4.0k stars · Apache-2.0 · v0.4.2 (2026-08-19), read from /releases/latest; the date showed no year and falls in the past, so 2026 · Track this in Scout
Gives each coding agent its own container with its own copy of the repository, plus a full log of every command it ran.
▶Repo detailsthe review · specs · pros & cons · install
What it is
A tool and an MCP server from the makers of Dagger. MCP is the standard way an assistant is given extra tools. Each task an agent starts gets a fresh container with a copy of your repository inside it, and the work is recorded as a git branch plus a full log of every command. You inspect an environment while it is running, and cu checkout brings its branch into your own repository when you are happy.What it is good for. Anyone who builds with a coding assistant and has learned to be nervous. The problem it removes is the review you cannot do: at the moment an assistant edits your files, you only see the result. Here you see the commands, the installed packages and the running process, and the work is a branch you can discard. You build everything with Claude Code, so this is the one entry today that changes how the building itself happens rather than adding another thing to run. Edition 19 gave you ccmanager, which runs several assistants side by side; this gives each of them its own machine.
- An assistant cannot damage your working folder, because it never has it.
- Every environment leaves a git branch and a complete command log, so you can see exactly what happened.
- Apache-2.0, and it works with any assistant that speaks MCP.
- The project calls itself experimental and says so on its own front page. Expect changes between versions.
- It needs Docker and the Dagger engine, which is a real amount of machinery for one person's laptop.
- Several agents at once means several containers at once, and that costs disk, memory and model spend.
kbwo/ccmanagerRuns several assistant sessions side by side across git worktrees, which organises the people rather than sealing off the machine.
Track this in Scout- e2b-dev/E2B
Provides sandboxes for AI-generated code as a service with its own SDKs, aimed at code your product runs rather than code you review.
Track this in Scout - microsandbox/microsandbox
Runs untrusted code in small virtual machines rather than containers, which is stricter isolation and heavier to start.
Track this in Scout
# macOS or Linux: curl -fsSL https://raw.githubusercontent.com/dagger/container-use/main/install.sh | bash # or: brew install dagger/tap/container-use # Tell Claude Code about it: claude mcp add container-use -- cu stdio # Then, while an agent is working: cu list # which environments exist cu log <id> # every command that agent ran cu checkout <id> # bring its branch into your repository
