Most AI coding extensions lock you to one vendor's key. The Pi coding agent was built differently: it is provider-agnostic, and PiLot Studio makes that multi-provider workflow visual. This guide covers how the pieces fit together — and what it means for privacy.

Bring your own key — any key

PiLot Studio speaks to every provider the pi CLI supports:

Anthropic — Claude models, including extended-thinking variants with adjustable reasoning levels.
OpenAI — GPT models.
Google — Gemini models.
xAI — Grok models.
OpenRouter — one key, hundreds of models.
Local — Ollama and llama.cpp for fully offline coding agents.

Keys live in your environment or PiLot's settings panel. The extension has no account system and no proxy: your requests go from VS Code to the provider you chose, nowhere else.

Mixing cloud and local in one session

The model dropdown sits above every conversation. A common pattern:

Plan and review with a frontier cloud model (e.g. Claude Sonnet).
Execute mechanical edits with a local model — free, fast, and private.
Switch mid-conversation; the session keeps its full context.

Privacy-oriented by design

BYOK — no middleman service ever sees your code or your keys.
Local models — with Ollama or llama.cpp, the entire loop (code + prompts + inference) stays on your machine.
Offline dictation — voice input runs Vosk locally; audio is never uploaded.
Light Mode — one toggle to run the agent with skills, extensions, and context files off, ideal for sensitive repos or small local models.

Try it

The provider matrix — including setup notes per vendor — lives on the providers page. To get running end-to-end, follow the install guide.