Build, run, and extend FRIDAY
FRIDAY is a local-first, voice-driven AI assistant for Linux and Windows. These docs take you from a fresh clone to writing your own capabilities.
Getting started
From clone to your first spoken command in a few minutes.
Installation
Linux & Windows setup, models, requirements, troubleshooting.
Architecture
The turn pipeline, stores, memory, and how the pieces connect.
How routing works
The deterministic 6-layer router and its confidence bands.
Capabilities
Everything FRIDAY can do, grouped by domain.
Commands & shell
The / slash, ! shell, and > follow-up prefix grammar.
Telegram & remote
Drive the full control surface from your phone.
Add a new tool
Register a capability and wire its intent pattern + tests.
What FRIDAY is
Say “Hey Friday” and ask it to set your brightness, find a file, summarize a PDF, look something up, or run a multi-step workflow — and it does the work on your machine, narrating progress out loud. It is built around three ideas:
- Local-first. Speech-to-text, the conversational model, the planning model, vision, and embeddings all run on your hardware. No account, no cloud inference, no telemetry. Online skills are opt-in and ask for consent.
- Capability-driven. Every skill is a self-contained capability in an MCP-compatible registry. A deterministic intent layer routes common phrasings instantly; anything else falls through to a local planning model.
- Cross-platform. One codebase runs on Linux and Windows, with platform-specific paths guarded throughout.