nc://acp

A Nodecode cell that runs a session in a coding agent: Claude Code, Codex, Gemini CLI, or any program that speaks the Agent Client Protocol on stdio. Pick acp/claude-code (or acp/codex, acp/gemini, acp/<your agent>) in /models or the page's model picker, and each turn of that session is one prompt to that agent.

The agent keeps the loop: its own tools, its own context, its own login. The session shows what it does:

• its words as the answer and its thinking (and its plan, as a checklist) as the turn's thinking;
• each call it makes as a row, in Nodecode's words where they fit (Read, Edited, Searched, Ran $ command), led by the agent's own words for it where it gives some: Claude Code's one-line description of each command (Run the test suite), a subagent's task, its title for a call Nodecode has no words for (Update TODOs: ...) - never the bare acp;
• what a call changed in the folder, as the row's diff, the same as a Nodecode tool's;
• its permission asks as the page's question card, with the agent's own options as the buttons (Allow / Always / Reject). No answer within ask_seconds is the agent's own "reject once";
• its token counts and, when the agent says so, its cost, on the turn's usage line and the Dashboard.

Stop cancels the agent's prompt (session/cancel); what it said and did up to then stays in the session. An agent that dies fails the turn with the last lines it wrote on stderr, and the next turn starts it again on the same session. The agent's session survives a restart of Nodecode too: the next turn loads it (session/load) where the agent offers that.

Install

Run nodecode add nc://acp, or copy this folder into ~/.nodecode/cells/nodecode-acp/ and restart Nodecode.

Each agent signs in with its own CLI, once, before its first turn here:

The two npx adapters need Node.js; the first turn fetches the adapter, which start_seconds allows for.

Configure

Optional. The presets above need no config; a top-level acp section adds an agent, changes a preset or takes one off:

{
  "acp": {
    "agents": {
      "claude-code": { "mode": "acceptEdits" },
      "opencode": { "command": "opencode", "args": ["acp"], "title": "OpenCode" },
      "gemini": { "enabled": false }
    },
    "ask_seconds": 600
  }
}

See config.example.jsonc.

What it does not do

• The agent's tools are its own: Nodecode's tools, MCP servers and context engine are not offered to it, and its turns are not compiled by Nodecode.
• The agent reads and writes files itself, under the session's folder; Nodecode does not offer it the client file system or terminal (both are false in initialize).
• A title or summary for the session is written by Nodecode's own side-call model, never by the agent.

Test

.github/test-cell.sh providers/nodecode-acp

The tests run a scripted agent, test/fake-agent.py, over the real wire and the real turn loop: words, thinking, a plan, a command, a write asked first and answered yes or no, a stop, a crash, a restart. No network, no model.