CmuxLayer — Terminal Agent Lifecycle MCP
MCP server that wraps the cmux CLI into 21 typed tools for terminal and agent lifecycle management. Spawn, monitor, communicate with, and kill AI agents (Claude, Codex, Cursor, Gemini, Kiro) running in parallel terminal panes. 11 core surface tools for pane control, 8 agent lifecycle tools for spawn/wait/send/read, and a V2 facade that consolidates agent interaction into 2 high-level tools (interact + kill). Includes agent hierarchy tracking, quality metrics, sidebar sync, and security hardening against command injection.
cmuxlayer
0
MCP tools
0
Supported CLIs
0
Test assertions
0
API layers
Built with
Key features
21 MCP Tools, 3 Layers
11 core surface tools (split, read, send, send_key, notify, rename), 8 agent lifecycle tools (spawn_agent, stop_agent, send_to_agent, wait_for), and 2 V2 facade tools (interact + kill).
Terminal Orchestration for AI
Spawn Claude, Codex, Cursor, Gemini, or Kiro agents into terminal panes. Each gets a tracked state machine: spawning → booting → ready → working → done. Native MCP with 1,423x socket speedup.
Playwright Browser Surface
How it works
CLI Command
spawn / interact / kill
MCP Server
21 typed tools
cmux Socket
Terminal control
Agent Engine
State machine lifecycle
State Store
JSON file persistence
CLI Command
spawn / interact / kill
MCP Server
21 typed tools
cmux Socket
Terminal control
Agent Engine
State machine lifecycle
State Store
JSON file persistence
Get started
git clone https://github.com/EtanHey/cmuxlayer.git
cd cmuxlayer && npm install && npm run buildThe journey
01
Extracted from Orchestrator
CmuxLayer started as a vendored copy inside the golems orchestrator repo at tools/cmux-mcp. As the tool count grew past 10, the MCP server needed its own repo for independent testing, CI, and versioning.
02
10 Core Surface Tools
Built the foundation: list_surfaces, new_split, send_input, send_key, read_screen, rename_tab, set_status, set_progress, close_surface, and browser_surface. These wrap the cmux CLI socket into typed, Zod-validated MCP tools.
03
Agent Lifecycle Layer
Added 8 agent lifecycle tools — spawn_agent, stop_agent, send_to_agent, read_agent_output, wait_for, list_agents, get_agent_state. Each agent gets a deterministic state machine: spawning → booting → ready → working → done.
04
Native MCP Socket
Replaced CLI subprocess spawning with direct Unix domain socket communication. 1,423x speedup. Every tool call serializes to JSON, sends through the socket, and receives a typed response — no shell overhead.
05
V2 Facade + Screen Parser
Consolidated 8 agent operations into 2 tools: interact (7 action types) + kill (single/batch/all). Added Cursor screen parser with 12 regex patterns for agent_type, status, token_count, context_pct. 259 test assertions.
More projects
Cantaloupe AI - Recruitment Platform
Enterprise-grade AI-powered recruitment platform with voice interviews, resume screening, multi-channel communication, and seamless ATS/HRIS integrations. Automates hiring workflows for Quick Service Restaurants, healthcare, and construction industries.
Hand Sign Detection Model
A production-ready computer vision model that detects and classifies hands, arms, and non-hand objects in real-time with 96% accuracy.
BrainLayer — Persistent Memory for AI Agents
11 MCP tools, 312K+ indexed chunks, hybrid semantic+keyword search, knowledge graph with entity resolution, local LLM enrichment via Groq/MLX/Ollama. pip install brainlayer.