HarnessRouter

# Fix list: HarnessRouter

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/harnessrouter, the October 2026 research run, assessed 17 October 2026. Grade D, 53.2 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 HarnessRouter: 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, 35 out of 100, up to 13 more on the total

Why it scored 35: Hosted reading, for HarnessRouter Cloud at api.harnessrouter.ai. No status page (status.harnessrouter.ai doesn't resolve and harnessrouter.ai/status answers 404), so 0 for the page and 5 for an unreadable incident record. No request rate limit with numbers was found. The pricing page states one task at a time on Free, unlimited concurrency on plans with brief queueing and a 30-minute or 1-hour task limit on Free, so 5 of 15. 429 is in the error table as rate_limited with Retry-After (UHP Errors §3), quota_exhausted is marked not worth retrying, and every task call takes an Idempotency-Key with a rule to recover the session rather than resend after a dropped stream (15). Terms §5.4 give no service level outside a signed order form (0). The task API isn't labelled beta, while tracing and agent memory are, and the protocol is a draft standard (10). 35 in all.

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. Payments & pricing, 25 out of 100, up to 9.4 more on the total

Why it scored 25: Published rubric, for the hosted service. No machine payment protocol (0). Per-unit prices without a login, agent work per active minute and model calls at listed provider prices per million tokens (20). The Free plan needs no card, but it only configures agents, and running a task draws on card-funded credit, so 5 of 20. A person signs up in a browser and creates the key in the console (0). 25 in all. Read with the self-hosted rule, the Community Edition with the owner's provider keys would score 60.

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.

## 3. Security & auth, 58 out of 100, up to 7.4 more on the total

Why it scored 58: Hosted reading. Bearer Workspace API keys, revocable at once through the API, held to one workspace or reaching the organisation, with harness-bound calibration credentials. No OAuth, per-endpoint scopes or expiry found (20 of 30). Least privilege 10 of 20. disabled_tools per configured agent, enforced by the runtime only where GET /v1/bases says hard, and per-turn cost and step caps, with no approval step for actions a harness takes. Prompt injection 10 of 15. UHP Security §4 puts injection in scope for the client, tells servers to keep harnesses that read untrusted input separate and forbids claiming an agent is safe on untrusted input, and secret values are redacted from responses, streams and traces. Operator visibility 10 of 15. A cost breakdown per task on every plan, and the step-by-step trace is beta, included in Scale or $50 a month. Security programme 8 of 20. A disclosure policy with safe harbour and no bounty, SECURITY.md with GitHub private reporting and a 10-working-day acknowledgement aim, GHSA identifiers cited in public fix commits, no security.txt and no SOC 2 or ISO 27001 found. 58 in all.

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. Agent ergonomics, 73 out of 100, up to 4.4 more on the total

Why it scored 73: API reading. Sessions, turns and files are separate reads, and list endpoints take limit and cursor, but there's no field selection on a response (15 of 25). Cursor paging on session and memory lists with a bounded limit, and few filters found (14 of 20). One error envelope with type, code, param and detail, a closed code set, a retry rule per code and a 422 that lists every failed field without echoing the body (18 of 20). Idempotency-Key on task creation, reuse only for the same request, recover from the session after a drop, and an explicit cancel because closing the stream doesn't stop work (18 of 20). Only input is required to run a task, and budgets default to 400 steps, 7,200 seconds and $10. No official SDK, only the OpenAI Responses shape, clients generated from the OpenAPI file and an Agent Skill (8 of 15). 73 in all.

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.

## 5. Schema & documentation, 83 out of 100, up to 2.8 more on the total

Why it scored 83: API reading. UHP's OpenAPI 3.1 file (61 paths) and JSON Schema 2020-12 are public and normative for structure. HarnessRouter's own path prefix /{harness_id}/v1/responses and GET /v1/bases aren't in them, and the hosted API publishes no OpenAPI of its own (/openapi.json answers 404), so 22 of 25. llms.txt, a 230 KB llms-full.txt, agents.md, docs.md, pricing.md and harnesses.md (10). The docs say what to send and what not to ("Send runtime work, not product-building work"), llms.txt says what the product isn't, and the schema carries 352 descriptions, but the docs pages are short (16 of 20). 26 enums, 134 required lists, bounded limits such as limit 1 to 100, with metadata left open by design (12 of 15). A request example on each docs page, a 422 example naming every failed field, and a table of status, type and code, but only 14 examples in the OpenAPI file (13 of 15). Date-based versions with a changelog and a versioning policy. Read strictly, seven same-day patches changed 2026-10-04's schema while the schema page calls a published version immutable, the hosted API still serves 2026-09-28, and the conformance page says 76 checks where the suite defines 98 (10 of 15). 83 in all.

