a login.
- 20, a free tier or trial that doesn't need a card.
- 20, autonomous onboarding, meaning an agent can get access without a person signing up in a browser (keyless use, x402, a programmatic key API).

Payment platforms and agent wallets rarely charge for their own API over a machine protocol, so the first line has steps for them, and the highest one that applies counts. 40 when x402, MPP or L402 runs on all their own endpoints, 30 when it runs on part of their own API, 25 when their merchants can accept one, 20 for running a facilitator, 15 for paying as a buyer, and 0 when the only protocol is their own. Merchant acceptance sits above a facilitator because the platform's own customers can charge agents through it, while a facilitator settles for sellers who wire up the protocol themselves. The counter-argument (a facilitator does more for the protocol as a whole) has a point. Each note says which step applied.

Open-source software you run yourself is scored on its hosted or paid option if it has one. A free, self-hosted package with nothing to buy gets 20, 20 and 20 for the last three lines, and 0 to 40 for the first only if it ships a payment protocol.

## 4. Transparency & trust, 61 out of 100, up to 3.4 more on the total

Made of editorial 48, provenance 73.

Why it scored 61: Closed service under terms effective 13 October 2026 that name Arcmira, Inc., San Francisco, California law and AAA arbitration; the MCP server's source is public under an all-rights-reserved notice, while the plugin, the TypeScript SDK and CLI and the Python SDK are Apache-2.0, and the Famous rankings are CC BY 4.0 (18 of 30). The privacy policy (effective 2 October 2026) states no retention periods, keeps limited usage and billing records after an account closes, discloses session replay and that private account data isn't used to train general models, and the MCP docs and README disclose that every tool call, including the first 4,000 characters of a program and the agent's intent, is logged to product analytics with credentials redacted and no opt-out; no DPA was found, and the terms' section 7A bars bulk use and model training on Arcmira data (14 of 30). Retirements are dated and typed in the changelog (POST /v1/trackers to 410 tracker_create_retired on 8 October, tool renames answering tool_retired), but the terms allow changes with or without notice, the quickstart says routes arrive without notice and no notice period is stated (8 of 20). The policy names service-provider categories (hosting, payment, analytics, AI processing) and US processing with other locations, and the docs name PostHog, Stripe, Cloudflare and Mintlify, but there is no subprocessor list and no data-location statement beyond the US (8 of 20). 18+14+8+8 = 48.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-transparency):

- 0 to 30, source availability and licence clarity. 30 for open source under an OSI licence, 15 for closed with clear terms, 0 for unclear terms.
- 0 to 30, data handling and retention statements that agree with each other (privacy policy, DPA, retention periods, subprocessors).
- 0 to 20, a deprecation policy or notices with dates.
- 0 to 20, telemetry disclosed with an opt-out (local software), or subprocessors and data locations disclosed (hosted).

The other half of Transparency and trust is the provenance score, computed from checked facts (below). The category score is the mean of the two.

Provenance checks not met in full (half of this category, computed from checked facts):

- Terms of service: read, states 5 of the 7 things a reader expects, and has 4 clauses that cost points (0.3 of 10)
- Privacy policy: read, states 6 of the 8 things a reader expects (8.5 of 10)
- Status page: not found (0 of 10)
- security.txt: not found (0 of 10)

