pi-lm-providers - Ollama & LM Studio for pi

Installation

pi install npm:pi-lm-providers

Alternatively, add "npm:pi-lm-providers" manually to packages in ~/.pi/agent/settings.json. Without npm: copy the .ts files into ~/.pi/agent/extensions/lm-providers/ (pi loads them automatically).

For local development a path package is enough (changes take effect directly after /reload, no reinstall):

pi install /path/to/pi-lm-providers

Registers two dynamic providers:

Provider	Endpoint	Login
ollama	Local server (http://localhost:11434) or Ollama Cloud (https://ollama.com)	/login ollama
lmstudio	LM Studio developer server (http://localhost:1234, incl. LM Studio Bionic)	/login lmstudio (optional)

Models are discovered live from the server - including real capabilities (tools/vision/thinking) and context windows where the server provides them. Models without tool support are hidden (pi is an agent and needs tool calling). Embedding models (LM Studio) are filtered out.

Ollama

Ollama Cloud (API key)

1. Create an API key: https://ollama.com/settings/keys
2. In pi: /login ollama -> Ollama Cloud (ollama.com) -> paste the key
3. /model -> e.g. ollama/gpt-oss:120b, ollama/deepseek-v4-pro:0813, ollama/qwen3.5:397b

Local Ollama server

No login needed - the provider counts as configured as soon as the server is reachable:

ollama serve                 # start the server
ollama pull gpt-oss:20b      # download a model
pi --provider ollama --model gpt-oss:20b

Environment variables (optional)

Variable	Meaning
OLLAMA_MODE	Force cloud or local
OLLAMA_API_KEY	Cloud API key; implies cloud mode when OLLAMA_MODE is unset
OLLAMA_BASE_URL	Override the full base URL (local & cloud, e.g. your own proxy)
OLLAMA_HOST	Local host in Ollama notation (127.0.0.1, host:11434, https://gpu.corp:8443)
OLLAMA_CLOUD_BASE	Override the cloud endpoint

A stored /login decision (local vs. cloud) takes precedence over environment detection.

LM Studio

Start the server in LM Studio (Developer tab -> Start Server, or lms server start). Default: http://localhost:1234, no authentication - no login needed.

pi --provider lmstudio --model google/gemma-4-26b-a4b

If the server requires an API token (Developer tab -> API tokens): /login lmstudio -> API token -> paste the token.

Variable	Meaning
LMSTUDIO_BASE_URL	Full base URL (e.g. http://mac-mini.local:1234)
LMSTUDIO_HOST / LMSTUDIO_PORT	Host and port separately
LMSTUDIO_API_KEY or LM_API_TOKEN	API token

Uses LM Studio's native /api/v1/models API (capabilities, loaded context length, quantization); older servers fall back to /v1/models automatically.

Notes

• Compat: the providers send system (not developer), max_tokens (not max_completion_tokens) and no reasoning_effort parameters - thinking models reason at the server default, and the reasoning trace is shown in pi.
• Context overflow: server errors like "prompt is too long" are normalized so pi can compact the conversation automatically and retry.
• Refreshing the model list: models are loaded at pi startup; after ollama pull ... / downloading a model in LM Studio, a restart or /reload is enough.
• Diagnostics: if a refresh fails (pi only shows "Could not refresh ollama"), the underlying cause is written to ~/.pi/agent/lm-providers.log.
• Cloud costs: usage tracking shows $0 - Ollama Cloud pricing is model-dependent (see https://ollama.com/cloud).
• Important for local Ollama models: Ollama allocates the runtime context server-side (model-dependent, configurable globally via OLLAMA_CONTEXT_LENGTH). On very long sessions the server can still throw "prompt is too long" - raise OLLAMA_CONTEXT_LENGTH and restart the server.