feat: add cmux MCP server for terminal control

README.md

@@ -107,6 +107,59 @@ The in-app browser has a scriptable API ported from [agent-browser](https://gith
 
 Everything is scriptable through the CLI and socket API — create workspaces/tabs, split panes, send keystrokes, open URLs in the browser.
 
+## MCP server
+
+cmux now includes an initial local MCP server for external agents that want to steer cmux programmatically instead of shelling out directly.
+
+What it wraps today:
+- `system.identify` and `tree` for discovery
+- window, workspace, pane, and surface control through the existing `cmux` CLI
+- terminal interaction via `read-screen`, `send`, and `send-key`
+- a first-pass browser wrapper for common `cmux browser ...` actions
+
+The server intentionally reuses the existing CLI and socket surface instead of introducing a second control plane, so MCP clients get the same handle model (`window:N`, `workspace:N`, `pane:N`, `surface:N`) and routing behavior that cmux already documents.
+
+Install the Node dependencies once:
+
+```bash
+npm install
+```
+
+Run over stdio:
+
+```bash
+npm run cmux:mcp
+```
+
+Run over HTTP:
+
+```bash
+npm run cmux:mcp -- --transport http --host 127.0.0.1 --port 8765
+```
+
+Useful environment variables:
+
+```bash
+export CMUX_MCP_CMUX_BIN=/path/to/cmux
+export CMUX_MCP_SOCKET_PATH=/tmp/cmux.sock
+export CMUX_MCP_SOCKET_PASSWORD=...
+export CMUX_MCP_ID_FORMAT=refs
+```
+
+Available MCP tools:
+- `cmux_identify`
+- `cmux_tree`
+- `cmux_list`
+- `cmux_control`
+- `cmux_terminal`
+- `cmux_browser`
+
+Term mapping for MCP clients:
+- `window` is a native macOS cmux window
+- `workspace` is the sidebar tab-like container
+- `pane` is a split region inside a workspace
+- `surface` is a tab inside a pane, usually the most stable automation target
+
 ## The Zen of cmux
 
 cmux is not prescriptive about how developers hold their tools. It's a terminal and browser with a CLI, and the rest is up to you.

docs/cmux-mcp-server.md

@@ -0,0 +1,111 @@
+# cmux MCP Server Recommendation
+
+Last updated: March 7, 2026
+
+## Recommendation
+
+Build a local MCP server for cmux, but keep it as a thin adapter over the existing v2 Unix socket API instead of adding a second control plane inside the app.
+
+This is justified because cmux already exposes most of the primitives an MCP host would need:
+
+1. `TerminalController` already implements a JSON v2 socket protocol with stable methods for `window.*`, `workspace.*`, `pane.*`, `surface.*`, and `browser.*` in [TerminalController.swift](../Sources/TerminalController.swift).
+2. `system.tree` and `system.identify` already provide the discovery and self-location data an agent needs to reason about windows, workspaces, panes, and surfaces.
+3. `TabManager`, `Workspace`, and `AppDelegate` already own the real mutations for selection, creation, moving, focusing, and cross-window routing, so an MCP server does not need to replicate that logic.
+
+## Codebase Findings
+
+### Existing control plane
+
+1. `TerminalController` is already the automation boundary:
+   - capability advertisement via `v2Capabilities()`
+   - hierarchy discovery via `v2SystemTree()`
+   - window/workspace/surface/pane/browser dispatch in the `window.*`, `workspace.*`, `surface.*`, `pane.*`, and `browser.*` handlers
+2. `AppDelegate` already exposes window-level lookup and mutation:
+   - `listMainWindowSummaries()`
+   - `tabManagerFor(windowId:)`
+   - `focusMainWindow(windowId:)`
+   - `createMainWindow()`
+   - `closeMainWindow(windowId:)`
+3. `TabManager` already owns workspace lifecycle and selection:
+   - `addWorkspace()`
+   - `selectWorkspace(_:)`
+   - `reorderWorkspace(...)`
+   - workspace title/color/pin operations
+4. `Workspace` already owns pane and surface topology:
+   - bonsplit pane state
+   - panel registry
+   - focused surface tracking
+   - terminal/browser surface metadata
+
+### Design implication
+
+The app already has a useful "cmux control kernel". The missing piece is a host-friendly MCP facade, not more app-side orchestration code.
+
+## Proposed Architecture
+
+Run the MCP server as an external local stdio process:
+
+1. MCP host starts `scripts/cmux_mcp_server.py`.
+2. The server connects to the configured cmux socket (`CMUX_SOCKET_PATH` or `--socket`).
+3. MCP tools translate directly to v2 socket calls.
+4. `TerminalController` remains the source of truth for capability routing and object identity.
+
+Why external is better than embedding inside cmux:
+
+1. MCP is a transport/protocol concern; cmux already has the stateful control API.
+2. External stdio avoids mixing MCP lifecycle with AppKit lifecycle.
+3. The same adapter can target different cmux sockets without changing the app.
+4. The adapter can stay intentionally small and experimental while the socket API evolves.
+
+## Initial Tool Set
+
+Recommend exposing a narrow first pass instead of mirroring every socket method one-for-one.
+
+### Discovery
+
+1. `cmux_socket_discover`
+   - Find candidate local cmux sockets under `/tmp`.
+2. `cmux_system_tree`
+   - Return windows, workspaces, panes, surfaces, and the active path.
+
+### Workspace steering
+
+1. `cmux_list_workspaces`
+2. `cmux_create_workspace`
+3. `cmux_select_workspace`
+
+### Surface steering
+
+1. `cmux_list_surfaces`
+2. `cmux_send_text`
+3. `cmux_read_text`
+
+### Escape hatch
+
+1. `cmux_socket_call`
+   - Raw pass-through to an existing v2 socket method.
+   - Useful while validating which abstractions should become first-class MCP tools.
+
+## What Not To Build Yet
+
+1. Do not duplicate the entire `window.*` / `workspace.*` / `surface.*` / `browser.*` matrix as separate MCP tools yet.
+2. Do not add a second IPC path inside cmux just for MCP.
+3. Do not invent new object IDs; reuse cmux UUID/ref handles from the socket API.
+4. Do not let the MCP layer bypass `TerminalController` focus rules or socket auth/access mode.
+
+## Risks and Guardrails
+
+1. Focus-stealing remains an app concern. The MCP adapter should respect the focus policies already enforced by `TerminalController`.
+2. Socket auth mode matters. If the socket is password-protected or disabled, MCP should fail clearly instead of trying to bypass it.
+3. Tool count can explode quickly. Keep MCP tools task-shaped and leave uncommon operations behind `cmux_socket_call` until repeated usage proves they deserve dedicated schemas.
+
+## Recommendation Summary
+
+Yes, build the MCP server.
+
+But build it as:
+
+1. an external stdio adapter
+2. backed by the existing v2 socket API
+3. with a small, high-value tool surface first
+4. and a raw-call escape hatch while the tool taxonomy settles