Hyperbrowser vs Steel (2026)

LangChain, LlamaIndex, Stagehand and AI function calling guides give an agent Hyperbrowser's browser, scrape and crawl as tools. HyperAgent's custom actions let a developer add their own tools, and x402 lets agents pay for fetches and searches.

Sourcehyperbrowser.ai/docs/integrations/langchainread 2026-09-23

Agents reach outside tools and systems through Steel in 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). 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

One API call starts a hosted HyperAgent, Browser Use or model native computer use task (Claude, OpenAI, Gemini, Meta, Grok, Jev) in a managed session. The agent works through its steps, and the customer polls status, reads results or stops it. The open source HyperAgent adds multi page management within one agent. Each task is one agent working through a multi step loop. Hyperbrowser has no workflow composition, step chaining or multi agent execution of its own, and action caching replays one agent's steps.

Sourcehyperbrowser.ai/docs/agents/overviewread 2026-09-23

Each browser session has a lifecycle Steel manages. Sessions are created, kept alive, and released on an inactivity timeout or an explicit call. Steel does not run multi step or multi agent workflows itself. Multi step recipes in the cookbook run in the customer's chosen engine (CrewAI, LangGraph, Temporal, Restate, Convex), and Steel has no workflow runtime or step chaining of its own.

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

Work starts when a caller creates a session, sandbox, scrape job or agent task through the API or SDKs. No event, schedule, webhook or inbound queue of Hyperbrowser's own starts agent work. Scheduled jobs are a listed use case for action caching replays, but the customer runs the schedule.

Sourcehyperbrowser.ai/docs/api-reference/create-new-sessionread 2026-09-23

A browser starts only when a caller creates a session through the API, SDK or CLI. Steel has no event, schedule, webhook or inbound queue of its own that starts agent work. 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

Fetch, search, scrape, crawl and AI extract endpoints turn public web pages into markdown, HTML or structured data for the customer's own pipeline. Hyperbrowser keeps no index, vector store or grounding layer of its own over customer content.

Sourcehyperbrowser.ai/docs/web/overviewread 2026-09-23

Retrieval over the customer's own content is not part of the product. Steel names no knowledge base, index, vector store or grounding layer of its own. Its scrape, screenshot and PDF tools turn live web pages into HTML, markdown and images for the customer's own pipeline. They extract from the public web and do not search the customer's documents.

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

Sandboxes carry state forward. A memory snapshot freezes a running sandbox's memory, processes and filesystem, and new sandboxes start from that state. Volumes persist data across sandboxes, so the agent's own code reads the restored state to carry on. A snapshot restores the machine's state whole, and volumes are storage the code reads. There is no separate memory layer with its own scope and lifetime. Browser profiles hold the saved login loaded at session start, and HyperAgent action caching replays recorded actions without LLM calls.

Sourcehyperbrowser.ai/docs/sandboxes/snapshotsread 2026-09-23

Between runs, Steel keeps 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. The browser loads each of these at session start, and the agent does not read them as context for its decisions. Steel names no memory the agent reads, and no scope or lifetime for one.

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

Every session carries a Live View URL that the customer can embed in its own application. Through it a person can watch and interact with the running browser, and human in the loop interactions are one of its named uses. There is no approval gate, pause primitive or policy guardrail, and the customer's code decides when to hand control to a person.

Sourcehyperbrowser.ai/docs/sessions/live-viewread 2026-09-23

A person can step into a running session. Embedding a session's live view with interactive=true and showControls=true lets a person take control of the browser to enter credentials, solve a CAPTCHA, or verify and correct what the agent did. Steel names no approval gate, pause primitive, action allow list or policy guardrail. The customer's own code decides when to hand control to a person.

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

The information security program is aligned with the SOC 2 Trust Services Criteria. It covers independent third party assessments, annual third party penetration testing, encryption at rest and in transit, vulnerability scanning, logging and monitoring, incident response, and backup and recovery. The Enterprise plan lists 'HIPAA/SOC 2', but the Security Commitments say only 'aligned with', so the attestation is unclear. Access controls (least privilege, quarterly access reviews, SSO where available) govern Hyperbrowser's own team. Customer accounts have no SSO, roles or permissions.

Sourcehyperbrowser.ai/securityread 2026-09-23

Enterprise SSO through the customer's identity provider comes on the Scale and Enterprise plans. 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. Scale and Enterprise also offer a HIPAA ready BAA. Steel publishes no trust center, security page, certification badge or attestation such as a SOC 2 report. SSO comes as a plan feature, and Steel names no roles or permissions.

