Supports browser automation capabilities by allowing the execution of JavaScript for DOM evaluation and interaction within virtual browser surfaces.
Click on "Install Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@cmux-agent-mcpOpen a new workspace and start Claude Code to refactor the login flow."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
cmux-agent-mcp
CMUX MCP Server — a programmable terminal control plane for multi-agent AI workflows.
Turns CMUX into a remote-controllable terminal multiplexer that any AI coding CLI can be orchestrated through. One orchestrator agent spawns, monitors, and communicates with any number of AI agents running in parallel across multiple projects.
What This Does
Spawn AI agents in CMUX workspaces — Claude Code, Gemini CLI, Codex CLI, OpenCode, Goose
Inject prompts into running agent sessions as if a human typed them
Read output from any pane — passive (fast) or deep (asks agents for status)
Orchestrate — send different plans to different agents in one call
Manage workspaces — create, rename, reorder, move between windows
Split panes — horizontal, vertical, grid layouts, drag-to-split
Browser automation — open URLs, navigate, snapshot DOM, evaluate JS, click/fill/type
Sidebar metadata — status pills, progress bars, log entries
Notifications — send alerts with blue ring indicators
Session recovery — save/restore full layouts including CLI session IDs and conversations
Auto-skip permissions — each CLI's autonomous mode is handled automatically
Auto-save — sessions are saved automatically after every layout change
Architecture
+---------------------------------------------------+
| Your AI Agent (Claude, etc.) |
| |
| "Launch 4 agents and distribute tasks" |
| | |
| MCP Tool Calls |
| | |
+---------------------------------------------------+
| cmux-agent-mcp |
| (this MCP server) |
| | |
| cmux CLI commands |
| | |
+---------------------------------------------------+
| CMUX |
| |
| +-- Sidebar ----+ +-- Tab Bar ---------------+ |
| | agents | | Claude 1 x | ~/p/sensei | |
| | > project-a | +-----------------------------+|
| | project-b | | +--------+ +--------+ | |
| | outputs | | | Claude | | Gemini | | |
| | edit | | | Agent | | Agent | | |
| +---------------+ | +--------+ +--------+ | |
| | | Codex | | Shell | | |
| | | Agent | | | | |
| | +--------+ +--------+ | |
| +-----------------------------+
+---------------------------------------------------+Quick Start
Prerequisites
Install
One command:
npm install -g cmux-agent-mcpInstall automatically registers the MCP server globally for all AI coding tools:
Config file | For |
| Claude Code, Codex |
| Cursor |
VS Code | VS Code |
| Gemini CLI |
| OpenCode |
One install, every project, every tool. Existing config files are merged — other MCP servers won't be touched.
To re-run manually or for per-project setup:
cmux-agent-mcp init # re-run global setup
cmux-agent-mcp init --project # per-project setup (current dir)
cmux-agent-mcp init --root /path # per-project setup (specific dir)Install from source
git clone https://github.com/multiagentcognition/cmux-agent-mcp.git
cd cmux-agent-mcp
npm install --ignore-scripts
npm run build
node build/cli.js init # register globally81 MCP Tools
Status & Discovery
Tool | Description |
| Check if CMUX is installed/running, show hierarchy summary |
| Full hierarchy tree of windows, workspaces, panes, surfaces |
| Context info for focused window/workspace/pane/surface |
| Search across windows and panes by content or title |
| Launch CMUX if not running |
| Capture screenshot of CMUX window |
Workspace Management
Tool | Description |
| List all open workspaces |
| Get the currently active workspace |
| Create a new workspace with optional cwd/command |
| Switch to a specific workspace |
| Close a workspace and all its panes |
| Rename a workspace (changes the sidebar name) |
| Reorder a workspace in the sidebar |
| Move a workspace to a different window |
Window Management
Tool | Description |
| List all open windows |
| Get the currently focused window |
| Create a new window |
| Focus a specific window |
| Close a window |
| Rename a window (changes the title bar) |
Surface (Tab) Management
Tool | Description |
| Create a new tab (terminal or browser) |
| Close a tab |
| Rename a tab (changes the tab bar name) |
| Move a tab to a different pane/workspace/window |
| Reorder a tab within its pane |
| Turn a tab into its own split pane |
Pane / Split Operations
Tool | Description |
| List all panes in a workspace |
| List pane surfaces with refs |
| List all panels in a workspace |
| Split a pane (left/right/up/down) |
| Create a new pane |
| Focus a specific pane |
| Resize a pane |
| Swap two panes |
| Move a pane to its own workspace |
| Merge a pane into another |
| Restart a pane's process |
Text I/O
Tool | Description |
