QwenPaw

# Fix list: QwenPaw

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/qwenpaw, the October 2026 research run, assessed 12 October 2026. Grade B, 62.2 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 QwenPaw: 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. Security & auth, 60 out of 100, up to 7 more on the total

Why it scored 60: Harness reading of the framework checklist, as for Goose and OpenHands. 30 for what leaves the machine by default, 20 for approvals and sandboxing, 15 for prompt-injection posture, 15 for audit and 20 for the security programme. Usage data is on by default. `qwenpaw init --defaults`, the documented quick start, sends it with no prompt, `qwenpaw app` sends it on first start, and the interactive prompt defaults to yes. The fields are system details and a random id. Conversation content goes only to the model provider the owner picks, and local models are supported (17). Four approval levels, File Guard, a kernel sandbox on three systems and per-tool MCP policies, with mail tools on `ask` by default. The default level AUTO asks only for flagged calls, `/proactive` skips prompts, channel policies default to `open`, Console login is off by default, and containers run with no isolation (13). SECURITY.md treats the model as untrusted, the mail docs treat message bodies as untrusted input, and a skill scanner runs in warn mode by default. All seven shell-evasion checks default to off (10). Approval cards, cron execution records, Inbox traces of automatic mail handling and a governance log (11). SECURITY.md with private reporting through the Alibaba Security Response Centre, a written trust model and a CodeQL workflow. No bounty, no security.txt, and no published advisory although security fixes ship in release notes (9).

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.

## 2. Reliability, 70 out of 100, up to 6 more on the total

Why it scored 70: Read with the local-software lines, as the Goose, OpenHands and Khoj dossiers are. `qwenpaw` is on PyPI for Python 3.11 to 3.13, with an install script, a Docker Hub image and a desktop app the README labels beta (20). The repository has 878 test files and Tests, E2E smoke, frontend, pre-commit and CodeQL workflows. On the head commit of 12 October the E2E smoke, frontend, pre-commit and CodeQL runs passed, but the 12 most recent Tests runs on main, back to 30 September, all showed `waiting` and none had completed (15 of 25). 671 open issues, 130 of them labelled bug, and 381 open pull requests on 12 October. New reports get a welcome-bot comment the same day, and we did not read maintainer replies (12 of 25). Dated release notes for every version under Added, Changed and Fixed, with no breaking-change section. The 2.0.0 notes list the removal of Plan Mode under Changed (8 of 15). 2.2.1 is a stable 2.x release (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.

## 3. Agent ergonomics, 67 out of 100, up to 5.4 more on the total

Why it scored 67: Harness reading of the framework line. 16 built-in tools that can be switched on or off one by one, per-tool enable lists on MCP clients, and skills loaded per agent (18 of 25). Every turn is persisted with older turns recalled on demand, and heartbeat and cron runs have timeouts. We did not find a turn or cost cap in the pages read (12 of 20). A REST API with streamed events, a CLI for cron, channels, chats and agents, and ACP for other runners (15 of 20). Sessions resume, workspaces have checkpoints, and cron keeps execution records (13 of 20). Three commands start it, but a model has to be configured first, there is no SDK, and the defaults leave approvals on AUTO and channels open (9 of 15).

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.

## 4. Schema & documentation, 68 out of 100, up to 5.2 more on the total

Why it scored 68: Harness reading, scored on what an owner or a pipeline has to read to drive it. `config.json`, `agent.json` and the MCP driver YAML have field tables with types and defaults, and the REST API has a tutorial page. We found no published OpenAPI file (14 of 25). No llms.txt (404). The docs are Markdown files in the repository, and the site draws them by script (3 of 10). The security pages say what each guard does and list what each sandbox backend does not enforce (16 of 20). Policies, approval levels and scanner modes are enumerated, with defaults stated (11 of 15). Worked examples for channels, cron, mail and the API, an FAQ and an API error section (11 of 15). Release notes for every version on the site and on GitHub releases (13 of 15).

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, 60 out of 100, up to 5 more on the total

Why it scored 60: Read with the self-hosted rule. No payment protocol (0). Free under Apache-2.0 with nothing to buy, so 20, 20 and 20 on the last three lines. It installs from PyPI or Docker Hub with no account and can run on a local model with no key. The README describes a hosted deployment on AgentScope Platform as free. Its pages are script-drawn and were not read, so it is not scored.

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. Transparency & trust, 56 out of 100, up to 3.9 more on the total

Made of editorial 58, provenance 53.

Why it scored 56: Apache-2.0 (30). The site publishes a Usage Policy of 26 August 2026 and no privacy policy or terms. The README says data stays on the owner's machine and lists the usage fields collected, with no retention period or recipient named. It says data is sent once a version during `qwenpaw init`, while the source also sends at `qwenpaw app` start and, on main since 17 September, once per active day (10 of 30). Release notes record removals such as Plan Mode in 2.0.0 and the mailbox docs name the `himalaya` skill as deprecated. No deprecation policy or dated notice period found (7 of 20). Usage data is disclosed in the README with its fields and an interactive opt-out. It is on by default, and `QWENPAW_TELEMETRY_DISABLED` appears only in the init prompt and the source, not in the README or docs (11 of 20).

