Container Use gives each coding agent its own containerized development environment and its own git branch, so several agents can work on a repository at the same time without overwriting one another. A failed attempt is discarded by dropping the branch, and finished work is reviewed with a normal git checkout .
It is an MCP server that also works as a command-line tool, so it pairs with Claude Code, Cursor, Goose, VS Code, and other MCP-compatible agents through the same step: registering container-use stdio as an MCP server. Containerized workflows are handled by Dagger and branching by git worktrees. The cu command is an alias for container-use.
Alongside isolation, the tool records what each agent did: you can read the full command history and logs for an environment, and drop into an agent's terminal to inspect its state or take over when it stalls. The project describes itself as in early development and actively evolving.
Features
- Isolated environments: every agent gets a fresh container on its own git branch, so parallel runs do not conflict
- Command history and logs: the full record of what an agent actually ran is available for review
- Direct intervention: open any agent's terminal to see its state and take control
- Git-native review: check out an agent's branch to inspect or keep its work, and discard failures by dropping it
- MCP setup: any MCP-compatible agent connects by adding container-use stdio as a server
- Agent rules file: an optional rules document can be appended to the agent's instructions in the repository
