With this, your mind can hand a coding job to Codex, Claude Code or opencode, each in a Linux sandbox of its own, and get the result back as a commit.

Why. The mind's own agents are good at running scripts against the mind. Coding agents are good at reading and changing a lot of code. We wanted both, under one list.

How it works. One call starts a session and answers its path. Then you message it like any other agent:

const path = await itx.acp.sandbox.start("codex"); // or "claude", "opencode"
await itx.agents.get(path).message("Fix the failing test in apps/docs.");

The agents speak the Agent Client Protocol (ACP), so one runner drives all three. The sandbox holds a clone of the config repo and two tools: run a script against the mind, and commit to its config repo. A commit names its parent, so an agent that is slower than another one is refused and tries again, instead of undoing someone's work.

Our files:

  • acp-sandbox.ts: the module behind itx.acp.sandbox; start(agent, model?).
  • acp-sandbox/runner.ts: starts a session in a sandbox and delivers each message to it.
  • acp-sandbox/host.ts (and its bundled host.generated.ts): what runs inside the sandbox and talks ACP to the agent.
  • acp-sandbox/worker.ts: the integration, one entry in the integrations list of worker.ts.
  • sandbox-tools/: the clone of the repo and the two mind tools.
  • agent-kinds/acp.ts: how a session shows up, and chats, in the Agents app.
  • claude-sandbox.ts and claude-sandbox/: the same for Claude Code's own headless mode.

What it needs. Sandboxes are Cloudflare Containers in your own account, billed there, and they sleep when idle. Each agent needs its own model account: a ChatGPT plan or an OpenAI API key for Codex (/secrets/chatgpt or /secrets/openai), a Claude account for Claude Code.

What surprised us. The list is the feature. Once every agent is on one page, "what is everyone working on?" is a page you open.