Ambient AI Agent
An ambient AI agent that sees what you see, remembers your day, and acts — live across your glasses, your phone, and a powerful backend brain.
Hawk wears your world: an iOS frontend on iPhone and Ray-Ban Meta glasses is the eyes and ears, while a long-lived backend agent does the thinking and the acting. One system, always circling — and six concrete flows it already runs.
Get started
Hermes has no app — copy the prompt and paste it into your Hermes session.
One system, two halves
Everything Hawk does comes back to three things — what it wears, what it runs, and what it can already do.
The Hawk iOS app turns your iPhone and Ray-Ban Meta glasses into a first-person camera and microphone. Live conversation, visual memory, a multi-panel carousel, and a lock-screen widget — the eyes and ears you carry into the room.
iOS frontendThe Hawk gateway is a long-lived process running a powerful coding-agent core. Agent loop, sessions, hybrid memory, heartbeat, cron, MCP tools, and sub-agents — always on, always watching, ready to act.
Backend agentNot a promise deck — real flows. From whispering names across a noisy room to shipping a feature you asked for out loud, Hawk runs six concrete demos today, with more circling in.
Multiple demo flowsSix flows, one agent
Six flows from the same agent — three live, two in progress, one in preview. Five are real recordings; tap any clip to play. The last is wired and circling in.
In a noisy party or meeting, Hawk listens, picks out who's talking, and quietly briefs you — names, context, and what to say next.
Hawk watches your surroundings and warns you about hazards in real time, before they reach you.
Hawk notices the commitments you make in conversation and surfaces the right reminder at the right moment.
A hands-free, no-audio mode. Hawk reads the scene and assists without speaking — a glance at your glasses display or lock screen is all it takes.
Ask Hawk out loud to build a feature. The backend agent writes and ships the code while you keep moving.
"What did I do yesterday?" "What did I just do?" Hawk answers from a searchable visual record of your day — down to "how many times did my dog appear in the last 10 minutes?"
Capabilities
Perception, memory, and action — the pieces that let Hawk see, remember, and strike.
Ray-Ban Meta glasses become a live input to the agent — first-person camera and microphone, streaming what you see and hear. Hawk's vision is whatever yours is pointed at, hands free.
Hawk keeps a searchable record of what you saw and did. Ask in plain language or query the video directly — "what did I just do," "where did I leave it," "how often did that appear" — and get an answer from your own day.
The realtime conversation is dumped to a session log, then a single Haiku call consolidates that session log into a daily log. At the start of each realtime session, Hawk's prompt is rebuilt from soul, identity, global memory, and daily memory — so it always wakes up knowing who it is and what just happened.
Four live modes for whatever the moment needs — meeting and cocktail briefings, real-time safety, in-the-moment reminders, and a silent, no-audio mode. Hawk shifts how it helps without you breaking stride.
Speak a request and the backend agent goes and does it — writing and shipping a real feature while you carry on. Hawk isn't a chat window; it's an agent with hands.
Phone mic, glasses camera, and more — Hawk ingests multiple concurrent input streams at once and routes them to the agent. Many eyes, many ears, one decision loop.
A prompt-defined soul plus an identity give Hawk a consistent character across every session. It doesn't reset between conversations — it stays itself.
Architecture
Your glasses and phone stream everything you see and hear to the Hawk gateway as multiple concurrent streams over a single WebSocket, tunneled through Tailscale. There the agent loop, memory, and background services do the thinking — pluggable across Hermes, Codex, and Claude Code.
Install
Stand up the backend in one line, then point the iOS app at it. URLs and names below are placeholders — keep them tidy and swap when you ship.
The long-lived brain. Requires Bun ≥ 1.3.
# install
bun install -g hawk # requires Bun >= 1.3
# or one-line:
curl -fsSL https://hawk.sh/install.sh | bash
# run
hawk gateway # start the long-lived backend
hawk # open the terminal clientThe eyes and ears you carry.
ws://<your-tailscale-ip>:4242Placeholder URLs and names (hawk.sh, ws://<your-tailscale-ip>:4242) — swap for your own when you deploy.
Roadmap
Where Hawk's vision and reach widen next.
Realtime-agent benchmark plus an auto-testing gym with synthetic tests.
Agent builds agent: prompt optimization, sharper tooling, live dev.
Hawk in the car — briefing and acting from the dashboard.
Sample perception at a rate tuned to live latency.
Android, Apple Watch, and more glasses beyond Ray-Ban Meta.
Face recognition and speaker diarization to map audio to who's talking.
Tunable soul/identity and shareable memory presets.
Cite / Source
Hawk is open at the link below. If it helps your work, cite it — the BibTeX is ready to copy.
github.com/hawk-agent/hawk@software{hawk_ambient_agent,
title = {Hawk: An Ambient AI Agent},
author = {The Hawk Team},
year = {2026},
url = {https://github.com/hawk-agent/hawk},
note = {Placeholder URL — see project page}
}