| Send text without pressing Enter |
| Send text and press Enter |
| Send a key press (enter, ctrl+c, escape, etc.) |
| Send text to a specific panel |
| Read terminal output from a surface |
| tmux-compatible capture |
Bulk Text Operations
Tool | Description |
| Send same text + Enter to ALL panes |
| Send DIFFERENT text to each pane (in order) |
| Send text to SPECIFIC surfaces only |
| Send a key to ALL panes (e.g., ctrl+c to cancel all) |
| Read output from all panes |
| Deep read — prompts idle agents for status |
| Full snapshot: tree + all output + sidebar state |
Sidebar Metadata
Tool | Description |
| Set a sidebar status pill (key-value badge) |
| Clear a status pill |
| List all status entries |
| Set a progress indicator (0.0-1.0) |
| Clear the progress indicator |
| Write a log entry to the sidebar |
| Get full sidebar state |
Notifications
Tool | Description |
| Send a notification (blue ring + sidebar highlight) |
| List all notifications |
| Clear all notifications |
Browser Automation
Tool | Description |
| Open a browser surface |
| Navigate: goto, back, forward, reload |
| DOM accessibility snapshot |
| Take a page screenshot |
| Execute JavaScript |
| Click an element |
| Fill an input field |
| Type text into an element |
| Wait for a condition (selector, text, URL, load state) |
| Get page data (url, title, text, html, value, attr, count) |
| Manage browser tabs (new, list, switch, close) |
| Get/clear console logs and errors |
High-Level Launchers
Tool | Description |
| Open a single AI CLI in a new or existing workspace |
| Create workspace with N agents in a grid layout |
| Create an exact rows x cols grid of panes |
| Launch different CLIs in one workspace |
| Send different prompts to specific surfaces in one call |
| Close all workspaces |
Session Management
Tool | Description |
| Save current state for crash recovery |
| Recreate layout and resume CLI conversations |
| Compare manifest vs live state, report drift |
Orchestration Workflow
The typical multi-agent workflow:
1. cmux_launch_agents(cli: "claude", count: 4)
→ Creates a workspace with 4 Claude Code agents in a 2x2 grid
2. cmux_orchestrate(assignments: [
{ surface: "surface:10", text: "Implement the auth module..." },
{ surface: "surface:11", text: "Write tests for the API..." },
{ surface: "surface:12", text: "Refactor the database layer..." },
{ surface: "surface:13", text: "Update the documentation..." },
])
→ Each agent receives its specific task
3. cmux_read_all()
→ Check all agents' progress at once
4. cmux_read_all_deep()
→ Ask idle agents to summarize their statusMixed-CLI orchestration:
cmux_launch_mixed(agents: [
{ cli: "claude", label: "Architect" },
{ cli: "gemini", label: "Reviewer" },
{ cli: "codex", label: "Implementer" },
])Session Recovery
Sessions are auto-saved after every layout change. If CMUX crashes or you restart your machine, call cmux_session_recover to recreate everything — including resuming CLI conversations.
CLI Session Resume
CLI | Resume Method |
Claude Code |
|
Gemini CLI |
|
Codex |
|
OpenCode |
|
Goose |
|
Session IDs are detected from each CLI's session storage:
Claude:
~/.claude/sessions/*.jsonCodex:
~/.codex/sessions/**/rollout-*.jsonlGemini:
~/.gemini/projects.json+~/.gemini/tmp/*/chats/*.jsonOpenCode:
~/.local/share/opencode/opencode.db(SQLite)Goose:
goose session list --format json
Auto-Trust & Permissions
When launching agents, cmux-agent-mcp automatically:
Claude Code: Sets
hasTrustDialogAccepted: truein~/.claude.jsonGemini CLI: Adds directory to
~/.gemini/trustedFolders.jsonCodex: Adds
trust_level = "trusted"to~/.codex/config.tomlOpenCode: Sets
permission: 'allow'in its config
No permission prompts — agents start working immediately.
ID Reference Format
CMUX uses ref format for IDs: workspace:5, surface:8, pane:3, tab:2, window:1. Always use the ref format, not bare numbers. Use cmux_identify or cmux_list_pane_surfaces to discover refs.
CMUX Hierarchy
Window → Workspace(s) → Pane(s) → Surface(s) → Panel(s)Window: macOS window
Workspace: Sidebar entry (like a "tab group")
Pane: Split region within a workspace
Surface: Tab within a pane (terminal or browser)
Panel: Content inside a surface
Naming Conventions
What you see | What to rename |
Sidebar entry |
|
Tab at top |
|
Window title bar |
|
Platform Support
macOS only — CMUX is a native macOS application.
Testing
The test/full-test-suite.md file contains a comprehensive test plan covering all 81 tools. Give it to any AI coding agent to execute:
Tag test/full-test-suite.md and say "Execute this test suite"Tests leave all workspaces open for human visual verification before cleanup.
License
This server cannot be installed
Resources
Unclaimed servers have limited discoverability.
Looking for Admin?
If you are the server author, to access and configure the admin panel.