# Fix list: CowAgent From Anchor Terminal's listing at https://www.anchorterminal.com/tools/cowagent, the October 2026 research run, assessed 12 October 2026. Grade C, 58.5 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 CowAgent: 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, 43 out of 100, up to 10 more on the total Why it scored 43: The console has no password by default and listens on 127.0.0.1. `web_password` adds one shared password, and `external_api_token` is a single Bearer token with no scopes. Model and MCP keys are plain text in `config.json` and `mcp.json`, while `~/.cow/.env` is closed to the `bash`, `read`, `edit` and `send` tools, with tests for the last three. No setting limits which Telegram, Slack or Discord users may message the bot (13 of 30). Three permission modes per session, with full-access the default for source and Docker installs and workspace-write for a fresh desktop install. The source says the checks are argument-level and not a sandbox. No action asks for approval, and the `bash` blocklist covers root deletion, disk wiping and power commands only (9 of 20). `web_fetch`, vision and browser navigation validate addresses against internal targets, and the agent may edit `mcp.json` only in full-access, but no guidance on injected content was found and issue #2998 is open (6 of 15). Tool calls are shown in the console, written to the session database after each step and returned as `tool_trace`, and channel credentials are masked in debug logs, with no separate audit log (9 of 15). GitHub reports no security policy, there is no security.txt, bounty or certification, one advisory is published, and release notes credit security fixes (6 of 20). 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. Agent ergonomics, 59 out of 100, up to 6.7 more on the total Why it scored 59: `agent_max_context_tokens` caps the prompt at 64,000 by default, history is compacted on a budget, a large knowledge index injects titles only, and on-demand MCP tool retrieval exists but is off by default (18 of 25). The endpoint returns the final message, with optional streaming and a tool trace, and nothing limits or selects the output (8 of 20). Documented status codes, a structured JSON 500 with a `timeout` code, error events inside a stream, and permission denials that name the reason (14 of 20). `conversation_id` keeps a stable session and 2.2.0 persists each step so an interrupted turn resumes, with no idempotency key or retry guidance (8 of 20). One command installs and starts it with a console on localhost, and any OpenAI client library reaches the endpoint, with no SDK of its own (11 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. ## 3. Schema & documentation, 67 out of 100, up to 5.4 more on the total Why it scored 67: Graded on the docs and on the interfaces an outside agent would use, the OpenAI-compatible endpoint, the CLI and the config. No OpenAPI file. The endpoint follows the chat completions shape and its additions (`conversation_id`, `tool_trace`, `cow_event`) are documented in prose (10 of 25). docs.cowagent.ai serves an llms.txt and a Markdown copy of each page (10). Each built-in tool has a page stating its purpose and uses, and the permission and MCP pages say when a mode refuses an action (13 of 20). Tool pages list parameters with types and required flags, and config tables give defaults and allowed values (10 of 15). The web channel page has curl and Python examples and documents 400, 401, 500 and 503 answers and the streaming error event (11 of 15). Release notes for every version on the docs site and on GitHub. The endpoint carries only a `/v1` prefix (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. ## 4. 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 in the docs or the source (0). Free under the MIT licence with nothing to buy, so 20, 20 and 20 on the last three lines. The README names LinkAI as a seller of managed hosting for CowAgent. link-ai.tech did not answer on 12 October 2026, so that service and its prices are unread and 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. ## 5. Transparency & trust, 45 out of 100, up to 4.8 more on the total Made of editorial 62, provenance 27. Why it scored 45: MIT (30). cowagent.ai links no privacy policy and no terms. The docs say where the desktop client, workspace and credentials are stored and warn that backups hold keys and conversations unencrypted, and a source comment says nothing is uploaded by default. Nothing states what the model providers receive or how to erase memory (12 of 30). The rename from chatgpt-on-wechat is announced with a redirect, and release notes are dated, with no deprecation policy (8 of 20). No telemetry was found in the Python source and the cloud client loads only when two settings are present. That is stated in a source comment and not in the docs, and the installer is fetched from LinkAI's CDN (12 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): - Legal entity named: not found (0 of 20) - Domain age: cowagent.ai, no registry record we could read (0 of 15) - Status page: not found (0 of 10) - security.txt: not found (0 of 10) ## 6. Reliability, 84 out of 100, up to 3.2 more on the total Why it scored 84: Read with the local-software lines. Installs by a script and a Compose file served from cdn.link-ai.tech, a Docker image and a desktop client, with Python 3.7 to 3.13 stated in the quick start. There is no versioned package on a registry that we checked, and `pyproject.toml` describes only the CLI at version 1.0.0 (15 of 20). The Tests workflow runs pytest on Python 3.11 and 3.13 for every push to master, and the 15 newest master runs we saw on 12 October 2026 passed (25). 