μ (mu)
AI agent harness with full local machine access. Wraps any OpenAI-compatible model behind Vercel AI SDK v6, executing tool calls directly on the host system with verbose step-by-step logging.
Built by Muse Mesh.
Power your agent
การติดตั้งจากผู้เผยแพร่ ไม่ใช่เอนด์พอยต์ของ PlugMCP
URL แบบรีโมตหรือคำสั่งบนเครื่องของคุณ? การติดตั้งสองแบบต่างกันอย่างไร
รันบน PlugMCP
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
เปิดจากผู้เผยแพร่
AI agent harness with full local machine access. Wraps any OpenAI-compatible model behind Vercel AI SDK v6, executing tool calls directly on the host system with verbose step-by-step logging.
Built by Muse Mesh.
Self-hosted AI agent framework with permission modes, audit logging and MCP. Runs on your machine, in Docker, or on a Raspberry Pi.
# Clone and install
git clone https://github.com/muse-mesh/mu.git && cd mu
pnpm install
# Configure
cp .env.example .env
# Edit .env — set MU_BOT_API_KEY at minimum
# Run interactively (CLI REPL)
pnpm dev
# Run with web UI
pnpm dev -- --web
# Open http://localhost:3141
git clone https://github.com/muse-mesh/mu.git && cd mu
corepack enable && corepack prepare [email protected] --activate
pnpm install
# Copy and configure environment
cp .env.example .env
# Edit .env with your API key and model settings
# Development mode (tsx, hot reload)
pnpm dev
# Production build
pnpm build
pnpm start # CLI mode
pnpm start:web # Web UI mode
cd mu
# 1. Configure environment (MU_BOT_API_KEY is the only required variable)
cp .env.example .env
# Edit .env — set MU_BOT_API_KEY at minimum, plus MU_BOT_API_BASE_URL and MU_BOT_MODEL if not using OpenAI
# 2. Build and start with docker compose (recommended)
docker compose up -d --build
# 3. Verify it's running
docker ps --filter name=mu
curl http://localhost:3141/api/health
# → {"status":"ok","uptime":...,"model":"..."}
The Web UI is available at http://localhost:3141.
To view logs: docker logs -f mu
To stop: docker compose down
Alternative — run without compose:
docker build -t mu:latest .
docker run -d \
--name mu \
-p 3141:3141 \
--env-file .env \
-e MU_BOT_WEB_UI_ENABLED=true \
mu:latest
The Docker image:
node:22-slim with common CLI tools (git, curl, wget, jq, ripgrep, tree, python3)mu user/app is read-only (owned by root)/ — full filesystem read access/home/mu/.mu (mount a volume to retain across restarts)GET /api/healthFor resource-constrained devices (1 GB RAM), run natively without Docker:
# On the Pi — install Node.js 22
curl -fsSL https://deb.nodesource.com/setup_22.x | sudo bash -
sudo apt-get install -y nodejs
# Transfer the built project (from your dev machine)
# On dev machine:
pnpm build
rsync -avz --exclude node_modules --exclude .git . user@pi-ip:~/mu
# On the Pi:
cd ~/mu
corepack enable && corepack prepare [email protected] --activate
pnpm install --frozen-lockfile --prod
# Configure
cp .env.example .env
# Edit .env
# Run
MU_BOT_WEB_UI_ENABLED=true node dist/index.js --web
# Access at http://<pi-ip>:3141
Optional: Create a systemd service for auto-start:
sudo tee /etc/systemd/system/mu.service > /dev/null << 'EOF'
[Unit]
Description=mu AI Agent
After=network.target
[Service]
Type=simple
User=pi
WorkingDirectory=/home/pi/mu
EnvironmentFile=/home/pi/mu/.env
ExecStart=/usr/bin/node /home/pi/mu/dist/index.js --web
Restart=on-failure
RestartSec=5
[Install]
WantedBy=multi-user.target
EOF
sudo systemctl enable mu
sudo systemctl start mu
Configuration loads from multiple sources (highest priority first):
--model, --web, etc.)MU_BOT_*).mu.json in cwd or ~/.mu.json)Copy .env.example to .env and customise:
Create .mu.json in your project root or ~/.mu.json:
{
"apiBaseUrl": "http://localhost:11434/v1",
"model": "llama3",
"maxSteps": 100,
"permissionMode": "default"
}
mu [prompt] [options]
Options:
-m, --model <model> Model identifier
-b, --base-url <url> API base URL
-k, --api-key <key> API key
-s, --max-steps <n> Max tool loop steps
-v, --verbose Verbose output
-d, --debug Debug output
-w, --web Enable web UI
-p, --port <port> Web UI port (default: 3141)
--permission <mode> Permission mode
--output <format> Output format: text|json|ndjson
--tools <tools> Comma-separated list of enabled tools
-c, --config <path> Config file path
Pass a prompt as an argument — mu executes the task and exits:
pnpm dev -- "list the files in the current directory"
pnpm dev -- -m gpt-4o "explain this codebase"
pnpm dev -- -s 5 "find all TODO comments in src/"
pnpm dev -- --tools file_read,grep "search for imports in src/"
Run without a prompt for multi-turn conversation:
pnpm dev
Browser-based interface with real-time streaming:
pnpm dev -- --web
# Open http://localhost:3141
Features:
Machine-readable output for scripting:
pnpm dev -- --output ndjson "echo hello" | jq '.type'
mu ships with 11 built-in tools:
# Read-only exploration
pnpm dev -- --tools file_read,grep,glob,list_dir "explore this codebase"
# File operations only
pnpm dev -- --tools file_read,file_write,file_edit "refactor the config module"
mu works with any OpenAI-compatible API endpoint:
# OpenAI
MU_BOT_API_BASE_URL=https://api.openai.com/v1
MU_BOT_MODEL=gpt-4o
# Ollama (local)
MU_BOT_API_BASE_URL=http://localhost:11434/v1
MU_BOT_MODEL=llama3
# OpenRouter
MU_BOT_API_BASE_URL=https://openrouter.ai/api/v1
MU_BOT_MODEL=anthropic/claude-sonnet-4-20250514
# LM Studio
MU_BOT_API_BASE_URL=http://localhost:1234/v1
MU_BOT_MODEL=local-model
# Mume Gateway
MU_BOT_API_BASE_URL=https://mume.ai/api/v1
MU_BOT_MODEL=z-ai/glm-5v-turbo
Connect external MCP servers for additional tools:
{
"mcpServers": [
{
"name": "my-tools",
"transport": "stdio",
"config": { "command": "my-mcp-server", "args": [] }
},
{
"name": "remote",
"transport": "sse",
"config": { "url": "https://mcp.example.com/sse" }
}
]
}
Set via .mu.json or the MU_BOT_MCP_SERVERS env var (JSON string). MCP tools are namespaced as mcp__{server}__{tool}.
