FRIDAY

Introduction

FRIDAY started as a question: what would a voice assistant look like if it actually lived on your machine, knew your context, and could execute multi-step CLI work without ever crossing a network boundary? The product is a locally-running AI agent featuring multi-layer routing, persistent memory, and autonomous task execution through an MCP-standardized tool surface.

The problem with cloud-bound assistants

Existing voice assistants rely heavily on cloud compute and treat the desktop as a microphone for someone else's API. They cannot execute complex, multi-step CLI operations, do not retain procedural memory across sessions, and have no concept of the local filesystem or shell. For a Linux power-user, that is the inverse of useful.

An assistant that lives on the machine should know the machine — its files, its shell, its history. Anything less is a hotline to a server.

Architecture

I engineered a three-tier routing architecture — deterministic execution for known intents, LLM tool selection for ambiguous ones, and a conversational fallback when neither applies. On top of that sits the TaskGraphExecutor: a planner that computes topological waves of tool dependencies, enabling parallel execution of independent tasks instead of the standard sequential agent loop.

Persistent memory is split into three layers — episodic events go to SQLite, semantic facts to ChromaDB with embeddings, and procedural sequences are surfaced as MCP tools the agent can re-invoke. The result is an assistant that gets faster the longer you use it, because every successful task becomes a primitive for the next one.