One day, one feature: Connect Mini Me to the world with MCP

The very small local models that I fondly like to call "tiny models" won't let you do vibe coding. They're not here to replace you (well, I mean the agent that will use a tiny model), they're here to give you a hand, explain a concept, search through documentation, ... And now many of them have excellent function calling support, so they're able to use the tools provided by MCP servers.

And of course, Mini Me brings MCP support.

Add an MCP configuration to your agent

To give Mini Me "MCP superpowers", you just need to add this to your agent's configuration file (here I want to connect to the Wikipedia MCP server):

mcp:
  enabled: true
  permissions: ask             # Ask before executing MCP tools (recommended for security), otherwise use `auto`
  discoveryTimeout: 10s        # Timeout for discovering tools at startup
  
  servers:    
    # Wikipedia MCP Server - Free Wikipedia access, no authentication required
    - name: wikipedia
      transport: http
      url: https://wikipedia-mcp.vercel.app/mcp
      timeout: 30s
      auth:
        type: none  # No authentication required
      headers:
        Accept: "application/json, text/event-stream"
      # Optional: filter which tools to expose
      # tools:
      #   - search
      #   - get_article

• ✋ tools: You can filter the tools by listing only the ones you need (with the tools: section), which can be handy if the MCP server exposes too many tools. Don't forget that the more you overload small models with too much information, the more they struggle to get their job done.
• For my agent I'm using Qwen3.5-4B-GGUF:Q4_K_M.

And now I can easily query Wikipedia from my agent:

Of course, you can add several MCP servers

For my little "coding helper" agents, I also need to search through documentation. For that I use Context7, a service that provides coding assistants with up-to-date documentation and code examples, with the ability to target specific library versions. You can install it locally or use the remote MCP server (which is what I'm using here). Add a new subsection to the mcp section:

mcp:
  enabled: true
  permissions: ask             # Ask before executing MCP tools (recommended for security)
  discoveryTimeout: 10s        # Timeout for discovering tools at startup
  
  servers:    
    # Context7 MCP Server - Provides documentation and code context
    # Requires authentication via API key
    - name: context7
      transport: http
      url: https://mcp.context7.com/mcp
      timeout: 60s              # Longer timeout for potentially complex queries
      auth:
        type: bearer
        tokenEnv: CONTEXT7_API_KEY  # Set this environment variable with your API key
      headers:
        Accept: "application/json, text/event-stream"
      # Optional: filter which tools to expose
      # tools:
      #   - resolve-library-id
      #   - query-docs
      
    # Wikipedia MCP Server - Free Wikipedia access, no authentication required
    - name: wikipedia
      transport: http
      url: https://wikipedia-mcp.vercel.app/mcp
      timeout: 30s
      auth:
        type: none  # No authentication required
      headers:
        Accept: "application/json, text/event-stream"
      # Optional: filter which tools to expose
      # tools:
      #   - search
      #   - get_article

✋ Important: you'll need to create an account (there's a free plan that lets you test in excellent conditions, and for now it's still enough for my use cases). Then you'll have to generate an API key and set it in the CONTEXT7_API_KEY environment variable. 🔥 And don't forget that your agent can see all your secrets, so protect yourself. In my case, I'm "dog fooding" and I use Docker sbx and its custom secrets feature to protect my credentials.

With this new server, I bring a significant amount of extra knowledge to my "little coding helper". For example, I'm currently learning Rust and I'd like to know how to connect to a Redis database in Rust:

The assistant will contact the Context7 MCP server, which will return the list of libraries matching the search criteria:

It can then query the documentation more specifically (still with the help of the MCP server):

And finally, from the collected information, our assistant can generate explanations showing you how to connect to Redis with Rust, based on official (and therefore relevant) sources, and no longer on its training corpus.

Once again, you can see that even a "modest size" model (Qwen3.5-4B-GGUF:Q4_K_M) has the sufficient and necessary capabilities to accomplish this kind of task, here excellent "function calling" support. So in a multi-agent setup, you can imagine delegating documentation search to an agent using a local model rather than to an agent using a remote, paid endpoint. Every little saving counts 😉.

That's it for this "feature of the day". You'll find the agent's configuration here: /06-mcp/agent.yaml. See you very soon.