AI analysis & local models

Analyze trends reads your sessions' telemetry with an AI model and summarizes the friction it finds. It's strictly bring-your-own: pick a provider in Settings and paste your own API key (saved on your device only, never stored by urTRACE) - or skip the AI companies entirely and point it at a model you run yourself.

• Ollama. Install from ollama.com, fetch a model (ollama pull llama3.1), then in Settings pick Local / private model and enter your endpoint - Ollama serves at /v1/chat/completions on port 11434 - plus the model name.
• LM Studio. Install from lmstudio.ai, download a model, and start the server from the Developer tab - it serves the same /v1/chat/completions shape on port 1234. Any other OpenAI-compatible endpoint (a company AI gateway, a self-hosted box) works the same way; the API key field is only for endpoints that actually require one.
• The endpoint must be reachable from urTRACE. Analysis runs through our server, so a plain localhost address on your laptop isn't visible to it (and private addresses are refused). Expose your machine through a tunnel you control - Cloudflare Tunnel, Tailscale Funnel or ngrok all work - or run the model on a box with its own public address.
• Mind the context window. Local models often run with small default windows (Ollama: 4096 tokens), and a model that overflows doesn't error - it silently ignores the earliest sessions while the summary looks complete. urTRACE asks your runtime for its real window and shows a warning next to the analysis whenever the data won't fit; raise the window (OLLAMA_CONTEXT_LENGTH, a Modelfile num_ctx, or LM Studio's context-length setting) or analyze fewer tests at once.
• Built for small models. Multi-test analysis summarizes each test on its own, then synthesizes across the summaries - so no single request needs a giant window, and modest local models stay usable.