pi-lm-studio

A pi package that discovers models from a running LM Studio server and registers them as a selectable lm-studio provider inside pi.

• Detects models from LM Studio's OpenAI-compatible and native model endpoints
• Registers them as a dynamic lm-studio provider in pi
• Adds /lm-studio-refresh to rescan models
• Adds /lm-studio-use to quickly switch to a detected LM Studio model
• Adds /lm-studio-context to show current context window usage
• Reports context usage in pi's status bar for active LM Studio models
• Enables streamed usage reporting by default so pi can track context more accurately
• Handles common local-server compatibility settings for pi

From GitHub:

pi install git:github.com/Acelogic/pi-lm-studio

Or add it to ~/.pi/agent/settings.json or .pi/settings.json:

{
  "packages": [
    "git:github.com/Acelogic/pi-lm-studio"
  ]
}

Pull the latest commit on the default branch:

pi update git:github.com/Acelogic/pi-lm-studio

Or update every installed package at once:

pi update

Reload the running session to pick up the new code:

/reload

pi update skips pinned sources. If you installed with a tag suffix (e.g. @v1.0.0), move to a newer tag by reinstalling:

pi install git:github.com/Acelogic/pi-lm-studio@v1.1.0

pi list
pi remove git:github.com/Acelogic/pi-lm-studio

Start the LM Studio local server first.

Default base URL:

http://127.0.0.1:1234

Optional override:

export LM_STUDIO_BASE_URL=http://127.0.0.1:1234

LM_STUDIO_URL is also supported.

After install, reload pi if needed:

/reload

Then:

• /lm-studio-refresh — detect available LM Studio models
• /lm-studio-use — pick a detected model from a selector
• /lm-studio-context — show current context usage for the active LM Studio model
• /model — your LM Studio models should also appear in the main model picker

• Embedding models are filtered out
• Models without tool tuning are labeled accordingly
• Vision support is inferred from LM Studio's native metadata when available
• Context usage in pi comes from streamed usage when LM Studio provides it, with pi falling back to estimation when needed
• If your LM Studio server does not support streamed usage, you can disable it with LM_STUDIO_STREAM_USAGE=false