skip to content
ksegit
LON --:--:--

Talos

A personal AI agent with a face - animated robot eyes mirror what the agent is doing, backed by a Dockerized agent stack with persistent memory.

[HW][DATA]private repo · walkthrough on request

flutter · dart · docker · postgres · redis · honcho · stt/tts

Problem

An always-on personal agent that thinks, edits files, runs commands and reads email is invisible between prompts. A terminal window is the wrong ambient interface - you want presence: a glance should tell you what the agent is doing, and voice should be a first-class channel.

Approach

Two halves. The Hermes backend runs the agent as a Docker stack - an OpenAI-compatible API with a dashboard, plus Honcho memory on Postgres and Redis. The Talos Face app is a cross-platform Flutter client whose code-drawn robot eyes are driven by a pure activity state machine: thinking, editing, running commands, reading - each activity has a face.

 voice ──► voicebox STT ──► Talos Face (Flutter)
                                │  activity state machine
                                │  code-drawn eyes · chat
                                ▼
                     Hermes agent API :8642
                        │            │
                  Honcho memory   dashboard :9119
                  (Postgres+Redis)

Development is plan-driven and test-first: phased implementation plans live in the repo (docs/plans/), each merged behind unit and widget tests.

Tradeoffs

  • Flutter over native. One codebase reaches macOS first and then Linux/Windows/Android/iOS; eyes are code-drawn rather than sprite assets, so the animation core stays testable as a pure state machine.
  • Vendored voice. STT/TTS is an upstream project (voicebox) kept out of git history and wired by port - replaceable without touching the agent.
  • Memory as a service. Honcho handles persistence instead of a bespoke store - one less thing to own, at the price of an extra container.

Outcome

Plan 1 - the eyes animation core and a wake-word feasibility spike - is merged and tested; the presence service that binds face to agent hooks is drafted as plan 2. The robot has eyes; next it learns to point them.