Add an agent to the dev image

The dev image already ships Claude Code, jcode, opencode, OpenChamber, and the GitHub Copilot CLI. This guide shows the pattern for adding another, using the conventions the existing five follow.

The pattern

Every agent addition has three parts: a pinned version as a build argument, an install step using the upstream-documented method, and a row in the README’s agents table covering the entry point and auth.

1. Add a version argument to the dev stage in sandbox/Dockerfile.

   ARG MYAGENT_VERSION=1.2.3

2. Install with the upstream method. Prefer a checksum-verified release binary for young projects, and npm with an exact pin for npm-distributed tools.

   RUN npm install -g --no-fund --no-audit "myagent-cli@${MYAGENT_VERSION}"

   For a release binary, follow the jcode block in the same file: download the tarball and the published checksums, verify with sha256sum -c, then extract into /usr/local/bin.

3. Rebuild the dev image and confirm the entry point resolves.

   docker compose --profile dev build dev
   docker compose --profile dev run --rm dev myagent --version

4. Document it. Add a row to the coding agents table in the README with the entry point, install source, and auth steps.

Caution

The dev container runs privileged for docker-in-docker. Installing an unreviewed third-party agent into it is a real trust decision, not a formality. Keep additions out of the base sandbox stage, pin exact versions so upgrades are deliberate edits, and prefer running new agents with the firewall applied.

Where things go

Agents belong in the dev stage only. The base sandbox stage is the execution contract for Anthropic sessions (bash, Bun, Python, the ant CLI) and stays minimal so session cold starts stay fast.