monkeybot

Multi-cloud agent harness — GCP-first docs, portable runtime

monkeybot is an owned agent runtime for tool-using LLMs: a FastAPI SSE gateway, pluggable storage, MCP tools, skills, and memory. Run locally with zero cloud dependencies, or ship the same container image anywhere Docker runs.

Docs and examples are GCP-first (Vertex, GCS, Cloud Run); AWS paths ship in the Pattern guides. See Cloud deployment — Positioning.

[!NOTE] Skills under SKILLS_PATH execute Python code — only add skills from trusted sources.

Quick start

Install uv, then:

uv tool install monkeybot-cli

monkeybot new --dest ./my-agent --provider openai --yes
cd my-agent
uv sync
cp .env.example .env   # add provider keys
monkeybot doctor
monkeybot chat

Upgrade the CLI with uv tool upgrade monkeybot-cli. Full walkthrough: Getting Started.

What you get

• SSE gateway — sessions, streaming, health; wire your own UI or use the CLI
• Realtime talk — WebSocket voice/text via monkeybot talk
• MCP + skills — stdio/HTTP tools and trusted Python skill bundles
• Pluggable storage — SQLite by default; Postgres, Firestore, GCS, S3
• Multi-provider — Gemini/Vertex, OpenAI, Anthropic, Bedrock, Ollama, and more

CLI

monkeybot chat            # spawn gateway + REPL
monkeybot run             # terminal 1: gateway with logs
monkeybot chat --attach   # terminal 2: attach to it

Provider extras live on the agent pyproject.toml (monkeybot[<provider>]), not the global CLI. See Getting Started and Features — CLI.

Agent skill

npx skills add human-plus-machine/monkeybot --skill monkeybot

Walks through install, scaffold, config, and first chat. Source: cli/skills/monkeybot/.

Documentation

Integrations matrix and roadmap: CHANGELOG · BACKLOG.

Contributing

Clone only if you are changing the harness. Setup, checks, and the release workflow are in CONTRIBUTING.md. Live eval smoke fixture: evals/smoke_agent/.

Support

Search or open an issue in the GitHub issue tracker.

Credits

MonkeyBot builds on two open-source projects:

• Browser Use — browser automation for the browser MCP (via browser-harness and the interactive-element DOM walker). MIT licensed.
• OpenSandbox — sandbox runtime used for isolated run_command / code execution. Apache-2.0 licensed.

License

MIT — see LICENSE.