Steel

# Fix list: Steel

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/steel, the October 2026 research run, assessed 8 October 2026. Grade BB, 70.4 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 Steel: 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, 58 out of 100, up to 8.4 more on the total

Why it scored 58: Graded as a hosted service, on Steel Cloud. status.steel.dev runs on Better Stack with Steel API, Dashboard, Proxies and Connections components and 90 days of history (20). The Main API shows 99.762 per cent over 90 days. The page records a 4 hour 10 minute outage for domain and SSL problems on 22 July 2026, 5 hours 41 minutes of capacity errors on session creation on 23 August and 1 hour 55 minutes of failing session creation on 1 October, so several majors (0). Seven other entries include two streaming and recording problems of more than an hour on 23 and 25 September, during which Steel said the API worked normally. Limits are published per plan, 60 and 600 requests a minute, 10 and 100 concurrent sessions, and 20 a minute on Browser Tools (15). Browser Tools document 429 and a 503 marked safe to retry, the Agent Traces API says to wait for `Retry-After`, and the SDKs retry 408, 409, 429 and 5xx twice with backoff. No idempotency keys, though session creation accepts a caller-chosen `sessionId` (10). Support and uptime SLAs are listed for Enterprise only, with no terms published (3). The sessions API is not marked beta (10).

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, 56 out of 100, up to 7.7 more on the total

Why it scored 56: Named API keys, shown once, revoked by deletion, with last-used times and a rotation guide. A global key covers the organisation and can create keys bound to one project through `POST /v1/api-keys`. No per-permission scopes were found (22). The documented CDP connection passes the key as `apiKey` in the `wss://connect.steel.dev` URL, a secret in a query string (minus 10). No read-only key. Projects isolate sessions, credentials and profiles, stored credentials are typed into pages without being shown to the agent, browser action endpoints reject localhost and private addresses since changelog 038, and the MCP server has a scrape-only profile and hands the browser to a person before logins and payments (13). The MCP server fences page content as untrusted and strips hidden text, and its README warns about injection. The API docs carry no such guidance (10). Session logs, recordings, Agent Traces, per-session cost and exports to JSON or CSV give the operator a per-call trail (13). A Vanta trust centre whose contents did not render for us, a HIPAA-ready BAA and SSO on Scale, and a disclosure policy with response times for the MCP server. No security.txt, bug bounty or certification report was found (8).

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, 79 out of 100, up to 3.4 more on the total

Why it scored 79: Scrape returns only the formats asked for (HTML, cleaned HTML, Markdown or readability text), lists take `limit`, and the MCP server holds its 16 tool definitions to a 17,000 byte budget checked in CI, with a three-tool scrape profile (20). Sessions list by cursor with a `status` filter and a default page of 50, other lists take `limit` and `offset`, and traces filter by time and event type (17). Errors return `message`, `linkToDocs` and validation `context`, with documented 401 messages. No stable machine-readable error codes were found (15). No idempotency keys. A caller-chosen `sessionId`, `inactivityTimeout` and release-all help with retries and clean-up, and every MCP tool carries read-only, destructive or idempotent annotations (12). Session creation needs no parameters, and there are SDKs for Node and Python, with Rust and Go per Steel's blog (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. Transparency & trust, 64 out of 100, up to 3.2 more on the total

Made of editorial 52, provenance 75.

Why it scored 64: Graded on the cloud service, which is closed with published terms. The browser runtime steel-browser and the SDKs are Apache-2.0, and the MCP server and CLI are MIT (24). The privacy notice (5 February 2026) keeps data while an account exists, the pricing page gives 7 and 14 days of data retention, recordings can be turned off per session, and a DPA is linked as a Google Drive file. The terms date from 5 February 2025, the docs Legal page links to Google Docs copies, and the FAQ still describes pricing by concurrent sessions (14). No deprecation policy. Legacy plans stay available after the pricing change, and one docs note says a playback mode will be deprecated without a date (4). The privacy notice names Anthropic, OpenAI and Google Cloud AI as AI processors and says servers are in the United States, and the docs name one region, us-east. No subprocessor list was found outside the trust centre, which did not render for us (10).

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: steel.dev, registered 2024-07-22 (2 years) (7 of 15)
- Terms of service: read, states 6 of the 7 things a reader expects, and has 3 clauses that cost points (3.1 of 10)
- security.txt: not found (0 of 10)

## 5. Payments & pricing, 75 out of 100, up to 3.1 more on the total

Why it scored 75: x402 version 2 at x402.steel.dev for scrape, screenshot and PDF, in USDC on Base or Solana. An unpaid request on 8 October 2026 returned 402 with terms. Browser sessions, the main product, aren't covered, and the payment header names api.paysponge.com as the resource, so partial (20). Per-hour, per-GB and per-1,000-call prices published without login (20). Launch plan with $30 of one-time credit valid 90 days, and the quickstart says no card is required (20). An agent with a wallet gets a scrape, screenshot or PDF with no account, but a browser session needs a person to sign up or run the Stripe CLI, so we award 15 of 20.

