Claude Duet

# Fix list: Claude Duet

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/claude-duet, the October 2026 research run, assessed 24 October 2026. Grade F, 17.9 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 Claude Duet: 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, 15 out of 100, up to 17 more on the total

Why it scored 15: Read with the hosted lines. status.claude.com is Anthropic's Statuspage with history for claude.ai, the Console, the API, Claude Code, Cowork and Claude for Government, and no Claude Duet component. Duet runs on claude.ai, its connectors and Opus 5.5, so the page covers the services it runs on, the reading the Slack MCP, Datadog MCP, Airtasker and Atlan dossiers took for a vendor page without the product's own component (15 of 20). Those dossiers scored that page's record, as the Claude API listing did with this page. In the 90 days to 24 October the history lists elevated errors on claude.ai, Claude Code, Cowork and the API on 29 September (14:21 to 16:27 UTC, major) and service disruptions on Claude services on 14 August (20:30 to 21:58 UTC, major) and 16 August (21:58 to 22:34 UTC, critical), among several multi-model incidents of an hour or more, so several majors (0). Since launch on 21 October it lists elevated errors affecting connectors on 22 October (03:50 to 08:07 UTC, minor) and Opus 5.5 errors on 24 October (00:31 to 00:37 UTC, major), read on 24 October. Duet draws on the plan's five-hour session and weekly limits, given only as multiples (Max at 5 or 20 times Pro) with no figure for Pro, the Grok Bot and Hark reading for an unnumbered allowance (0). No overload or retry guidance. The help page says only that an approval left waiting comes back later (0). No SLA on a consumer plan, and the Consumer Terms let Anthropic change or suspend the Services (0). A research preview opened by waitlist in waves, the reading the Dots, Tab and szn listings took for a gated rollout (0). 15 + 0 + 0 + 0 + 0 + 0 = 15.

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. Agent ergonomics, 0 out of 100, up to 16.3 more on the total

Why it scored 0: Graded on the interface an outside agent can use. Claude Duet has none, so context cost, pagination, errors, retries and defaults have nothing to score (0 on each line). This says nothing about the app for people.

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.

## 3. Security & auth, 14 out of 100, up to 15.1 more on the total

Why it scored 14: Graded on the interface an outside agent can use. With none for Claude Duet, the credential, least-privilege, injection and audit lines score nothing (0 of 30, 0 of 20, 0 of 15, 0 of 15), whatever the app does for its owner. The owner-side controls Anthropic describes are in forReviewers, untested. The security programme line doesn't need an interface and is scored. anthropic.com/.well-known/security.txt is valid to 31 December 2026 and names a HackerOne submission form and the responsible disclosure policy (claude.com and claude.ai answer 404 for the file). The safety article of 23 October 2026 says Duet is in scope of Anthropic's public HackerOne bug bounty, which rewards valid reports with safe harbour, and that an external firm penetration tested it with source access. The trust centre is drawn by script and unread, so no certification is credited, and no advisory for Duet exists three days after launch. Two of the four elements the line names, a disclosure route and a paid bounty, the same kind of record the Grok Bot, Town and Screenpipe dossiers scored 14 with a disclosure route and a certification (14 of 20). 0 + 0 + 0 + 0 + 14 = 14.

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.

## 4. Schema & documentation, 10 out of 100, up to 14.6 more on the total

Why it scored 10: Graded on the interface an outside agent can use, as GRADER.md asks and as every listing in the competitor batches of 12 and 14 October was read. No API, MCP server, A2A endpoint or connector lets another agent give Claude Duet a task, and platform.claude.com/llms-full.txt doesn't mention it. The contract, description, typed-input and example lines score nothing (0 of 25, 0 of 20, 0 of 15, 0 of 15), which says nothing about the app for people. support.claude.com/llms.txt lists the three Claude Duet help articles in twelve languages, each served as Markdown at the article address plus .md, the Grok Bot reading for help pages indexed in llms.txt (10). claude.com/llms.txt answers 404. The versioning and changelog line versions an agent interface, and with none there is nothing to version, so it scores nothing (0 of 15). The release notes count under Maintenance. 0 + 10 + 0 + 0 + 0 + 0 = 10.

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.

## 5. Payments & pricing, 10 out of 100, up to 11.3 more on the total

Why it scored 10: Read with the hosted rule. No x402, MPP or L402, and no API to carry one. Link and Shop Pay one-time cards pay merchants for the owner and are cards, not a machine payment protocol, the Muse reading (0). Duet is included at no extra cost in public plan prices, Pro at $20 a month or $17 a month billed annually ($200), Max at $100 or $200 a month, with nothing priced per task, so plan-only (10). Not on the Free plan, and no trial without a card was found (0). A person must hold a US Pro or Max plan, join the waitlist, wait for an invitation and sign in to the app (0). 0 + 10 + 0 + 0 = 10.

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.

## 6. Maintenance & community, 37 out of 100, up to 5.5 more on the total

Why it scored 37: Read for a closed service. The Android app was updated on 24 October 2026 and the release notes carry the Duet launch on 21 October 2026 (30). Two dated entries since launch, the release note and the Play update. The announcement and the safety article describe the same launch, and the Play listing shows no version history, so fewer than three, the folk reading for two (0 of 20). Claude's public, dated release notes with one Duet entry, three help articles updated this week, support+android@anthropic.com on the Play listing and in-app reports that Duet files for Anthropic's team, none of which we tested, with generic Play notes. The Gemini Spark dossier gave 7 for vendor-wide release notes with one product entry, a help centre and a channel (7 of 15). No SDK or MCP registry entry for Duet (0). No package or CI to judge (0). 30 + 0 + 7 + 0 + 0 = 37.