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. Maintenance & community, 72 out of 100, up to 2.5 more on the total

Why it scored 72: Open-source reading of the repository behind the hosted service. Last tag v0.34.2 on 16 October 2026 (30). 202 tags between 15 August and 16 October (20). GitHub's issue pages and API refused this session, so open issues and reply times are unchecked. Pull requests from at least ten outside authors were merged in the 30 days to 17 October, by the git history (15 of 25). No official SDK and no MCP server to register (0). A tests workflow runs every suite on each pull request and push to main, with a release workflow and a Docker image, but whether the default branch passes is unchecked (7 of 10). 72 in all.

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 78, provenance 65.

Why it scored 72: Editorial half. The server, specification, schema and conformance suite are Apache-2.0. The Cloud's own billing and deployment wasn't confirmed to match the open code, and the Agent Skill has a custom licence for some guides (25 of 30). The privacy policy, DPA, subprocessor list and a retention page agree. Customer content isn't used for training without an opt-in, and periods are stated (tasks deleted at once, agents and files after 30 days, backups within 14 days, credentials within 90, free-plan task records 7 days). The privacy policy gives no periods of its own (24 of 30). UHP's versioning page says how a field is deprecated and that the previous version SHOULD be supported for six months. No dated product deprecation notices found (14 of 20). Subprocessors listed with purpose and location (Azure, Stripe, Vercel, Supabase, all United States), but the routing and inference gateway provider is named only on request (15 of 20). 78 in all.

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: harnessrouter.ai, registered 2026-06-18 (under a year) (0 of 15)
- Status page: not found (0 of 10)
- security.txt: not found (0 of 10)

## Deductions

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

- 2026-10-06, -5. Privately reported flaws fixed in public commits #416 to #429 on 6 and 7 October 2026. The gateway source names them GHSA-p6cq-54cg-8mpv (a bearer token that failed verification still reached internal routes, so anyone could rewrite any organisation's connections and routing; "found live on the hosted service and closed there the same way, 2026-10-06"), GHSA-qrr8-m92f-7vfp (identity headers trusted by default), GHSA-mw9j-m5jf-5r56 and GHSA-c5g7-c47h-f6c3 (server-side request forgery through git imports and upload addresses) and GHSA-rmh9-whjx-gfrc (secrets with quotes or non-ASCII characters left unredacted). A workspace key could also read, change and delete another workspace's harnesses, sessions and responses and mint an organisation-wide key ("reported privately three times"). All were fixed within days and described in the source. The advisory pages couldn't be read (GitHub answered 403), so whether they're published is unchecked, and the deduction is lower than an open cross-tenant flaw would get. Source https://github.com/HarnessRouter/harnessrouter/blob/main/gateway/app.py

## 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.

- unchecked: GitHub issues, pull request reply times and the published state of the GHSA advisories. GitHub's web pages and API answered 403 to this session
- unchecked: whether the tests workflow passes on the default branch
- unchecked: Qoder's own pages on its UHP support. The claim is HarnessRouter's, and the repository has a Qoder driver
- unchecked: what implementation version 0.3.0 in the hosted discovery document means, given repository tags at v0.34.2 and a 23 August commit that stopped hard-coding 0.3.0
- No request rate limit was found for the hosted API
- firstReleased is 2026-08-11, the first published UHP version, which its changelog says was taken from an implementation already in production. The first repository tag is v0.4.2 on 15 August 2026 and the history starts on 9 August
- The scout reported 98 checks on the spec site. The conformance page read today says 76, and 98 is the count in the suite's source
- api.harnessrouter.ai was fetched before its robots.txt was read. That robots.txt answers 404, so nothing was disallowed

## Weaknesses

- No status page found (status.harnessrouter.ai doesn't resolve and /status answers 404), and the terms give no service level without a signed order form
- No request rate limit is published. Only concurrency (one task at a time on Free) and task length (30 minutes, or 1 hour after a top-up) are stated
- Running a task on the hosted service draws on credit, and top-ups and plans are card-only, so an agent can't start without a person
- Privately reported access-control, SSRF and redaction flaws were fixed on 6 and 7 October 2026, one of them found live on the hosted service
- The hosted API serves UHP 2026-09-28 while the spec site's current version is 2026-10-04, and the conformance page counts 76 checks against 98 in the suite

## 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.

- Name the harness yourself. The task call is POST /{harness_id}/v1/responses, and HarnessRouter doesn't choose a harness for you
- Send a fresh Idempotency-Key with each new task, reuse it only for a retry of the same request, and recover from the session if a stream drops
- Call cancel to stop work. Closing the SSE connection leaves the agent running and billing
- Keep the Workspace API key on the server. The docs forbid calling the API from a browser with it
- Check each tool's enforcement in GET /v1/bases. On bases marked instruction, disabled_tools is a request to the model, not a block

## What the review panel asked for

- a new date for each spec change
- a product changelog
- Publish an OpenAPI for the hosted API
- Reconcile the conformance check counts
- Publish a status page
- State request limits for hosted plans
- scoped and expiring keys
- runtime-enforced tool blocks

## 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.