## 5. Schema & documentation, 87 out of 100, up to 2.1 more on the total

Why it scored 87: MCP reading, with the API the tools call. A public OpenAPI 3.1 file at api.arcmira.com/v1/openapi.json (31 paths, 35 operations, 75 schemas, one bearer scheme, also served at arcmira.com/docs/openapi/arcmira-v1.json), and the four tools declare zod input schemas in src/tools.ts (25). llms.txt routed by situation, a .md twin of every docs page, developers.md, auth.md, a published Agent Skill, an RFC 9727 api-catalog and a .well-known/mcp card (10). Tool descriptions say what each does, that arcmira_describe is called once and bills nothing, that the write tool never deletes and when to send feedback; operation descriptions in the OpenAPI file say to call coverage before claiming a show isn't indexed and to resolve names before filters, and the server instructions carry the id rule, the 90-day default and the usage rule. When not to use a method is covered by the id rule and the resolver note, little else (16 of 20). The two execute tools take a JavaScript program as one string, checked only when it runs; the feedback tool has an enum category, a regex call_id and length caps; the API has 102 parameters, 37 with enums, ranges or patterns, and ids typed by prefix (9 of 15). Worked example programs in the reference, curl examples on every guide, an error catalogue of about 60 codes with a cause and fix each, doc_url anchors and request ids on every error, and typed unlock and details.limit objects; only 9 parameters in the OpenAPI file carry an example (14 of 15). /v1 in the path, a dated docs changelog from April 2026 with a 14-row table for the 5 to 8 October breaking changes, a per-version CHANGELOG.md in the repository and 30 registry versions; the OpenAPI file marks retired fields deprecated, but the quickstart says new fields and routes arrive without notice and nothing states a notice period (13 of 15). 25+10+16+9+14+13 = 87.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-schema):

APIs and MCP servers.

- 25, a machine-readable contract (a public OpenAPI file or similar; for MCP, typed JSON Schema inputs on every tool).
- 10, llms.txt or Markdown docs served for agents.
- 0 to 20, descriptions that say what a tool is for, when to use it and when not to, read from the tool definitions in the source or the API reference.
- 0 to 15, typed inputs with enums, constraints and required fields, and no free-form JSON blobs.
- 0 to 15, examples and documented error responses.
- 15, versioning and a public changelog.

Models are read from the API reference, the OpenAPI file, llms.txt, the structured-output and tool-use docs and the model cards. Frameworks from docs a model can follow, typed interfaces, examples and the API reference.

## 6. Agent ergonomics, 91 out of 100, up to 1.5 more on the total

Why it scored 91: MCP reading. Four tools whose definitions are short, but arcmira_describe returns a reference the vendor measures at 23,124 characters, loaded once a session, and the server instructions run to 1,995 characters; the vendor's own measurement for the change to code mode was 1,123 tokens of definitions against 8,163 for the earlier ten tools (18 of 25). limit 1 to 20 on search (default 5), cursors and has_more on lists, after and before windows, about, by, kind, channel and entity filters, start and end on transcripts, and program output capped at 12,000 characters, 3,000 per string and 100 per array, so a program returns only the fields an answer needs (20). One error envelope with a stable code, a message, param, gate, unlock.url, details.limit, retry_after_seconds and request_id, surfaced to the model with isError and the call_id to report (20). Idempotency-Key on every POST and PATCH with a replay header, a Premium read that is never charged twice, writes that pause rather than delete, and all four annotations declared on every tool; arcmira_execute_read is marked not read-only because a Premium read can start a transcription that uses credits, which a strict host may treat as a write (18 of 20). Few required inputs (one search string, ids resolved for free), a 90-day default window, defaults on limit, and official Apache-2.0 SDKs for TypeScript (npm arcmira 0.6.2) and Python (pip arcmira, repository v0.6.2) plus a CLI, with the caveat that the MCP client's argument names differ from the HTTP and SDK names (15). 18+20+20+18+15 = 91.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-ergonomics):

- 0 to 25, context cost. For MCP, the number and size of the tool definitions (25 for ten or fewer compact tools, 15 for 11 to 30, 5 for more than 30, plus up to 10 back for toolsets, dynamic loading or read-only subsets). For APIs, whether responses can be sized (field selection, limits, summaries).
- 20, pagination, filtering and output-size controls.
- 20, actionable, documented error responses, codes and messages an agent can recover from.
- 20, idempotency or safe retries, and for MCP the `readOnlyHint` and `destructiveHint` annotations.
- 15, sensible defaults, few required parameters, and official SDKs in at least two languages.