12 open issues and one open pull request, most of them feature requests. Issue #2998, a security report of 31 July, is still open (21 of 25). Dated notes for every release, with behaviour changes described inline and a setting to restore the old behaviour, but no breaking-change section, and patch-level versions carry new products such as the desktop client in 2.1.3 (8 of 15). Version 2.2.0 (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. ## 7. Maintenance & community, 83 out of 100, up to 1.5 more on the total Why it scored 83: Version 2.2.0 was tagged on 30 September 2026, 12 days before this check (30). Seven tagged releases in the 90 days to 12 October (20). The maintainer merged pull requests on 12 October, the tracker holds 12 open issues and one open pull request, and the security report in #2998 has waited ten weeks with its proposed fix unmerged (20 of 25). Not an MCP server and no SDK. The Docker image, CLI and desktop client are current (7 of 15). Tests pass on master, requirements are mostly lower bounds with no lock file, and the image is built on `python:3.10-slim-bullseye` (6 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. - 2026-07-31. Issue #2998 reports that a member of a Feishu group asked the bot for local files and it sent the owner's documents into the group. This is the reporter's account and we did not reproduce it. We confirmed in the master source on 12 October 2026 that group channels accept every group, that full-access is the default and that the proposed sender gating is absent. Open and unfixed, -3. https://github.com/zhayujie/CowAgent/issues/2998 - 2026-07-29. Version 2.1.5 fixed a chain in which injected content led the agent to rewrite the MCP config, run arbitrary commands and take keys, and fixed the `edit` tool reaching `~/.cow/.env`. Both are described in the release notes with no advisory. Fixed and documented, -2. https://docs.cowagent.ai/releases/v2.1.5 ## 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. Recorded as `harness`, the closest kind the briefs define for agent software the owner runs, since there is no assistant kind. The README calls it an agent harness - CowAgent is a personal assistant by the category's description, with channels, memory and unasked background work. It has no built-in mail or calendar connector, so `assistant.accounts` rests on the optional Agent Mail overlay and MCP OAuth - Unchecked: LinkAI's hosted CowAgent, its prices, terms and privacy policy. link-ai.tech answered our robots.txt request with a connection failure twice on 12 October 2026, so the host was treated as closed - Unchecked: whether the bot sends the owner's files to a group as issue #2998 reports. We read the source and did not run the software - Unchecked: the installer script, the Compose file and the desktop client served from cdn.link-ai.tech and cowagent.ai/download, and whether they match the repository - Unchecked: the desktop client's source for telemetry. The search for telemetry covered the Python source - Unchecked: Tencent's Agent Mail service at agent.qq.com, its scopes and its terms - Unchecked: the Docker Hub image tags and the registration date of cowagent.ai - Whether a legal entity stands behind the project. None is named, and the enterprise contact is an address at simple-future.tech - The advisory GHSA-qwjc-xv6c-mqfx was published 16 months after the release it names as the fix. No deduction was taken because the fix predates the last 12 months - The star count is the 47.4k shown on the repository page on 12 October 2026 ## Weaknesses - Source and Docker installs default to full-access, and no mode asks the owner before a shell command, a file send or an MCP call - The permission checks read arguments and are not an OS sandbox, as the source says. The `bash` blocklist covers only root deletion, disk wiping and power commands - Issue #2998, open since 31 July 2026, reports the bot sending the owner's local documents into a Feishu group. Its fix is not on master - No setting limits which Telegram, Slack or Discord users may message the bot, and those channels reply in every group - No SECURITY.md, security.txt, privacy policy or terms. Model keys sit in `config.json` and MCP keys in `mcp.json` as plain text ## 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 `agent_permission_mode` to `workspace-write` or `read-only` before connecting any channel. Source and Docker installs default to `full-access` - Keep the bot out of group chats that include other people. Any member who mentions it reaches the same tools as the owner - Set `web_password` before changing `web_host` to `0.0.0.0`. The password is empty by default - Set `external_api_token` to enable `POST /v1/chat/completions`, and send `conversation_id` to keep one session. The `model` field does not select a model - Run it in a container for a hard boundary. The docs and the source both say the permission modes are not isolation ## 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.