Skip to content

What is COMSTAR

Zlatko Lakisic edited this page Aug 3, 2026 · 2 revisions

What is COMSTAR

COMSTAR is a physical AI terminal — an always-on, embodied presence for a room. Not a phone app. Not a browser tab. A small screen that lives in your space, recognises who you are, and talks back.

Embodied presence vs screen


The pitch in one paragraph

A Raspberry Pi with a camera, mic, speaker, and HDMI panel runs continuously. When you walk into frame, CodeProject.AI detects a person and recognises your face. That identity opens an agentic-orchestration session scoped to you. You say hey comstar (or continue in a follow-up window). Speech is transcribed, the planner answers, Piper (or Ada TTS) speaks, and a live SVG emblem on the kiosk animates with the audio.


Why it exists

Proto Hologram proved the thesis: an embodied presence in a room is categorically different from a video call or a text box. People walk up to it. People talk to it without being prompted. But Proto is $10k–$250k of proprietary hardware running a closed OS, and you cannot put your own agent logic behind it.

COMSTAR is the open version. Commodity parts, a GPU you already own, and an orchestration engine you can rewrite.

Proto inspiration vs COMSTAR


What it is not

Not this Because
A cloud chatbot with a pretty face Phase 1 stays on your LAN
A full robot No locomotion — presence through attention and voice
A dashboard of sensors One job: be addressable in a room
A wrapper around ChatGPT The brain is agentic-orchestration with MCP tools

The three machines

Machine Role
Raspberry Pi 4 Thin I/O terminal — capture, wake, VAD, kiosk, playback
AI server (Ada) Orchestration, vision (CPAI), preferred speech sidecars
Mac (dev) Editor, bridge-dev, kiosk hot reload, fake hardware

Details: Architecture · Hardware · Dev Loop.

Three machines topology


Identity is load-bearing

Face recognition is not decoration. The userid from CodeProject.AI becomes:

x-agentic-user-name: <userid>
x-agentic-session-id: comstar-<userid>

Session memory and knowledge-base scoping are therefore per person for free. An unrecognised face opens an anonymous guest session with a restricted overlay.

More: Vision & Face Enrollment · Attention Model.


Related pages

Clone this wiki locally