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):

- Domain age: agentscope.io, no registry record we could read (0 of 15)
- Status page: not found (0 of 10)
- security.txt: not found (0 of 10)

## 7. Maintenance & community, 73 out of 100, up to 2.4 more on the total

Why it scored 73: 2.2.2-beta.4 was tagged and published as a GitHub pre-release on 30 September 2026, 12 days before this check. The last stable release, 2.2.1, is from 11 September, 31 days before. Recency is scored on the last stable release (20). Four stable releases in the 90 days to 12 October (2.0.1, 2.1.0, 2.2.0, 2.2.1) and more than 20 tagged pre-releases (20). Commits land on main daily and new issues get a bot reply the same day, against 671 open issues and 381 open pull requests. We did not sample maintainer reply times (14 of 25). PyPI, Docker Hub, an Alibaba Cloud registry image and desktop builds ship from the same tags (12 of 15). Dependencies carry version bounds with reasons in comments, and CodeQL, pre-commit and nightly test workflows exist, while the Tests workflow on main has not completed since 30 September (7 of 10).

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.

- 2026-09-11. Version 2.2.1 began enforcing each MCP server's tool allowlist on the agent runtime path (commit f2a292d of 4 September, #7504). Before it, a tool the owner had disabled on an MCP client could still be invoked by the agent. Fixed and listed in the release notes, with no security advisory, -2. https://github.com/agentscope-ai/QwenPaw/pull/7504
- 2026-09-03. Version 2.2.0 blocked shell line-continuation variants that bypassed sensitive-path protections (#7472). Fixed and listed in the release notes, with no security advisory, -1. https://github.com/agentscope-ai/QwenPaw/blob/main/website/public/release-notes/v2.2.0.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. We used `harness`, the closest kind the briefs define for a self-hosted agent program, with `metrics.kind` local. Reliability is read with the local-software lines and Payments with the self-hosted rule
- Maintenance recency is scored on the tagged pre-release 2.2.2-beta.4 of 30 September 2026. A strict reading on the last stable release, 2.2.1 of 11 September, 31 days before the check, gives 20 on that line and 73 overall
- The lead names the vendor as AgentScope (Alibaba). First-hand we found the AgentScope team credited in the README, Qwenpaw PRIVATE LIMITED in the site footer and the Alibaba Security Response Centre in SECURITY.md, and no direct statement of ownership
- Unchecked: the rendered pages of qwenpaw.agentscope.io, which are script-drawn. Docs, release notes, the Usage Policy and the footer were read from the site's source in the repository
- Unchecked: AgentScope Platform's terms, privacy policy, limits and what its free hosted deployment stores (platform.agentscope.io is script-drawn)
- Unchecked: whether the Tests workflow passes on main. Its 12 most recent runs showed `waiting`
- Unchecked: maintainer reply times on issues and pull requests, PyPI download counts and the first release date
- Unchecked: whether channel bot tokens in `agent.json` are encrypted. The encrypted store covers model keys, environment variables, mail and MCP credentials in the source we read
- Unchecked: what the telemetry endpoint logs. The README says no IP addresses are collected, and the endpoint is an HTTPS function on Alibaba Cloud
- The daily usage event is on main and in the 2.2.2 pre-releases and is absent from the README's Telemetry section. It was not in the stable 2.2.1 release, so no deduction was taken
- No calendar connector was found in the docs or the built-in skills, though the default heartbeat checklist mentions checking a calendar
- No documented way to delete a single memory other than editing or removing its Markdown file, and nothing on whether index data in `mem_metadata/` is purged when a file is removed

## Weaknesses

- Usage data is on by default, `qwenpaw init --defaults` accepts it without a prompt, and the opt-out variable is absent from the README and docs
- The default approval level AUTO asks only for tool calls a guard rule flags, and `/proactive` skips approval prompts entirely
- Channel direct-message and group policies default to `open`, and Console login is off by default
- No privacy policy, terms of service or security.txt on qwenpaw.agentscope.io, and no published GitHub security advisory
- Inside containers the sandbox falls back to `mode=none`, which the docs say enforces no isolation
- 671 open issues and 381 open pull requests on 12 October 2026, and the Tests workflow on main had not completed a run since 30 September

## 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.

- Set `QWENPAW_TELEMETRY_DISABLED=true` before `qwenpaw init --defaults` or `qwenpaw app`. Both send usage data without asking
- Set `approval_level` to `STRICT` or `SMART` in `agent.json`. The default `AUTO` runs every tool call no guard rule flags
- Set `dm_policy` and `group_policy` to `allowlist` with `allow_from` on each channel. Both default to `open`
- Set `QWENPAW_AUTH_ENABLED=true` before binding beyond 127.0.0.1. Requests from localhost skip login even then
- Keep mail tools on `ask` for `send_message`, `delete_message` and `delete_thread`, and turn on Mail Access Control with automatic handling

## 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.