MCP Development

MCP Development Skill

Guide for building and integrating MCP servers with LlamaFarm.

When to Load

Load this skill when the user:

• Wants to build a custom MCP server
• Needs to expose business logic as AI tools
• Is configuring MCP servers in LlamaFarm
• Asks about inline tool definitions
• Needs help with tool access control

MCP Overview

MCP (Model Context Protocol) is a standardized protocol for giving AI models access to external tools, APIs, and data sources.

LlamaFarm supports MCP both as:

1. Client - Connect to external MCP servers
2. Server - Expose LlamaFarm's API as MCP tools

Quick Reference

Connect to MCP Server

mcp:
  servers:
    - name: filesystem
      transport: stdio
      command: npx
      args: ['-y', '@modelcontextprotocol/server-filesystem', '/data']

Assign to Model

runtime:
  models:
    - name: assistant
      mcp_servers: [filesystem]

Define Inline Tools

runtime:
  models:
    - name: assistant
      tools:
        - type: function
          name: my_tool
          description: Does something useful
          parameters:
            type: object
            properties:
              input:
                type: string

Progressive Disclosure

For detailed documentation:

• python-mcp-server.md - Build MCP servers in Python
• inline-tools.md - Define tools in YAML configuration
• transport-types.md - STDIO, HTTP, SSE transports
• access-control.md - Per-model tool access

Common Use Cases

1. File System Access

Give models access to read/write files:

mcp:
  servers:
    - name: filesystem
      transport: stdio
      command: npx
      args: ['-y', '@modelcontextprotocol/server-filesystem', '/allowed/path']

runtime:
  models:
    - name: assistant
      mcp_servers: [filesystem]

2. Database Queries

Let models query SQLite databases:

mcp:
  servers:
    - name: database
      transport: stdio
      command: npx
      args: ['-y', '@modelcontextprotocol/server-sqlite', './data.db']

3. Custom Python Tools

Expose your own business logic:

mcp:
  servers:
    - name: my-tools
      transport: stdio
      command: python
      args: ['-m', 'my_mcp_server']

4. Remote API

Connect to cloud-hosted MCP servers:

mcp:
  servers:
    - name: company-api
      transport: http
      base_url: https://api.company.com/mcp
      headers:
        Authorization: Bearer ${env:API_TOKEN}

Tool Call Flow

1. User sends message to LlamaFarm
2. Model decides to call a tool
3. LlamaFarm executes tool via MCP
4. Result returned to model
5. Model generates final response

User → LlamaFarm → Model → Tool Call → MCP Server → Result → Model → Response

Security Considerations

Least Privilege

Only grant tools the model actually needs:

runtime:
  models:
    # Research model: read-only access
    - name: researcher
      mcp_servers: [filesystem]  # Read files only

    # Admin model: full access
    - name: admin
      mcp_servers: [filesystem, database, api]

    # Chat model: no tool access
    - name: chat
      mcp_servers: []

Environment Variables

Never hardcode secrets:

mcp:
  servers:
    - name: api
      transport: http
      base_url: ${env:API_URL}
      headers:
        Authorization: Bearer ${env:API_TOKEN}

Path Restrictions

Limit file system access to specific directories:

mcp:
  servers:
    - name: docs
      transport: stdio
      command: npx
      args: ['-y', '@modelcontextprotocol/server-filesystem', '/safe/docs/only']