"A web browser for AI agents & applications."
| Website | https://browserbase.com |
| Docs | https://docs.browserbase.com |
| GitHub | https://github.com/browserbase |
| Stars |  |
| MCP | https://www.browserbase.com/mcp |
| Classification | agent-native |
| Category | Browser & Web Execution Services |
| Compliance | SOC-2 Type I · HIPAA |
https://github.com/browserbase/stagehand — Stagehand (open-source NL web automation SDK built on Browserbase)
https://github.com/browserbase/mcp-server-browserbase — Official MCP server
Agent Skills are portable SKILL.md instruction sets following the AgentSkills open standard that teach AI coding assistants (Claude Code, Cursor, Codex, Windsurf, etc.) how to use this service correctly.
Status: ✅ Official skills published at browserbase/skills
npx skills add browserbase/skills| Skill | What It Teaches the Agent |
|---|---|
browser |
Automate web interactions via CLI with remote Browserbase sessions, stealth mode, CAPTCHA solving, and residential proxy support |
functions |
Deploy serverless browser automation functions to Browserbase cloud |
Compatibility: Claude Code, Cursor, Codex, Windsurf, and all AgentSkills-compatible tools.
Tip: After installation, set
BROWSERBASE_API_KEYandBROWSERBASE_PROJECT_IDin your agent's environment settings.
Status: ✅ Available
Browserbase provides an official MCP server that allows LLMs to control remote browser sessions directly through the Model Context Protocol.
| Detail | Value |
|---|---|
| MCP Page | https://www.browserbase.com/mcp |
| MCP Docs | https://docs.browserbase.com/integrations/mcp/introduction |
| MCP Repo | https://github.com/browserbase/mcp-server-browserbase |
| Transport | stdio / Streamable HTTP |
| Compatible Clients | Claude Desktop, Cursor, OpenAI Agents, any MCP-compatible client |
Browserbase provides managed, cloud-hosted headless browser infrastructure specifically for AI agents. Agents spin up isolated browser sessions via API, control them through Playwright, Puppeteer, or Selenium — or through the Stagehand SDK using natural-language web actions — and destroy them when done. The entire browser lifecycle is agent-controlled, with no human sitting at a keyboard.
| Criterion | Evidence |
|---|---|
| Agent-first positioning | Homepage: "A web browser for AI agents & applications"; Stagehand SDK, MCP server, and agent integration guides are first-class products |
| Agent-specific primitive | Remote browser session with per-agent isolation; natural-language action via Stagehand ("click the login button"); session recording for agent trajectory audit |
| Autonomy-compatible control plane | Thousands of parallel sessions; stealth capabilities prevent bot detection; agents operate entirely without human intervention |
| M2M integration surface | REST API, Playwright/Puppeteer/Selenium driver, Chrome DevTools Protocol (CDP), MCP server |
| Identity / delegation | Per-session proxy and fingerprint configuration; session IDs enable per-agent audit |
| Primitive | Description |
|---|---|
| Remote Browser Session | Isolated cloud browser instance created and destroyed via API |
| Stagehand Actions | Natural-language web commands (act("click signup button"), extract(...)) |
| Session Recording | Full replay of agent's browser actions for debugging and audit |
| Stealth Mode | Fingerprint randomization and proxy rotation to avoid bot detection |
| Parallel Sessions | Spin up thousands of simultaneous browser contexts in milliseconds |
| MCP Server | Direct LLM tool use via Model Context Protocol |
- Agent calls
POST /sessions→ receives a remote browser session endpoint - Agent connects via Playwright/CDP or Stagehand
- Agent issues actions (navigate, click, extract, screenshot) without human involvement
- Session is automatically recorded for later inspection
- Agent calls
DELETE /sessions/{id}when done
There is no human browser window. The agent is the only actor in the session.
- Each session is isolated — no shared state between agent sessions
- Proxy and fingerprint configuration is per-session, enabling agent-specific identity on the open web
- Session IDs provide per-agent attribution in logs
- API key authentication scopes access per project
| Interface | Detail |
|---|---|
| REST API | Session lifecycle management (create, list, delete) |
| Playwright | Standard Playwright over CDP pointing to remote session |
| Puppeteer | Standard Puppeteer over CDP pointing to remote session |
| Selenium | Standard Selenium WebDriver pointing to remote session |
| Chrome DevTools Protocol | Low-level CDP access for custom tooling |
| Stagehand SDK | Natural-language action layer built on Browserbase |
| MCP Server | Model Context Protocol for direct LLM browser control |
None required by default. Session recordings can be reviewed by humans post-hoc for debugging. HumanLayer can be composed on top for approval gates on specific high-risk actions (e.g., before the agent submits a form with financial implications).
| Alternative | Why It Fails |
|---|---|
| Local Playwright/Puppeteer | Runs on developer's machine; no cloud scaling, no per-agent session isolation, no stealth, no agent-optimized SDK |
| BrowserStack | Built for human QA engineers running test suites; not designed for agent-initiated autonomous navigation at scale |
| Selenium Grid | Human test infrastructure; no natural-language action interface, no agent-oriented session management |
| AWS Lambda + Chromium | DIY infra with no agent-specific primitives; developer builds all session management, stealth, and recording from scratch |
- Web automation — agents log in to third-party services, extract data, submit forms, take screenshots
- Research agents — agents navigate multi-step research workflows across multiple sites
- Testing agents — agents validate UI behavior as part of an agentic CI pipeline
- E-commerce agents — agents monitor prices, fill carts, complete checkouts on behalf of users
- Data extraction — agents navigate dynamic, JavaScript-rendered pages and extract structured information