Skip to content
Ptah Documentation

LM Studio

LM Studio runs models on your own machine and serves them over an OpenAI-compatible HTTP API. Ptah talks to that API through its translation proxy, so any model you load in LM Studio becomes available in the model picker.

Nothing leaves your machine. There is no API key and no account.

What you need

Section titled “What you need”
  • LM Studio installed.
  • At least one model downloaded inside LM Studio.
  • LM Studio’s local server running.

Setup

Section titled “Setup”
  1. Open LM Studio and load a model.
  2. Start the local server. LM Studio listens on http://localhost:1234 by default.
  3. In Ptah, open Settings → Providers → LM Studio.
  4. Leave the key field alone. LM Studio needs no credential.
  5. Click Test connection.

Models

Section titled “Models”

LM Studio has no static model list in Ptah. The catalog is fetched live from LM Studio’s /v1/models endpoint, so the picker shows exactly the models you have loaded — no more, no less.

Load a new model in LM Studio and it appears in Ptah’s picker on the next refresh. Unload one and it disappears.

Model tiers

Section titled “Model tiers”

Ptah maps three tier words — opus, sonnet, and haiku — onto real model ids so that “Default (recommended)” resolves to something your endpoint can serve.

LM Studio ships no tier map, deliberately. Ptah does not know which of your local models is the strong one. Instead it derives the mapping from your live catalog, ranking by context length.

If the derived mapping is wrong for your setup, set the tiers explicitly from the model picker. An explicit choice always outranks the derived one.

A different endpoint

Section titled “A different endpoint”

To use a non-default port or a remote LM Studio instance, set the base URL:

Terminal window
ptah provider base-url set http://192.168.1.20:1234/v1 --provider lm-studio

Or edit it in Settings → Providers → LM Studio.

Cost

Section titled “Cost”

$0. Ptah reports zero cost for LM Studio because the inference runs on your hardware.

Troubleshooting

Section titled “Troubleshooting”

“Connection refused” — LM Studio’s local server is not running. Start it from the server tab in LM Studio, not just the chat tab.

The model picker is empty — Ptah reads the catalog from /v1/models. If that endpoint is unreachable, there is nothing to list. Confirm the server is up and the base URL matches.

A 404 on the first message — the tier word did not resolve to a servable model. This happens when /v1/models is down while inference is up. Pick a model explicitly instead of leaving it on “Default”.

Slow first token — a large model on modest hardware is simply slow. Try a smaller quantization, or shorten your context.

Section titled “Related”