Moonbridge

🚀 Moonbridge

Your MCP client just got a team. Spawn AI coding agents from Claude Code, Cursor, or any MCP client. Run 10 approaches in parallel for a fraction of the cost.

uvx moonbridge

🚀 Quick Start

Step 1: Install at least one supported CLI

Step 2: Add to MCP config (~/.mcp.json)

{
  "mcpServers": {
    "moonbridge": {
      "type": "stdio",
      "command": "uvx",
      "args": ["moonbridge"]
    }
  }
}

Step 3: Use it

Your MCP client now has spawn_agent and spawn_agents_parallel tools.

⚠️ Security Warning (Read First)

Moonbridge executes agentic CLIs (Kimi/Codex/OpenCode/Gemini). A malicious or careless prompt can cause an agent to run shell commands, read accessible files, or exfiltrate data via network calls.

Moonbridge adds guardrails (MOONBRIDGE_ALLOWED_DIRS, environment allowlists, optional MOONBRIDGE_SANDBOX=1), but these are not equivalent to OS-level containment.

For untrusted prompts or shared environments, run Moonbridge inside a container or VM with least-privilege filesystem and network access.

🔄 Updating

Moonbridge checks for updates on startup (cached for 24h). To update manually:

# If using uvx (recommended)
uvx moonbridge --refresh

# If installed as a tool
uv tool upgrade moonbridge

Disable update checks for CI/automation:

export MOONBRIDGE_SKIP_UPDATE_CHECK=1

💡 When to Use Moonbridge

Best for: Tasks that benefit from parallel execution or volume.

🔍 How it Works

Connection Flow

1. MCP client (Claude Code, Cursor, etc.) connects to Moonbridge over stdio.
2. Client discovers available tools via list_tools.
3. Client calls spawn_agent or spawn_agents_parallel.

Spawn Process

1. Moonbridge validates the prompt and working directory.
2. Resolves which adapter to use (Kimi, Codex, OpenCode, Gemini).
3. Adapter builds the CLI command with appropriate flags.
4. Spawns subprocess in a separate process group.
5. Captures stdout/stderr, enforces timeout.
6. Returns structured JSON result.

Parallel Execution

• spawn_agents_parallel runs up to 10 agents concurrently via asyncio.gather.
• Each agent is independent (separate process, separate output).
• All results returned together when the last agent finishes (or times out).

MCP Client → stdio → Moonbridge → adapter → CLI subprocess
                                          → CLI subprocess (parallel)
                                          → CLI subprocess (parallel)

🛠️ Tools

Example: Parallel Exploration

{
  "agents": [
    {"prompt": "Refactor to React hooks"},
    {"prompt": "Refactor to Zustand"},
    {"prompt": "Refactor to Redux Toolkit"}
  ]
}

Three approaches. One request. You choose the winner.

Tool Parameters

spawn_agent

spawn_agents_parallel

check_status

list_models

📄 Response Format

All tools return JSON with these fields:

When output is too large, Moonbridge truncates it and adds raw.output_limit metadata with original sizes.

⚙️ Configuration

Environment Variables

🔒 Security

1. Threat Model (Prompt Injection Included)

If an attacker can influence prompt input sent through MCP, they can attempt to make the agent:

• read sensitive files (for example ~/.ssh or .env),
• run destructive shell commands,
• exfiltrate data over the network.

Moonbridge does not inspect prompt intent. Treat prompt input as potentially untrusted.

2. Directory Restrictions (MOONBRIDGE_ALLOWED_DIRS)

Default: agents can operate in any directory. Set MOONBRIDGE_ALLOWED_DIRS to restrict: colon-separated allowed paths. Symlinks resolved via os.path.realpath before checking. Strict mode (MOONBRIDGE_STRICT=1) exits on startup if no valid allowed directories are configured.

export MOONBRIDGE_ALLOWED_DIRS="/home/user/projects:/home/user/work"
export MOONBRIDGE_STRICT=1  # require restrictions

3. Environment Sanitization

Only whitelisted env vars are passed to spawned agents. Each adapter defines its own allowlist (PATH, HOME, plus adapter-specific like OPENAI_API_KEY for Codex). Your shell environment (secrets, tokens, SSH keys) is not inherited by default.

