Add MCP Capability

Add MCP Capability to Jarvis

This skill guides the process of adding new MCP (Model Context Protocol) server capabilities to the Jarvis Signal AI assistant.

Overview

Jarvis loads MCP servers from data/mcp.json. MCP servers are installed as npm packages in data/node_modules/ and configured to run when the agent starts. The agent then has access to the tools provided by those MCP servers.

Key Architecture Points

• Config file: data/mcp.json defines which MCP servers to load
• Package location: MCP servers installed in data/node_modules/
• Environment: Same data/ directory is mounted in both local dev and Docker
• Env var substitution: Config supports ${VAR_NAME} syntax for secrets

Adding an MCP Server

Step 1: Research the MCP Server

Before installing, understand what the MCP server provides:

• What tools does it expose?
• What environment variables or credentials does it need?
• Is it a Node.js package (preferred) or Python?
• Does it return file paths or inline data? (File paths work better for Jarvis)

Step 2: Install the Package

Install into the data/ directory:

cd data
npm install <package-name>

Important for Docker compatibility: If installing from a local path, use npm pack first:

# From the local package directory
npm pack

# Then install the tarball in data/
cd /path/to/signal-ai-assistant/data
npm install /path/to/package-name-1.0.0.tgz

Direct local path installs (npm install /path/to/local) create symlinks that don't work across Docker mount boundaries.

Step 3: Configure in mcp.json

Add the server to data/mcp.json:

{
  "servers": {
    "server-name": {
      "type": "stdio",
      "command": "${DATA_DIR}/node_modules/.bin/server-binary",
      "args": ["--optional-flag"],
      "env": {
        "API_KEY": "${API_KEY_ENV_VAR}"
      },
      "enabled": true
    }
  }
}

Key fields:

• type: Usually "stdio" for npm packages
• command: Path to the executable, use ${DATA_DIR} prefix
• args: Optional command-line arguments
• env: Environment variables passed to the server (supports ${VAR} substitution)
• enabled: Set to false to disable without removing

Step 4: Add Environment Variables

Add any required credentials to .env:

echo "API_KEY_ENV_VAR=your_api_key" >> .env

For Docker, also add to docker-compose.yml environment section:

environment:
  - API_KEY_ENV_VAR

Step 5: Restart and Test

For local development:

npm run dev

For Docker:

docker compose restart

Check the logs to verify the MCP server loaded:

[mcp] Loaded 1 MCP server(s) from /path/to/data/mcp.json
[mcp]   - server-name: /path/to/data/node_modules/.bin/server-binary

Common Issues

MCP Tool Returns Image Data Instead of File Path

Problem: Claude can see MCP image responses visually but cannot extract raw bytes to save to disk.

Solution: Fork the MCP server and modify it to:

1. Save the image/file to disk (e.g., /tmp/)
2. Return the file path as text instead of inline data

See references/meme-mcp-fix.md for a detailed example.

Tool Not Available in Docker

Problem: Error: No such tool available: mcp__server__toolName

Causes:

1. Symlink issue: Local path install created symlinks that don't exist in container
   • Fix: Use npm pack + install from tarball
2. Missing dependencies: Package has native deps not in container
   • Fix: Add to Dockerfile or use pure JS alternative
3. Env vars not passed: Server fails silently due to missing credentials
   • Fix: Add env vars to docker-compose.yml

Environment Variables Not Resolving

Problem: Server gets empty credentials

Check:

1. Variable is in .env file
2. Variable is listed in docker-compose.yml environment section
3. Variable name in mcp.json matches exactly (case-sensitive)

MCP Server Types

Stdio (Recommended)

Most npm MCP packages use stdio transport:

{
  "type": "stdio",
  "command": "${DATA_DIR}/node_modules/.bin/mcp-server",
  "args": []
}

SSE (Remote)

For remote MCP servers:

{
  "type": "sse",
  "url": "https://mcp-server.example.com/sse"
}

HTTP

For HTTP-based MCP servers:

{
  "type": "http",
  "url": "http://localhost:3000"
}

Best Practices

1. Always test in Docker after local testing - mount boundaries cause issues
2. Prefer file paths over inline data for binary content (images, PDFs)
3. Use ${DATA_DIR} for portability between environments
4. Keep secrets in .env and reference via ${VAR} syntax
5. Set enabled: false to temporarily disable servers without removing config

Additional Resources

Reference Files

• references/meme-mcp-fix.md - Detailed walkthrough of fixing an MCP server to return file paths
• references/mcp-json-schema.md - Full schema documentation for mcp.json

Example Files

• examples/mcp.json - Example configuration with multiple servers

Scripts

• scripts/test-mcp-server.sh - Test an MCP server standalone

Converted and distributed by TomeVault — claim your Tome and manage your conversions.