Features

browser-mcp focuses on practical browser automation: direct Playwright-style tools for precise steps, plus an agent tool for tasks that need planning across multiple page states.

Core capabilities

Capability	What it does	Typical use
Page observation	Reads current URL, title, and visible body text.	Decide the next selector or summarize the page state.
Navigation	Opens URLs, scrolls pages, and waits for selectors, text, or timeouts.	Move through multi-page flows reliably.
Interaction	Clicks selectors, types text, and presses keyboard keys.	Fill forms, submit flows, open menus, or trigger UI actions.
Screenshots	Captures full-page or element screenshots as PNG artifacts.	Evidence, debugging, reports, before/after checks.
Artifact access	Lists and reads files from the runtime data directory.	Retrieve screenshots or generated downloads from the container.
Agent tasks	Runs a browser-use agent from a start URL and instruction.	Ask for a higher-level browser outcome when direct steps are too verbose.

Runtime model

The server starts a headless Chromium runtime and exposes MCP tools over SSE. Direct tools use Playwright through BrowserRuntime; agent tasks use browser-use with the configured OpenAI-compatible model.

Generated artifacts are written under /data in the container. Docker Compose binds that to ${DATA_PATH}/browser-mcp on the host, so useful files survive container restarts without exposing the rest of the host filesystem.

Safety boundaries

• Runtime browser sessions are headless.
• Artifact writes use safe generated filenames.
• Artifact and download reads validate the requested path stays inside /data.
• type_text returns only the number of typed characters, not the typed value.
• Credentials, cookies, auth headers, passwords, and TOTP secrets should not be logged, returned, or persisted.