使用说明
🐙 Octo MCP
Give AI agents access to Octo Browser profiles and browser actions.
Start a profile, connect over CDP, interact with pages and collect results through one MCP server.
⚡ What it does
- 🗂️ Profiles — find, start and stop profiles; create temporary sessions.
- 🌐 Browser actions — navigate, click, type, switch tabs and take screenshots.
- 📄 Page data — read text, HTML and attributes, or run JavaScript.
- 🔌 Integrations — Octo Local API, Cloud API and Playwright over CDP.
🔧 Built for integration
- 37 MCP tools, exposed over stdio.
- Async API clients with bounded retries for HTTP 429 responses.
- Remote browser support with configurable host and CDP endpoint rewriting.
- Checks in CI for linting, types and tests.
🚀 Quick start
Requires Python 3.10+ and a running Octo Browser installation.
git clone https://github.com/mazamaka/octo-mcp.git
cd octo-mcp
python3 -m venv .venv
source .venv/bin/activate
pip install -e .
Connect your MCP client using stdio and the absolute path to .venv/bin/octo-mcp. On Windows, use .venv/Scripts/octo-mcp.exe.
For Claude Code on macOS / Linux, from the project directory:
claude mcp add octo-mcp -- "$PWD/.venv/bin/octo-mcp"
Start with “Check whether Octo Browser is running” (octo_health_check). Profile search and team resources use the Cloud API: configure OCTO_API_TOKEN in your MCP client's environment. Local operations use the running desktop app.
🧪 Development
The CI workflow runs Ruff, mypy and pytest. Run the same checks locally:
pip install -e ".[dev]"
ruff check src/ tests/
ruff format --check src/ tests/
mypy src/ tests/
pytest
Python · MCP · Playwright · httpx
MIT license · Built by Maksym Babenko
