Steel

Steel connects agents to outside tools and systems two ways. A named integration catalog gives the agent a Steel browser as a tool inside more than twenty frameworks and agents (Vercel AI SDK, LangGraph, CrewAI, Pydantic AI, Mastra, OpenAI Agents SDK, Claude Agent SDK, Browser Use, Stagehand, Hermes Agent and others), and the Credentials API lets the agent act in outside systems under stored logins: credentials are stored with envelope encryption, injected into a session's logins with TOTP support, and isolated so the agent cannot read them back.

Sourcedocs.steel.dev/integrationsread 2026-09-23

Multi-step and multi-agent workflows are not something Steel runs itself. It manages a browser session's lifecycle (create, keep alive, release on an inactivity timeout or an explicit call), which is how one browser runs. The cookbook's multi-step recipes run in the customer's chosen engine (CrewAI, LangGraph, Temporal, Restate, Convex), and no workflow runtime or step chaining of Steel's own is documented.

Sourcedocs.steel.dev/overview/sessions-api/session-lifecycleread 2026-09-23

Retrieval over the customer's own content is not part of the product: no knowledge base, index, vector store or grounding layer of Steel's own is documented. Steel's scrape, screenshot and PDF tools turn live web pages into HTML, markdown and images for the customer's own pipeline, which is extraction from the public web rather than retrieval over the customer's documents.

Sourcedocs.steel.dev/overview/browser-tools/overviewread 2026-09-23

Human intervention is built in: Steel's human-in-the-loop guide embeds a session's live view with interactive=true and showControls=true, so a person can take control of a running browser to enter credentials, solve a CAPTCHA, or verify and correct what the agent did. No approval gate, pause primitive, action allow list or policy guardrail is documented, and when to hand control to a person is decided in the customer's own code.

Sourcedocs.steel.dev/overview/sessions-api/human-in-the-loopread 2026-09-23

Enterprise SSO through the customer's identity provider comes on the Scale and Enterprise plans, per Steel's pricing and limits page. The Credentials API protects each credential with its own AES-256-GCM key wrapped by an organization-specific KMS key and keeps secrets out of the agent's reach after injection, and Scale and Enterprise offer a HIPAA-ready BAA. No trust center, security page, certification badge or attestation such as a SOC 2 report is published, a BAA is a contract rather than an attestation, and SSO appears only as a plan feature with no roles or permissions documented.

Sourcedocs.steel.dev/overview/pricinglimitsread 2026-09-23

The agent's own behavior is recorded step by step. Agent Traces turns each recorded session into a timeline of agent activity (one row per action with verb, target, page URL and timestamp, expandable to page, element, pointer, keyboard or error detail), synced to the session recording and built from what happened in the browser rather than what the agent claimed.

Traces export as markdown, JSON or a ZIP with screenshots and are served by a REST endpoint, sessions carry console and network logs, and data retention runs by plan (7 days on Launch, 14 on Scale, custom on Enterprise). Capture of the agent's prompts and an audit log separate from the traces are not documented.

Sourcedocs.steel.dev/overview/agent-traces/overviewread 2026-09-23

No memory the agent reads, with a stated scope and lifetime, is documented. What Steel keeps between runs is the browser's state: Profiles persist cookies, local storage, extensions and auth context across sessions, session context can be captured from one run and replayed into the next, and the Credentials API stores logins for injection. Each is loaded by the browser at session start rather than read by the agent as context to decide, so it works as a saved browser login rather than as agent memory.

Sourcedocs.steel.dev/overview/profiles-api/overviewread 2026-09-23

Steel can run in the customer's own environment: the open source Steel Browser runtime and API self-host in the customer's infrastructure with Docker, a one-click Railway deploy, or a clustered setup, with guides for each. The managed cloud itself is a fixed single region. The session configuration page says Steel runs browsers in one region, us-east, and rejects any other region value, so a buyer who needs a different hosting location gets it only by self-hosting. Proxy geolocation changes the IP a site sees, not where the browser runs.

Sourcedocs.steel.dev/overview/self-hosting/dockerread 2026-09-23

Ready-made starting points ship in two forms: five installable agent skills for coding agents (steel-browser, steel-developer, steel-reliability, steel-session-debugging and a skill creator) and a cookbook of starter projects across frameworks and languages. The skills teach the customer's own agent to use Steel and the starters are examples a developer builds from; neither is a catalog of agents a buyer adopts to do work.

Sourcedocs.steel.dev/overview/skills/available-skillsread 2026-09-23

A browser starts only when a caller creates a session through the API, SDK or CLI; no event, schedule, webhook or inbound queue of Steel's own starts agent work. The scheduled recipes in the cookbook run on the customer's scheduler (Convex crons, Temporal workflows), not on Steel.

Sourcedocs.steel.dev/overview/sessions-api/overviewread 2026-09-23

Steel runs no model the buyer chooses. The model lives in the customer's agent, whether Claude, OpenAI or Gemini computer use or a framework the customer runs, and each integration page shows that agent calling a Steel browser, so model choice and routing sit with the customer's agent rather than with Steel.

Sourcedocs.steel.dev/integrationsread 2026-09-23

Steel is callable from outside through a documented REST API (api.steel.dev, with an API reference covering sessions, profiles, files, extensions, credentials, CAPTCHAs, browser tools and agent traces, authenticated by API key), official Node and Python SDKs, a CLI, and CDP connections for Playwright, Puppeteer and Selenium. The open source Steel Browser exposes the same API when self-hosted.

Sourcedocs.steel.dev/api-referenceread 2026-09-23

Debugging after a run goes wrong is supported: the Session Debugging skill has a coding agent collect evidence on a failed session and diagnose it, handing off to the Reliability skill when the fix is bot detection, CAPTCHA, proxy, profile or retry strategy. This is debugging rather than testing, and no evaluation harness, scored test cases or quality gate is documented.

Sourcedocs.steel.dev/overview/skills/available-skills/steel-session-debuggingread 2026-09-23

Operating a real browser for agents is Steel's product: isolated cloud browser sessions an agent drives to navigate, click, fill forms and read pages, through CDP with Playwright, Puppeteer or Selenium or through Claude, OpenAI and Gemini computer-use agents, with mobile mode, stealth, proxies, CAPTCHA solving and persistent profiles for production use.

Sourcedocs.steel.dev/overview/intro-to-steelread 2026-09-23