Acp Cockpit — AI skill for Claude Code

acp-cockpit

A **browser client for AI coding agents** that speaks the [Agent Client Protocol](https://github.com/agentclientprotocol/agent-client-protocol) directly — no terminal, no TUI, nothing scraped off a screen. Windows-native (no WSL, no pty), agent-agnostic, and built around one guarantee: *nothing that crosses the wire is lost or silently misread*. The first configured agent is Claude Code, through the [`claude-agent-acp`](https://github.com/agentclientprotocol/claude-agent-acp) adapter.

The client shows itself as **ClaudIU** out of the box — that name is one line of configuration (`uiname.toml`), yours to change. This project is not affiliated with, endorsed by, or sponsored by Anthropic or Zed Industries.


*The model's own thinking, then its answer — with the tool, event and harness lanes switched off (struck through, bottom left) so that only the conversation is left. Nothing is hidden by default; you decide what is on the page.*

What you get

**Nothing is left out, because nothing is redrawn.** A terminal client shows you a screen; whatever is not on the screen is gone, and a client that scrapes one is guessing. acp-cockpit is an ACP client: the conversation arrives as structured JSON-RPC, is written to a per-session record *before* it is interpreted, and is then rendered as itself. The model's **thinking**, tool calls with their **diffs**, **permission requests** with the agent's own typed options, the agent's **own questions** as forms, **plans**, **subagent transcripts**, slash commands, token usage, cost, and your account's rate-limit windows — all of it is data, and all of it is on the page. Anything the client does not recognise becomes a visible `unrecognized` element rather than a gap.

**The view holds still.** This is a reading tool, so it behaves like