Skip to main content

Connect a host application

You interact with the data through your MCP host application; there is no separate chat UI to install. This section uses LM Studio as the worked example. Any MCP-compatible host (Claude Desktop, Ollama, OpenCode, OpenWebUI, and others) works the same way once pointed at the server; only the location of the MCP configuration differs, so consult that application's MCP documentation for where to put the server entry.

Before you start, make sure the testing environment is running.

Install LM Studio​

Download and install LM Studio for macOS, Windows, or Linux. LM Studio bundles a local model runtime with a chat UI, so the model runs on your machine and nothing leaves it.

Choose a local model​

The workflow requires a model that supports tool calling: the model calls the server's tools to discover the schema and run queries. In LM Studio, tool-capable models are marked with a tool-use badge. The target is open-weights models in roughly the 4B to 70B parameter range, Q4-quantized, so the whole workflow runs locally on accessible hardware.

  1. Open the Discover tab (magnifying glass icon).

  2. Search for a tool-capable model. Reasonable starting points:

    ModelNotes
    Qwen2.5-7B-Instruct (Q4, ~4.7 GB)Good balance of speed and reasoning; strong tool-calling
    Qwen2.5-3B-Instruct (Q4, ~2 GB)Smaller option for machines with limited RAM
    important

    These are provisional starting points, not a benchmarked recommendation. We are currently working on establishing a defensible, benchmarked model recommendation.

  3. Click Download and wait for it to complete.

  4. Open the Chat tab and load the downloaded model; it should appear in the top status bar.

Configure the MCP connection​

The server is setup for Streamable HTTP MCP transport at http://localhost:3100/mcp. In LM Studio:

  1. Select the Chat tab from the left-hand menu.
  2. In the right-hand panel, under Integrations, select + Install. This opens your mcp.json file.
  3. Add the MCP server entry.
  4. Click Save and confirm the server status shows Connected.

If you set a custom MCP_PORT in the demo's .env, use that port in the URL instead of 3100.

With a host connected, work through the demonstrated use cases.