4. Input Validation

Model parameters are validated to prevent flag injection (values starting with - are rejected). Prompts are capped at 100,000 characters and cannot be empty.

5. Process Isolation and Sandbox Mode

Agents run in separate process groups (start_new_session=True). Orphan cleanup on exit. Sandbox mode available (MOONBRIDGE_SANDBOX=1) for copy-on-run isolation.

Not OS-level sandboxing. Agents can still read or write arbitrary host files they can access.

6. Hardened Deployment Checklist

• Set MOONBRIDGE_ALLOWED_DIRS to the smallest possible set.
• Enable MOONBRIDGE_STRICT=1 so missing restrictions fail closed.
• Enable MOONBRIDGE_SANDBOX=1 to avoid direct workspace mutation.
• Run Moonbridge in a container/VM for strong isolation.
• Do not expose Moonbridge to untrusted clients without additional auth controls.

🐞 Troubleshooting

"CLI not found"

Install the CLI for your chosen adapter:

# Kimi
uv tool install --python 3.13 kimi-cli
which kimi

# Codex
npm install -g @openai/codex
which codex

# OpenCode
curl -fsSL https://opencode.ai/install | bash
which opencode

# Gemini
npm install -g @google/gemini-cli
which gemini

"auth_error" responses

Authenticate with your chosen CLI:

# Kimi
kimi login

# Codex
export OPENAI_API_KEY=sk-...

# OpenCode
opencode auth login

# Gemini
gemini  # complete login flow, or set GEMINI_API_KEY

Timeout errors

Adapters have sensible defaults: Codex = 1800s, Kimi = 600s, OpenCode = 1200s, Gemini = 1200s.

For exceptionally long tasks, override explicitly:

{"prompt": "...", "timeout_seconds": 3600}

Or set per-adapter defaults via environment:

export MOONBRIDGE_CODEX_TIMEOUT=2400  # 40 minutes
export MOONBRIDGE_KIMI_TIMEOUT=900    # 15 minutes
export MOONBRIDGE_OPENCODE_TIMEOUT=1200  # 20 minutes
export MOONBRIDGE_GEMINI_TIMEOUT=1200  # 20 minutes

⏱️ Timeout Best Practices

Priority resolution: explicit param > adapter env > adapter default > global env > 600s fallback

"MOONBRIDGE_ALLOWED_DIRS is not set" warning

By default, Moonbridge warns at startup if no directory restrictions are configured. This is expected for local development. For shared/production environments, set allowed directories:

export MOONBRIDGE_ALLOWED_DIRS="/path/to/project:/another/path"

📦 Sandbox Mode (Copy-on-Run)

Enable sandbox mode to run agents in a temporary copy of the working directory:

export MOONBRIDGE_SANDBOX=1

When enabled:

• Agents run in a temp copy of cwd.
• Host files stay unchanged by default.
• A unified diff + summary is included in raw.sandbox.

Optional:

export MOONBRIDGE_SANDBOX_KEEP=1       # keep temp dir
export MOONBRIDGE_SANDBOX_MAX_DIFF=200000
export MOONBRIDGE_SANDBOX_MAX_COPY=300000000

Limitations: this is not OS-level isolation. Agents can still read/write arbitrary host paths if they choose to. Use containers/VMs for strong isolation.

To enforce restrictions (exit instead of warn):

export MOONBRIDGE_STRICT=1

Permission denied on working directory

Verify the directory is in your allowlist:

export MOONBRIDGE_ALLOWED_DIRS="/path/to/project:/another/path"

Debug logging

Enable verbose logging:

export MOONBRIDGE_LOG_LEVEL=DEBUG

🖥️ Platform Support

macOS and Linux only. Windows is not supported.

📄 License

MIT. See LICENSE.