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. Schema & documentation, 86 out of 100, up to 2.3 more on the total

Why it scored 86: OpenAPI 3.0.3 at docs.steel.dev/openapi.json with 51 paths and 75 operations and the `steel-api-key` header scheme (25). llms.txt, llms-full.txt and a Markdown twin of every docs page (10). 65 of 75 operations carry a description, and the MCP server's tool descriptions say when to use each tool and when to pick another (16). Inputs are typed with enums, UUID formats and list limits of 1 to 100, but `timeout` and `concurrency` carry no real bounds and `useProxy` and `region` have no stated type (11). One shared error schema is attached to most operations and the docs have error tables for Browser Tools and Agent Traces. The spec holds one example, and 429 is declared on session creation only (11). Paths are versioned under /v1 and the changelog is public and numbered, with dates in page metadata. The spec's own version reads 0.0.1 (13).

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.

## 7. Maintenance & community, 82 out of 100, up to 1.6 more on the total

Why it scored 82: Changelog 038 is dated 2 October 2026, and steel-browser's latest commit is 6 October (30). Changelog entries on 24 July, 18 and 25 September and 2 October 2026, plus steel-browser v0.5.4-beta on 25 August and MCP server 3.0.0 on 9 September (20). Public changelog, Discord and email support, and steel-browser merged ten fixes between 20 July and 6 October. GitHub issue reply times were not read (15). Official SDKs exist, but the newest stable releases are npm 0.18.0 from 16 March 2026 and PyPI 0.19.0 from 23 June, with previews in September, and the MCP server is not in the MCP registry (9). CI workflows in steel-browser and a test, byte-budget and conformance suite in the MCP server. Whether they pass was not read (8).

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.

## 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: the contents of the Vanta trust centre at trust.steel.dev (certifications, subprocessors, policies), which needs JavaScript and did not render for our reader
- unchecked: GitHub issue counts, reply times, CI results and the star count, because the GitHub API refused us for its rate limit and the issue pages did not parse. Steel's site states 7.8K stars for steel-browser
- unchecked: the DPA, linked from the privacy notice as a Google Drive file, was not opened
- unchecked: the Rust and Go SDKs, taken from Steel's blog index and not read in a registry
- Whether the Launch signup asks for a card. The quickstart says no credit card is required and the pricing page does not say
- Who operates x402.steel.dev. The 402 header names https://api.paysponge.com/v1/scrape as the resource and the docs do not mention a third party
- The authentication page says keys are created and deleted only in the dashboard, while the OpenAPI document has `POST /v1/api-keys` for project-bound keys
- The Node SDK repository has tags v0.19.0 and v0.20.0 (23 June 2026) that are not on npm, where `latest` is 0.18.0
- Steel's llms.txt calls its speed results independent benchmarks while linking to benchmark code in its own steel-dev/browserbench repository. We did not assess the claim and made no deduction
- The docs name us-east as the only region while the OpenAPI description of `region` lists six. The FAQ describes pricing by concurrent sessions, which the pricing page no longer does

## Weaknesses

- Three incidents of more than an hour hit the API or session creation in the 90 days to 8 October 2026, the longest 5 hours 41 minutes
- The documented CDP connection passes the API key as `apiKey` in the `wss://connect.steel.dev` URL
- x402 covers the three one-shot endpoints only, so a browser session needs an account and a key
- The MCP server is not on npm or in the MCP registry, and `mcp.steel.dev` is not live per its README
- No deprecation policy or security.txt found, and the trust centre did not render for our reader

## 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.

- Use `POST /v1/scrape` for pages you only need to read. It starts no session and costs $5 per 1,000 calls, or $0.01 a call by x402
- Release every session with `POST /v1/sessions/{id}/release` and set `inactivityTimeout`. Browser time bills by the minute, rounded up
- Build the CDP URL as `wss://connect.steel.dev?apiKey=<key>&sessionId=<id>` and keep it out of logs, because it carries the key
- Launch sessions end after 15 minutes at most, and `timeout` cannot be changed on a live session
- CAPTCHA solving and Steel's managed proxies on Launch need $10 of paid balance. Free credit does not count

## 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.