Introduction

What is claude-forge?

claude-forge is a CLI that runs Claude Code inside an isolated Docker container, so the agent can work autonomously without access to your host machine.

Claude Code is most productive when it can act without asking permission for every step (--dangerously-skip-permissions) — but granting that on your host means an autonomous agent with your credentials and your filesystem. claude-forge resolves that tension with containers:

• The agent only sees the project you mount at /home/user/work, never your host filesystem, host Docker daemon, or unrelated credentials.
• All git and GitHub API traffic flows through a gateway proxy that lets the agent read any repository but write only to the current project’s — a compromised or confused agent cannot push to your other repos.
• Each run is a named session you can list, resume, and prune, and multiple sessions run in parallel across projects.

Tool integrations use MCP (Model Context Protocol), the standard way to give Claude Code extra capabilities: a GitHub MCP server scoped to your repository comes built in, first-party Kubernetes and Google Cloud servers ship in the repo, and you can add your own servers.

Next steps