cp .env.example .env # configure API key & model
docker compose up -d # builds, starts on port 3141
docker compose logs -f # watch logs
docker compose down # stop
# Prerequisites: Node.js 22, pnpm
# Transfer built artifacts, install prod deps, run
# See "Raspberry Pi / ARM64" under Installation for full steps
When deployed in Docker:
/app) is owned by root:root with 755 permissions — the agent user (mu) can read but cannot modify its own code/ — the agent can browse and read the entire filesystem/home/mu/.mu (logs, sessions, memory)mu userEvery session creates a JSONL audit log at ~/.mu/logs/{sessionId}.jsonl:
# View the latest log
cat ~/.mu/logs/$(ls -t ~/.mu/logs/ | head -1)
# Filter for tool calls
cat ~/.mu/logs/*.jsonl | jq 'select(.type == "tool_call_start")'
Event types: session_start, session_end, user_message, step_start, step_finish, tool_call_start, tool_call_finish, model_response.
src/
├── index.ts # CLI entry point (Commander)
├── agent.ts # Agent loop (streamText + tools)
├── config.ts # Zod-validated config loader
├── logger.ts # Pino structured logger + JSONL audit
├── state.ts # Reactive state store
├── types.ts # Shared types (MuToolDef, LogEvent, etc.)
├── shutdown.ts # Graceful SIGTERM/SIGINT shutdown
├── cost.ts # Cost tracking with model pricing
├── retry.ts # Retry with exponential backoff
├── context.ts # Context window sizes + token estimation
├── cli/
│ ├── renderer.ts # Terminal output formatter
│ └── repl.ts # Interactive REPL mode
├── compaction/
│ ├── index.ts # Context compaction pipeline
│ └── prompts.ts # Compaction & memory extraction prompts
├── memory/
│ └── index.ts # 3-tier memory system (MEMORY.md, entries, session)
├── tools/
│ ├── index.ts # Tool registry + execution pipeline
│ ├── build-tool.ts # buildTool() factory with safe defaults
│ ├── shell-exec.ts # Shell command execution
│ ├── file-read.ts # File reading with line ranges
│ ├── file-write.ts # File writing
│ ├── file-edit.ts # Find-and-replace editing
│ ├── glob.ts # File glob matching
│ ├── grep.ts # Regex search (ripgrep + fallback)
│ ├── list-dir.ts # Directory listing
│ ├── http-fetch.ts # HTTP requests
│ ├── system-info.ts # System information
│ ├── think.ts # Reasoning scratchpad
│ └── task-complete.ts # Task completion signal
├── permissions/
│ ├── index.ts # 4-mode permission system
│ └── prompt.ts # CLI permission prompting
├── mcp/
│ └── client.ts # MCP server client
└── web/
├── server.ts # Hono API server + static serving
└── sessions.ts # Session store (JSON file-backed)
web/ # React frontend (Vite)
├── src/
│ ├── App.tsx # Main app component (useChat hook)
│ ├── main.tsx # React entry point
│ └── app.css # Styles (dark/light themes)
├── index.html
└── vite.config.ts
# Dev mode (tsx, auto-reload)
pnpm dev
# Dev with web UI
pnpm dev -- --web
# Frontend dev server (hot reload, proxied to backend)
pnpm dev:web
# Type-check
pnpm exec tsc --noEmit
# Production build (frontend + backend)
pnpm build
# Production run
pnpm start # CLI mode
pnpm start:web # Web UI mode (port 3141)
Apache 2.0 — see LICENSE for details.
Copyright 2025 Muse Mesh.
Self-hosted AI agent framework with permission modes, audit logging and MCP. Runs on your machine, in Docker, or on a Raspberry Pi.
เครื่องมือ
ยังไม่มีรายการเครื่องมือ เซิร์ฟเวอร์แบบรีโมตจะได้รายการหลังการตรวจสอบสถานะครั้งแรก
FAQ
คัดลอกคอนฟิก stdio สำหรับ Cursor, Claude Desktop, Claude Code หรือ VS Code ได้ในหน้านี้ คอนฟิกนี้ใช้การติดตั้งของผู้เผยแพร่ ไม่ใช่พร็อกซีของ PlugMCP
ระเบียนนี้ไม่ได้ระบุตัวแปรสภาพแวดล้อมที่จำเป็น
Mu ทำงานแบบโลคัลผ่าน stdio
ตรวจสอบล่าสุดเมื่อ 2026-10-28 Self-hosted AI agent framework with permission modes, audit logging and MCP. Runs on your machine, in Docker, or on a Raspberry Pi.