Claude Code as a workspace agent

workspace.agent could run opencode or codex. Now it can run Claude Code too.

What's new

• agent: claude. Pick it on workspace.create and the sandbox is built with Claude Code in it. Everything else is the same: workspace.agent hands it the job, continue: true carries on its session, and each call is kept as an agent session.
• On your own models. Claude Code speaks the Anthropic Messages API, so it takes an Anthropic model, or a model you host on Ollama, which answers on that API too. It stays on the one model you name — the smaller model it would reach for in the background is pointed at the same one.
• Your repository's settings stay out of it. Claude Code reads no .claude/settings.json or .mcp.json from the repository it is working in, so a cloned project cannot redirect where your key goes or start tools nobody chose.

Murmurator AI models aren't accepted by Claude Code yet, for the same reason codex refuses them: they are served on chat completions, and Claude Code doesn't speak it. Use opencode to run one.