Use SourceShelf with LM Studio

Requirements #

• A current LM Studio release with MCP support (LM Studio documents MCP support from 0.3.17).
• A downloaded and loaded model that supports tool use.
• A current SourceShelf share created through Packs > More > Local AI Access….

Install the SourceShelf connection #

1. In SourceShelf, open the pack’s Local AI Access sheet.
2. Select Copy MCP Configuration.
3. In LM Studio, open the Program tab.
4. Select Install, then Edit mcp.json.
5. Paste or merge the copied mcpServers entry. Do not remove other servers you want to keep.
6. Save the file and enable the SourceShelf integration.

The result has this shape:

{
  "mcpServers": {
    "sourceshelf-municipal-research": {
      "command": "/Applications/SourceShelf.app/Contents/MacOS/SourceShelfMCPServer",
      "args": ["--share", "<share-id>"]
    }
  }
}

Use the exact command and share ID copied by SourceShelf. Do not manually substitute an application or helper path.

Test the tools #

Load a tool-capable local model and start with a narrow, explicit request:

Use the SourceShelf tools to search the shared pack for tree canopy targets. Read the best matching resource, summarize the targets, and cite the SourceShelf resource URI. Do not answer from general knowledge.

For a smaller model, make the sequence explicit:

First call search_pack with the query tree canopy targets and a limit of 5. Then call read_pack_resource for the best result. Base the answer only on the returned text.

You should see one search call followed by one or more bounded reads. If the model describes the tools instead of calling them, try a model with stronger function-calling support or keep the procedural wording.

LM Studio as a local API server #

LM Studio can also make configured MCP servers available to an API-driven agent. Current LM Studio documentation requires server configuration and authorization settings before API clients can invoke servers from mcp.json. This is an advanced route; the built-in chat is the simplest functional test.

Common LM Studio errors #

“Plugin process exited unexpectedly with code 1” #

The helper could not start or validate the share. Recopy the configuration after moving or updating SourceShelf, confirm sharing is enabled, and confirm the pack still shows a current snapshot.

“Unknown method: tools/list” #

The client reached a resource-only or older helper. Install the current SourceShelf version, recopy the configuration, and restart the integration. Current SourceShelf exposes both search_pack and read_pack_resource.

The context window fills quickly #

Ask the model to search first and read only one or two results. SourceShelf’s paged read tool exists specifically to avoid loading every resource into the prompt.

Official references: LM Studio MCP servers, LM Studio MCP API use, LM Studio server settings.