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

Every session is recorded. Web recordings (rrweb) capture DOM changes, interactions and network requests for replay, and MP4 video recordings are available too. Both are fetched through the API, and every session has a live view. Data retention runs by plan, with 7 days on Free, 30 on Startup and Scale, and 180 or more on Enterprise. A customer can replay what the agent did in the browser. There is no structured log of each agent action, no audit log separate from the recording and no export to a monitoring system.

Sourcehyperbrowser.ai/docs/sessions/recordingsread 2026-09-23

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. The timeline is synced to the session recording and built from what happened in the browser, not from what the agent claimed.

Traces export as markdown, JSON or a ZIP with screenshots, and a REST endpoint serves them. Sessions carry console and network logs, and data retention runs by plan (7 days on Launch, 14 on Scale, custom on Enterprise). Steel does not say whether it captures the agent's prompts, and names no audit log separate from the traces.

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

Customers choose a region for each browser session from a published list of us-central, us-east, us-west, europe-west and asia-south. Production data is hosted in United States databases unless otherwise agreed in writing. There is no self hosted option.

Sourcehyperbrowser.ai/docs/sessions/regionsread 2026-09-23

Customers can run Steel in their 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 runs browsers in one fixed region, us-east, and rejects any other region value. A customer 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

A templates gallery offers ready to run sample projects built on the Hyperbrowser APIs, such as a news sentiment aggregator, product review analyzer, shopping recommenders, and API servers wrapping Hacker News, Product Hunt and Craigslist, among others. An examples repository sits alongside. A developer clones and adapts these projects, and there is no catalog of agents a customer adopts as they are.

Sourcehyperbrowser.ai/templatesread 2026-09-23

Five installable agent skills for coding agents (steel-browser, steel-developer, steel-reliability, steel-session-debugging and a skill creator) come with Steel, along with a cookbook of starter projects across frameworks and languages. The skills teach the customer's own agent to use Steel. The starters are examples a developer builds from. Steel offers no catalog of agents a customer adopts to do work.

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

The customer chooses the model behind the hosted agents. Each task starts with a chosen agent, either HyperAgent, Browser Use or a model native computer use agent from Anthropic (Claude), OpenAI (CUA), Google (Gemini), Meta, xAI (Grok) or Jev. Each is billed by its own model's tokens or steps, and HyperAgent supports OpenAI, Anthropic and Google Gemini models as configured providers.

Sourcehyperbrowser.ai/docs/agents/overviewread 2026-09-23

The model lives in the customer's agent, whether Claude, OpenAI or Gemini computer use or a framework the customer runs. In each integration, that agent calls a Steel browser. Model choice and routing sit with the customer's agent, and Steel runs no model of its own for the customer to choose.

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

A REST API makes the platform callable from outside, covering sessions, profiles, extensions, sandboxes, volumes, snapshots, fetch, search, scrape, crawl, extract and agent tasks, with a full API reference. Official Python (pip install hyperbrowser) and Node.js (@hyperbrowser/sdk) SDKs sit alongside it, with the hx CLI for sandboxes and an open source MCP server.

Sourcehyperbrowser.ai/docs/sdks/introductionread 2026-09-23

Outside systems call Steel through a REST API at api.steel.dev, authenticated by API key. Its reference covers sessions, profiles, files, extensions, credentials, CAPTCHAs, browser tools and agent traces. Steel also offers 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

Agents run on the platform, but it does not score them. There are no fixtures, scoring, quality gates or comparison of agent runs. Session recordings and live view serve for debugging, and action caching replays recorded actions deterministically to cut cost.

Sourcehyperbrowser.ai/docs/introductionread 2026-09-23

The Session Debugging skill has a coding agent collect evidence on a failed session and diagnose it. It hands off to the Reliability skill when the fix is bot detection, CAPTCHA, proxy, profile or retry strategy. Steel names no evaluation harness, scored test cases or quality gate.

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

The product is a real cloud browser that agents operate. Agents drive cloud browser sessions through Playwright, Puppeteer or Selenium, use computer actions (click, type, scroll, screenshot), or run as hosted computer use agents from Anthropic, OpenAI, Google, Meta, xAI and Jev. Stealth, proxies, CAPTCHA solving, profiles and extensions support production use.

Sourcehyperbrowser.ai/docs/introductionread 2026-09-23

Agents drive isolated cloud browser sessions from Steel to navigate, click, fill forms and read pages. They work through CDP with Playwright, Puppeteer or Selenium, or as Claude, OpenAI and Gemini computer use agents. Operating a real browser for agents is Steel's core product, with mobile mode, stealth, proxies, CAPTCHA solving and persistent profiles for production use.

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