Models are read for tool use, structured output, prompt caching, context length, batch and SDKs. Frameworks for how much code and how many defaults a tool-calling agent with MCP needs.

## 7. Maintenance & community, 84 out of 100, up to 1.4 more on the total

Why it scored 84: The MCP server's newest release is 0.10.7, published to the registry on 12 October 2026 and to GitHub on 15 October, with repository commits to 30 October; the SDKs reached 0.6.2 on 15 October and the last dated API change in the docs changelog is also 15 October (30). 30 registry versions from 0.3.0 on 4 September to 0.10.7, 24 GitHub releases, 15 npm versions of arcmira since 22 August and 10 dated docs changelog entries in October (20). Issues on arcmira/mcp are mostly the maintainer's own (18 of 19 items are their own pull requests or notes); the one outside issue, #13 of 18 October, had no reply by 30 October; support is an email address, hi@arcmira.com, and a free feedback endpoint whose review states can be read back; no Discord, forum or community channel (11 of 25). io.github.arcmira/arcmira is in the official MCP registry under the GitHub-verified organisation namespace, current at 0.10.7, and the TypeScript and Python SDKs are current (15). CI on every push and pull request (typecheck, tests, manifest check against the OpenAPI contract, signed-commit check) passing on master on 30 October, pinned dependencies and a lockfile; the server itself is not a package anyone installs, so package health is the SDKs', whose CI we didn't read (8 of 10). 30+20+11+15+8 = 84.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-maintenance):

- 0 to 30, time since the last release, or the last published model or API change for a closed service. 30 within 30 days, 20 within 90, 10 within 180, 0 older.
- 20, at least three releases or dated changelog entries in the last 90 days.
- 0 to 25, responsiveness. Issues and pull requests answered on GitHub (the open issues and how recent the replies are). For closed services, a public changelog and a support or community channel that answers, 0 to 15.
- 15, presence in the official MCP registry under a verified namespace (MCP servers), or current official SDKs (APIs and models).
- 10, package health, current dependencies and CI.

Models are read for deprecation notice periods and model churn rather than release counts.

## Deductions

Each comes off the total. A fixed and documented problem counts for less at the next check.

- Between 5 and 8 October 2026 the v1 API retired POST /v1/trackers (410 tracker_create_retired) and six tracker delivery fields for the team account model. The changelog says the first seven rows can break a client, and its entry is dated 8 October, the day the change finished, with no advance notice found. Documented the same day with typed codes, a 410 and a migration table, so deducted at the low end (-3). https://arcmira.com/docs/changelog.md

## What we couldn't check

What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it.

