Can LM Studio API trigger local MCP tool calls from external clients?

[pepel101]

I’m trying to access local MCP servers installed on my PC, which are accessible from LM Studio and work relatively well in my home lab.

However, when I try to access my home lab from a custom Android client via the LM Studio API, with authentication enabled and MCP.json calls enabled with authentication, I cannot get the MCP tool calls to actually execute when sending queries or chats to the LM Studio server.

I’ve been working on this for two days and cannot figure out a way to make it work.

Is it actually possible to call local MCPs on a PC via the LM Studio API from an external client? Is there an issue with this or documentation I may have missed?

[tedchirvasiu]

Have you ever figured it out?

[principalwater]

Yes, it is possible — but not through /v1/chat/completions. That endpoint never attaches MCP servers on its own; a client has to send tools itself. Servers listed in mcp.json are reachable only through POST /api/v1/chat, which is a different request shape:

curl http://<host>:1234/api/v1/chat \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "<model>",
    "input": "What is the population of Reykjavik? Search the web.",
    "integrations": ["mcp/<server-name-from-mcp.json>"],
    "context_length": 8000
  }'

Three things that are easy to get wrong here:

• the field is input (a string), not messages
• context_length is required
• the integration id needs the mcp/ prefix — plain "searxng" is rejected

Requires LM Studio 0.4.0+. Docs: https://lmstudio.ai/docs/developer/core/mcp

The part that probably blocked you. If you get

Permission denied to use plugin 'mcp/x'. Ensure that the server configuration allows plugin usage and, if using an API token, it has the necessary permissions.

note that this points at two separate grants, and per-token permissions being correct is not enough. Both live in .internal/permissions-store.json:

{"json":{
  "tokens":[{ "...": "...", "permissions":{"pluginUse":"allowAll"} }],
  "serverPermissions":{"dynamicRemoteMcpServer":"allowAll","pluginUse":"deny"}
}}

In my case both tokens already had pluginUse: "allowAll" and the blocker was serverPermissions.pluginUse sitting at "deny". Setting that to "allowAll" and restarting the app fixed it — after which the model issued a real tool call and answered from live search results. Since you mention having "MCP.json calls enabled with authentication", the server-level grant is worth checking; it is a different toggle from the per-token one.

One thing that does not work, in case you try it from a mobile client: chats served over LM Link / Locally never get the schemas attached, regardless of these permissions. I measured prompt sizes and opened #2313 for that separately.