Updated

Local models: Ollama + Goose

A local model does not connect to md-doc directly: the client that uses it does. With Ollama serving the model and Goose as the MCP client, your local model reads and writes the same notes, logs and issues as Claude or ChatGPT.

It is tested with this combination. Other clients with MCP-over-HTTP support should work the same way, but we have not verified them.

Requirements

  • Ollama installed and running.
  • A model that supports tool calling. Without it, the model cannot use MCP.
  • Goose installed.
  • An MCP-type API key: Getting started.

Steps

  1. Pull a model with tool support, for example:

    ollama pull qwen3
  2. Configure Goose with Ollama as the provider:

    goose configure

    Choose Configure Providers → Ollama and the model.

  3. Add md-doc as a remote extension. In goose configure → Add Extension → remote extension over HTTP, with:

    • Name: md-doc
    • URL: https://app.md-doc.com/api/mcp
    • Header: Authorization: Bearer mcp_YOUR_KEY

    If you prefer to edit Goose’s configuration file, the extension looks like this (check the field names for your version):

    extensions:
      md-doc:
        type: streamable_http
        uri: https://app.md-doc.com/api/mcp
        headers:
          Authorization: Bearer mcp_YOUR_KEY
        enabled: true
  4. Start a session with goose session and ask something like “list my open issues in md-doc”.

Tips

  • Small models make more mistakes choosing tools. If in doubt, try simple requests first (search, list).
  • To use secrets, the model needs a terminal; see the mddoc CLI.

Frequently asked questions

How do I give a local model persistent memory?

Use a client with MCP support, such as Goose, that talks to your model (Ollama) and to the md-doc MCP server. The model reads and writes notes, logs and issues through the client.

Does md-doc connect directly to Ollama?

No. What connects to the md-doc MCP is the client (Goose), not the model. Ollama only serves the model.