IronClaw

# Fix list: IronClaw

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/ironclaw, the October 2026 research run, assessed 12 October 2026. Grade B, 66.7 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 IronClaw: 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, 67 out of 100, up to 6.6 more on the total

Why it scored 67: Read with the local-software lines, as the Goose and OpenHands dossiers are. GitHub releases carry shell, PowerShell and MSI installers for seven targets, and source builds need Rust 1.96 and Node.js 22 (20). Push runs on `main` that we saw, dated 3 to 5 September 2026, passed Code Style, Reborn E2E and Platform & Compat and failed Tests (Reborn) and Code Coverage on all five commits. A scheduled Live Canary run failed on 1 October (10 of 25). 1,363 open issues and 183 open pull requests on 12 October. Reports filed since mid-September show no comments, among them a start-up failure (#8122) and a silent web chat status bug (#8124) (8 of 25). CHANGELOG.md follows Keep a Changelog and semantic versioning, with upgrade sections that name removed settings and the Slack re-consent in 1.3.0, and the 1.0.0 entry states there is no migration from 0.29 (14 of 15). 1.4.1 is a stable 1.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.

## 2. Security & auth, 66 out of 100, up to 6 more on the total

Why it scored 66: 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. The README states no telemetry, we found no analytics library in the 1.4.1 source, and Trace Commons is opt-in with no default endpoint. Message content goes to the chosen model provider (28). Tools carry allow, ask or deny defaults, owners can disable a tool or force a prompt each time, WASM tools run in Wasmtime with declared network targets, secrets are injected outside the sandbox, and the web interface binds to loopback. Against that, global auto-approve defaults to on and covers tools marked ask, so sends, trashes and deletes run unprompted, and only financial, approval and budget effects always ask. The published docs do not state that default (11 of 20). Channel context is framed as untrusted quoted content, a leak detector scans traffic for secret patterns, and tool output is wrapped, per the security page and release notes (12 of 15). Metadata-only audit events for approvals, run transcripts, and an operator debug inspector with sanitised tool arguments (11 of 15). No SECURITY.md, no security.txt on ironclaw.com, docs.ironclaw.com or near.ai, no published GitHub advisory and no bounty found. The 1.4.1 notes do record the Wasmtime security update (4 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.

## 3. Agent ergonomics, 69 out of 100, up to 5 more on the total

Why it scored 69: Harness reading, scored on what a caller driving it has to supply. Extensions install one at a time, but GitHub alone declares 49 tools and turn-0 tool selection is still an open proposal (#8113) (15 of 25). Context compaction, a resource governor, a cap on autonomous wakes and bounded finalisation are in the release notes. No cost cap was found in the docs (15 of 20). `ironclaw run --message` and the Responses API serve scripts, with `previous_response_id` for state, but approval and OAuth gates fail the turn over the API (14 of 20). Lease expiry recovers safe runs, an unavailable tool call is repaired, and parked approvals persist in a notification inbox (16 of 20). One binary with guided onboarding and an OpenAI-compatible endpoint, no SDK of its own, and auto-approve on by default (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. Payments & pricing, 60 out of 100, up to 5 more on the total

Why it scored 60: Read with the self-hosted rule, scoring the hosted option where one is sold. No payment protocol (0). ironclaw.com shows three monthly plans for hosted instances without a login and near.ai/pricing lists per-token inference prices (20). The software is free under MIT or Apache-2.0 and runs with a local Ollama model (20). An agent can install and run it with no sign-up (20). The Starter plan shows $5 beside $0 and the page does not say whether a card is needed.

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. Schema & documentation, 74 out of 100, up to 4.2 more on the total

Why it scored 74: Harness reading, as for Goose. Extension manifests are typed (`manifest.toml` with a JSON Schema reference per tool), and the Responses API follows the OpenAI wire format with a contract test, but the repository has no OpenAPI file (18 of 25). llms.txt with Markdown twins at docs.ironclaw.com (10). Tool descriptions say when not to use a tool (`gmail.send_message` warns against using it for the final reply), but the heartbeat, cron, shell and WASM pages of the published docs name settings that are absent from the 1.4.1 source, and the quickstart gives two different token paths (9 of 20). Permissions, effects and origin gates are enums, and network targets and scopes are declared per tool (12 of 15). The Responses page lists status codes 400, 429 and 503 with an error envelope and troubleshooting steps (10 of 15). Dated changelog in the repository and on the docs site, which says it tracks the latest stable release (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.

