What is agent-browser?
agent-browser is a browser automation CLI with companion skills for AI agents. It controls Chrome or Chromium through the Chrome DevTools Protocol and represents pages as accessibility-tree snapshots with element references. Agents can use these references to click controls, complete forms and extract content without processing the full page HTML.
The skill entry point directs agents to workflow documentation served by the installed CLI. Specialized skills cover exploratory testing, Electron applications and cloud browser environments. The underlying tool runs through a native Rust CLI and daemon on macOS, Linux and Windows, with an optional MCP interface for compatible clients.
Deployment requires installing the CLI and a compatible browser, plus browser libraries where needed. Sessions, profiles and authentication-state persistence support ongoing workflows, but transient state is discarded on shutdown unless restoration is configured. Website content and page-provided tools remain untrusted and do not replace host permissions or approval controls.
Key capabilities
- Browser interaction: Navigates pages, clicks elements, fills forms, uploads files and manages tabs and frames.
- Page inspection: Produces accessibility-tree snapshots with element references and extracts rendered text, HTML and attributes.
- Visual evidence: Captures screenshots, annotated screenshots, PDFs and video recordings for inspection and debugging.
- Session management: Supports named sessions, persistent profiles and saved authentication state for separate workflows.
- Agent workflows: Supplies core and specialized skills through the CLI, alongside an MCP interface for compatible clients.
Where agent-browser fits
agent-browser fits AI-assisted web interaction, data extraction and exploratory testing workflows. Developers and QA teams can use it to inspect applications, exercise forms and capture visual evidence. Specialized skills extend the workflow to Electron applications and supported cloud browser environments.
Before deployment, teams should verify browser availability, operating-system dependencies, session isolation and authentication-state handling. They should configure trusted access policies and confirmation requirements for consequential actions. Agents need fresh snapshots after page changes, and obstructed, hidden or disabled controls can prevent interactions.
Positioning and limitations
agent-browser suits developers and teams whose AI agents can execute shell commands or connect to an MCP server. Its skill layer provides workflows around a locally installed automation tool rather than a standalone hosted application. Browser setup, permissions, session persistence and approval policies remain deployment responsibilities; page-provided metadata is not a security boundary.
Why it matters
Browser interaction requires agents to connect page observations with specific actions. agent-browser combines compact accessibility snapshots, element references and workflow guidance to support that loop, while session controls and visual evidence help teams inspect what the automation is doing.
agent-browser: key facts
- First published
- Profile updated
- Access
- Self-hosted
- Source model
- Commercial
- GitHub repository stars
- 45,056
- Repository forks
- 3,084
- Main repository language
- Rust
- Recorded profile views
- 8
- Repository license
- Apache-2.0 ↗
- Pricing model
- See vendor
- Repository release
- v0.41.0 ↗ ·
- Website
- github.com ↗
- GitHub
- Source repository ↗
Compare agent-browser and explore alternatives
Compare deployment, pricing model and available integrations with tools that match your workload.
Open a comparison with agent-browser → · Browse related profiles →
What to verify
Start from the users, systems and operating responsibilities the tool needs to support.
- Confirm current features, licensing and support terms.
- Validate deployment, data location and access control.
- Test integrations, exports and recovery workflows.
Reviews of agent-browser
No published reviews yet.
Loading review form…