Home

agentbox documentation

agentbox lets a frontier model in Claude Code (the planner) hand well-specified coding tasks to another model (the executor), usually an open-weight model on your own GPU. Each task runs Claude Code headless inside a hardened, network-less container on a scratch clone of your repo. The result comes back as a branch agent/<id> that the planner verifies, inside the same sandbox, before you merge it.

These docs follow the Diátaxis structure, like Canonical's documentation:

Tutorials: learn by doing, start here	How-to guides: steps for a specific goal
Reference: exact settings, tools and APIs	Explanation: how it works and why

Tutorials

• Your first delegation: install, run the smoke test, delegate a real fix, verify it and merge it

How-to guides

• Install agentbox
• Delegate work from Claude Code
• Write a spec a local model gets right
• Prepare a Python environment for tasks
• Add a local model (vLLM or Ollama)
• Connect OpenAI and other remote APIs
• Run agents under gVisor
• Operate agentbox: start, stop, upgrade, logs, uninstall
• Troubleshoot

Reference

• MCP tools
• Configuration: .env, secrets.env, models.yaml
• Command line
• Sandbox specification
• Internal APIs: launcher, gateway, llm-bridge

Explanation

• How agentbox works
• Life of a delegation
• How a network-less agent reaches a model
• Security model
• Collecting results safely
• Remote models and llm-bridge
• GPU arbitration
• Design decisions