- Kind is `mcp`. The docs quickstart, home page and docs index lead with connecting an agent to mcp.arcmira.com, the registry entry and the repository are the server and /mcp is the first product page, so the server is graded as the main surface and the HTTP API, CLI and SDKs are described alongside it. The API reading of each checklist line would move little, since the limits, errors and idempotency rules are the API's.
- unchecked: the live tools/list, the server's protocol behaviour and the size of the arcmira_describe reference in tokens. No MCP session was opened; the four tools, their annotations and schemas are read from src/tools.ts at the 30 October master, which package.json says is 0.10.7, and from the server card.
- unchecked: whether an API sign-up or OAuth flow works as documented. No account was created, no code requested and no client registered. The one unauthenticated GET to /mcp is the 30 October listing check's.
- unchecked: the PyPI page for arcmira, which served a bot check. The Python SDK's version and licence are read from github.com/arcmira/python (v0.6.2, Apache-2.0).
- unchecked: whether a status page exists under another name. None is linked from any page read, https://arcmira.com/status answers 403 and status.arcmira.com does not answer, so it is scored absent.
- unchecked: the hosted server's deployed version. The registry's latest is 0.10.7 of 12 October; the repository's master has later commits (docs, listing proofs, plugin links) without a version bump, and the CHANGELOG says a 0.10.8 change was reverted so the server stays on 0.10.7.
- firstReleased is 2026-08-22, the first npm publication of arcmira and the start of the SDK repository's history. The vendor's changelog dates the first public API to April 2026 and a v1.1.0 to May 2026 without days, so the API is older than that date; the MCP server's first registry version is 4 September 2026. The batch file suggested the registry date; the npm date is earlier and first-hand.
- The vendor's index size, 'search engine for the spoken web' and 'SF-based' are the vendor's claims; the terms name Arcmira, Inc. in San Francisco. The 'Used by people at' logos on the pricing page (OpenAI, PlanetScale, TBPN, T3, Convex, Rainmaker) are unverified claims.
- The MCP registry's 0.7.x descriptions mentioned code mode; the repository's 0.7.0 entry (29 September 2026) describes it as ten tools replaced by describe and execute over a typed client, with the vendor's own measurement of 1,123 tokens of definitions against 8,163 and 59 of 60 tasks against 56. Those are the vendor's figures.
- The sign-up route sends a six-digit code by email, so autonomous onboarding is scored 15 of 20 rather than 20. A mailbox-equipped agent can complete it; one without cannot.
- The 8 October v1 changes are deducted -3 as a breaking change with same-day notice. The vendor's position that the API is v1 without a versioning or deprecation policy is recorded in transparency; a published notice period would lift both.
- GitHub shows the arcmira/mcp repository created on 12 October 2026 with commit history from 28 September and 23 of its 24 releases published that day and 0.10.7 on 15 October, so the earlier registry versions were published from a repository that wasn't public. Recorded, not deducted.
- Arcmira doesn't do LocalGhost's or letme's job. It returns transcript passages, mentions and sponsor data for an agent to cite; it neither hides an agent's identity nor chooses tools for a task.

## Weaknesses

- No status page, incident history or SLA was found; the terms disclaim uninterrupted service and Enterprise lists priority support without an uptime commitment
- No security.txt on any of the three hosts, no disclosure policy, bug bounty or certification, and the terms forbid probing for vulnerabilities
- The v1 API retired `POST /v1/trackers` and six tracker fields between 5 and 8 October 2026 with the changelog entry dated the day the change finished; the quickstart says routes arrive without notice
- The MCP server is at 0.10.7 after 30 releases in six weeks and two rounds of tool renames; the Free plan withholds the newest 30 days of media and pages past the first
- Every tool call, including the first 4,000 characters of the program and the agent's intent, is logged to the vendor's analytics with no opt-out, and the privacy policy states no retention periods
- Results are third-party speech with no prompt-injection guidance beyond checking the linked source, and support is one email address with no community channel

## What costs an agent a turn today

The notes we give agents before they call it. Each one is a workaround an agent shouldn't need.

- Call `arcmira_describe` once per session and write one program per question for `arcmira_execute_read`; return only the fields the answer needs, since output is cut at 12,000 characters and 3,000 per string.
- Resolve names with `arcmira.resolve` (free) before any `about`, `by`, `entity_ids` or `channel` filter; a name in a filter fails with `id_required`. Plain keyword search needs no resolver call.
- Check `arcmira.status()` or GET /v1/me for the plan, scopes and `usage.credits.available` before metered reads; on Free, media newer than 30 days and pages past the first fail with `freshness_requires_paid` and `pagination_gated`.
- A Premium read of an untranscribed video starts a transcription that uses credits (300 per started 15 minutes) and may return `pending`; read the same video again later, which is never charged twice, and send `retry: true` only when the user asks.
- Over HTTP send `Authorization: Bearer arc_sk_...`, an `Idempotency-Key` on every POST and PATCH, honour `Retry-After` on 429, and switch on `error.code`, following `unlock.url` on 402 and 403 rather than re-authenticating.
- Use `arcmira_execute_write` only after listing the user's monitors; it creates, follows and pauses but never deletes, and a connection made before server 0.9.0 answers `insufficient_scope` until reconnected.

## When it's done

Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.