Home

agentbox wiki

agentbox lets a frontier model in Claude Code (the planner) hand well-specified coding tasks to an open-weight model on your own GPU (the executor). 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>, which the planner reviews and merges.

Why do it this way? Local models are free per token and keep code on your machine. They are also slower, and they fail in predictable ways. A strong planner that writes precise specs, plus a reviewer that runs the tests before merging, turns them into useful workers. Running them in a sandbox means a confused or prompt-injected model can't damage your machine or leak your code.

Pages

• Architecture: the three services and the life of a delegate_task call
• Security Model: what each layer enforces, the attack tests, residual risks
• Installation: requirements, setup, smoke test
• Configuration: .env, config/models.yaml, adding vLLM and Ollama models
• Using with Claude Code: registering the MCP server, the workflow, CLI and benchmarks
• Writing Good Specs: how to get correct results from a local executor
• Troubleshooting
• FAQ