## 6. Maintenance & community, 74 out of 100, up to 2.3 more on the total

Why it scored 74: 1.4.1 was published on 29 September 2026, 13 days before this check (30). Six stable releases in the last 90 days, from 1.0.0 on 27 July to 1.4.1 (20). Responsiveness is weak on the evidence we saw. The `main` branch has no commit after 10 September, issue numbers moved from #8072 to #8131 in a month, and the reports filed since mid-September have no comments (8 of 25). Release binaries for seven targets are current, though we confirmed no package-manager listing (10 of 15). Dependabot runs weekly and the Wasmtime fix shipped, but Tests (Reborn) was failing on `main` when we looked (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.

## 7. Transparency & trust, 75 out of 100, up to 2.2 more on the total

Made of editorial 79, provenance 71.

Why it scored 75: MIT or Apache-2.0, with the whole runtime in one public repository (30). The NEAR AI privacy policy of 4 September 2026 covers the IronClaw framework and agent hosting and names PostHog and Stripe, with no retention periods. The release notes state what the Telegram link retains. The home page credits NEAR Foundation while the legal documents name Jasnah, Inc. d/b/a NEAR AI (17 of 30). Removals are dated in the changelog with upgrade steps, and the docs note the channels dropped in 1.0, with no written deprecation policy (12 of 20). No telemetry per the README and the source, and the one upload path, Trace Commons, is documented as opt-in with local preview and revocation (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: ironclaw.com, registered 2018-01-04 (8 years) (11 of 15)
- Privacy policy: read, states 7 of the 8 things a reader expects (9.3 of 10)
- 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.

- 2026-10-12. The ironclaw.com home page lists Discord and Signal among integrations, while the channel docs for the current release say Discord, Signal, WeChat and WeCom are not part of the 1.0 release, which shipped on 27 July 2026. The docs site also says it documents the latest stable release while its heartbeat, cron, shell and WASM pages name settings absent from the 1.4.1 source. A misleading claim, at the minimum, -2. https://www.ironclaw.com/; https://docs.ironclaw.com/channels/overview.md; https://github.com/nearai/ironclaw/blob/ironclaw-v1.4.1/docs/capabilities/routines/heartbeat.mdx

## 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 harness. The briefs define no assistant kind, and a self-hosted agent program the owner runs is closest to the harness listings (Goose, OpenHands), so the harness readings of Schema, Ergonomics and Security were used
- Whether development has paused or moved. `main` has no commit after 10 September 2026 while the repository's last push was 11 October and 1.4.1 shipped from another branch
- Whether the web interface tells a new owner that auto-approve is on. The default was read from the source, and the software was not run
- Whether heartbeat exists in 1.x under another name. The docs page describes it with settings the 1.4.1 source does not contain
- Unchecked: the hosted service at agent.near.ai, its sign-up, card requirement and what the Starter plan costs. Nothing was sent to it
- Unchecked: trust.near.ai and any certification for NEAR AI Cloud
- Unchecked: whether GitHub private vulnerability reporting is enabled for the repository
- Unchecked: crates.io, npm, Homebrew and the MCP registry were not searched for an IronClaw package
- Unchecked: how memory is deleted. No delete tool is declared and the docs describe none
- The NEAR AI terms forbid automated extraction of data and outputs except through the published APIs. Recorded as a fact with no deduction, and it matters before any probe is run
- The privacy policy and terms were read through a summarising reader, so their wording should be checked before any clause is quoted
- The repository was cloned before github.com's robots.txt was read. The clone path is not one that file disallows for Anthropic agents

## Weaknesses

- Global auto-approve defaults to on in the 1.4.1 source, so tools marked ask (send, trash, delete) run without a prompt. Only financial, approval and budget effects always ask
- The `main` branch has had no commit since 10 September 2026, and open reports filed since mid-September show no replies. 1,363 issues and 183 pull requests are open
- Docs pages for heartbeat, cron, shell and WASM limits name settings such as `HEARTBEAT_ENABLED` and `ALLOW_LOCAL_TOOLS` that do not appear in the 1.4.1 source
- No SECURITY.md, security.txt or published GitHub advisory was found
- The home page lists Discord and Signal integrations, while the channel docs say both are not part of the 1.0 release

## 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.

- Open Settings and turn off global auto-approve, or set send and delete tools to ask each time, before connecting mail or chat accounts
- Install with the release installer and run `ironclaw onboard`. 1.x does not migrate 0.29 data, so use a fresh home directory
- Call `POST /api/v1/responses` with the bearer token from `~/.ironclaw/reborn/webui-token`. Keep under 30 requests in 60 seconds
- Resolve approval and OAuth gates in the web interface. The Responses API fails the turn instead of surfacing them
- Connect MCP servers over HTTP only. The stdio transport is rejected

## 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.