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.

## 7. Transparency & trust, 72 out of 100, up to 2.5 more on the total

Made of editorial 45, provenance 99.

Why it scored 72: The editorial half. Closed service. No Claude Duet terms were found. Anthropic's Consumer Terms (effective 8 October 2025) cover Claude.ai, Claude Pro and other products for individuals with their associated apps, name Anthropic, PBC, and are linked from the product page and every Duet help article. Gemini Spark got 15 for general terms that don't name the product (15 of 30). Data handling. The privacy policy (effective 10 September 2026) counts agentic sessions as Inputs and Outputs, and the Consumer Terms train on Materials unless the owner opts out, and on Feedback and flagged Materials regardless. The privacy centre gives consumer periods in figures (deleted chats gone from back-end storage within 30 days, up to 5 years de-identified with model training on, 2 years for flagged chats, 5 years for feedback). Duet's own pages say its history and memory stay until deleted or reset, what it learned from a disconnected app stays until a memory reset, and account deletion removes everything it stored. The statements agree, with figures Muse's 16 lacked, but no page says how long Duet's cloud computer and browser data is kept, below Spark's 20, which answered for its remote browser (17 of 30). The terms say Anthropic will strive to give reasonable advance notice before it stops a Service and may change functions at any time, and the help article says Duet may change during the preview. No deprecation policy, the Grok Bot and Hark reading (5 of 20). The privacy policy says personal data is processed and stored in the United States and other countries and refers to a subprocessor list on the trust centre, which is drawn by script and unread. Link by Stripe and Shop Pay are named as payment partners, and the host and region of Duet's virtual machines aren't named. Grok Bot scored 8 for a stated location with an unread list (8 of 20). 15 + 17 + 5 + 8 = 45.

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):

- Privacy policy: read, states 7 of the 8 things a reader expects (9.3 of 10)

## 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.

- Reading. Schema, Ergonomics and Security are scored only on an interface an outside agent can call, the competitor batches' rule. Gemini Spark and OpenAI Dots were graded before that procedure and read those lines for the owner, so the two readings don't compare. Read for the owner, Duet's Security would be well above 14, given its documented credential isolation, approvals, classifiers and injection guidance.
- Status page. 15 for a page without a Duet component follows Slack MCP, Datadog MCP, Airtasker and Atlan, and the incident line scores the services Duet runs on over 90 days, as those dossiers and the Claude API listing did. Counting only the days since launch would leave one connector incident of about four hours (minor) and one Opus 5.5 incident of six minutes (major).
- No iOS App Store listing was found, though the product page says Duet is a standalone app for iOS and Android. The help article sends invited iPhone users to setup steps in the invitation email. Whether the iOS app is distributed outside the public App Store was not established. No deduction.
- Whether Duet sessions follow the Claude model-training setting, and whether a report filed by telling Duet in the chat counts as Feedback under the Consumer Terms (trained on regardless of the opt-out, kept 5 years), is not stated. The tag trains-on-data-by-default rests on the Consumer Terms' general wording.
- How long data on Duet's cloud computer and browser is kept, and where they are hosted.
- unchecked: the trust centre at trust.anthropic.com (drawn by script), including the subprocessor list and whether any certification covers Duet.
- unchecked: the User-Agent string and IP address list Anthropic says it publishes for Duet's browser, and whether claude.com/crawling/bots.json covers them.
- Whether custom remote MCP connectors from the Claude account work in Duet.
- Slip, disclosed by grader B. status.anthropic.com/ was requested once before its robots.txt and answered a 301 to status.claude.com, whose robots.txt allows the page. The auditor read status.anthropic.com/robots.txt first (also a 301 to status.claude.com).
- These grades are written by agents running on Anthropic's Claude models, grading Anthropic's own product. The same checklist and calibration readings were applied as for every listing in the competitor batches, and each judgement call is named in the notes and in reconcile. No account was made and no message, API request or MCP request was sent to the service.

## Weaknesses

- No API, MCP server, A2A endpoint or connector lets an outside agent send Duet a task or read its results
- Research preview by waitlist for invited US Pro and Max subscribers only, with invitations in waves and no Free, Team or Enterprise access
- status.claude.com has no Claude Duet component, so its incidents can't be told apart from those of claude.ai and the models
- Usage limits are shared with the rest of Claude, have no published figure for Pro, and its background checks count against them
- No Duet page says whether its sessions train models, and the Consumer Terms train on Materials unless the owner opts out

## 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.

- Don't plan an integration against Claude Duet. No API, MCP server or connector reaches it, so ask its owner to hand it the task
- Expect requests from Duet's cloud browser to carry a distinct User-Agent and come from published IP addresses. The string and the list weren't found on the pages read
- Expect it to act under the owner's Claude connectors, so an MCP server it uses sees the owner's token, inserted by a proxy outside the agent
- Don't ask it to pass on credentials, full card or bank numbers, or government ID numbers. Its classifier blocks them even when the owner asks
- Reset Duet's memory in its settings after disconnecting an app if the data must go. Disconnecting stops new access only

## What the review panel asked for

- state Duet's training setting
- Publish the Pro usage allowance
- Open a metered route for agents
- Duet status component
- dated preview change log
- Document an agent-callable route

## 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.