agent-browser-mcp-codex

Expose local agent-browser capabilities as MCP tools and bootstrap Codex integration quickly.

Quick Start for Codex

Prerequisite: agent-browser must already be installed locally.

1. Run npx agent-browser-mcp-codex init-codex

2. Restart Codex or open a new session

Related MCP server: OpenCode in Chrome

What agent-browser already provides

agent-browser already provides a broad browser automation surface, including:

• Navigation and interaction

• Keyboard, mouse, upload, download, and scrolling controls

• Page inspection and info retrieval

• Tabs, sessions, storage, and network operations

• Debugging, diffing, recording, and streaming

• Auth, chat, dashboard, and setup commands

What this MCP wrapper currently exposes

This package currently exposes a focused subset of agent-browser commands:

• Navigation: open, back, forward, reload

• Interaction: click, fill, type

• Read and page state: get_title, get_text, get_html, snapshot, screenshot, wait

• Tabs: tab_list, tab_new, tab_close

• Runtime: eval

• Storage and network: cookies_get, network_requests

• Sessions and raw passthrough: session_current, session_list, session_close, run_raw_command

Manual Setup

If you want to wire the plugin in manually, the local Codex templates live under codex/plugin/.

Limitations

• This package currently exposes a subset of the upstream agent-browser CLI, not the full surface.

• Some tool families are intentionally narrow right now.

• Broader coverage may be added over time, but the current surface should be treated as evolving.

Development

For local verification:

• npm test

• npm run typecheck

Release

• Configure an NPM_TOKEN secret in the GitHub repository before the first automated publish.

• Create and publish a GitHub Release to trigger npm publication for the current package version.