A realtime voice in front of the agent
Gemini Live or OpenAI Realtime follows everything the agent does and tells you: its replies, its questions, problems and progress. Ask "how's it going?" mid-task and you get an answer at once.
For Claude Code, OpenCode and Codex
Give your coding agent a face.
OpenFace floats over your desktop and talks with you. Say what you want: your agent does the work in the background, and the face tells you how it's going, out loud, as it happens.
Free, for Windows. Works with any model your agent can run.
The face here is the app's own engine, running in your browser: the same skins, teeth, expressions and lip sync. Turn the sound on to hear it.
Hear the lip sync
Claude CodeOpenCodeCodexGemini Live
OpenAI RealtimeElevenLabsmodels.dev
Watch it work
Real agent, real voice, real time
Claude Code finds and fixes two bugs in a checkout module. The face asks before it runs anything, then tells you what was wrong. Recorded in one take, nothing sped up.
What it does
A colleague you can talk to while it works
Any agent, any model
Claude Code, OpenCode or Codex behind the face. Switch in Settings, and each agent keeps its own conversation. Pick from the models.dev catalog, or type any model name.
Your words reach the agent verbatim
Speech is recognised on your PC and goes to the agent word for word. The voice model never hears you and never does the work: every request is answered once, by the agent.
Say its name to cut in
A small keyword spotter listens for the face's name. Say "Claude" while it talks and it stops within about half a second and listens for what you want next.
Hears you, not the room
Fans, hum and typing are taken out before your words are recognised, and voices much quieter than yours, like a TV or people across the room, are ignored.
Watch every step
The activity panel shows the session as a timeline: tool calls, results, subagents, cost per request, and Claude Code's full transcript, live. Dock it to the edge of the screen as a drawer.
It can use your PC
Claude Code can take screenshots and use the mouse and keyboard in desktop apps, drive a Chrome window of its own through Playwright, or work in your own Chrome. Clicking and typing ask first.
Expressions
Happy, curious, concerned, amused, sad: Claude Code sets its own expression with a face tool, and the face takes cues from what it is saying.
Local where it counts
Speech recognition runs on your PC, so no audio leaves it. API keys stay in a local file, never in the page or the logs, and the local server makes a fresh access token every run.
How it works
The agent does the work. The face does the talking.
- 1
You speak
Speech recognition on your PC turns it into text. Or just type, or drop a file on the face.
- 2
The agent works
Your words go straight to Claude Code, OpenCode or Codex as its prompt, in a background session.
- 3
The voice follows
Every tool call, result and reply becomes a short note for the realtime voice model.
- 4
The face tells you
Replies, questions and progress, in its own words, lip-synced. Answer "yes" or "no" to a permission prompt.
Permissions stay yours
Claude Code and OpenCode ask before they run a command or change a file. The face asks out loud and shows Allow, Always and Deny. Skipping permissions is one switch away, and the face wears a red rim while it's on.
No API key? It still talks
Without a realtime voice key, the classic voice reads the agent's replies with text-to-speech that runs on your PC, or with ElevenLabs if you have a key.
Faces
Wear any face
A glowing hologram, a sculpted head, or a photo: pick one of eight portraits, or drop any picture of a face onto it to wear it. Robots, monsters and counts from Transylvania work too. Change the colours, the teeth (sharp and vampire are options), and how it moves. Say "go away" and it burns away in a shower of sparks. Say "come back" and it does.
Rooms
Several agents, one conversation
Put faces in the same room and they behave like a team around you. The room owns the microphone, so you are heard once, and only one face speaks at a time.
- One round per question. The first agent with something to say leads. Each of the others adds a sentence or two, or passes. Then the lead concludes. Time limits on every turn, so they can't loop.
- Say a name to choose. "Codex, what do you think?" goes to Codex first. "Everyone, ..." goes to all. Say a name, or "stop", while anyone is talking and they all stop.
- Faces only hear you. Agents never hear each other's voices. What one says reaches the others as text, labelled with its name.
OpenFace.cmd --face claude --room team
OpenFace.cmd --face codex --room team
OpenFace.cmd --face opencode --room teamAgents
Pick the mind behind the face
| Claude Code | OpenCode | Codex | |
|---|---|---|---|
| Connects through | Claude Agent SDK | Agent Client Protocol (opencode acp) | codex exec --json |
| Asks before acting | Yes: Allow, Always, Deny | Yes: Allow, Always, Deny | No: its sandbox decides |
| Permission modes | default, acceptEdits, plan | default, own | default, read-only |
| Runs as | One long session | One long session | One run per message, resumed |
| Models | opus, sonnet, haiku or any id | Any provider/model | Any Codex model |
Get started
Running in a few minutes
The first run installs everything it needs inside the folder: a Python environment, Electron, and the speech models (about 650 MB, downloaded once).
- Windows 10 or 11
- Python 3.11 or newer, and Node.js
- Claude Code, OpenCode or Codex, installed and signed in
- For the realtime voice: a Gemini or OpenAI API key (optional)
git clone https://github.com/compsmart/ai-faces openface
cd openface
OpenFace.cmdThen right-click the face for settings, turn on listening, and say hello. Several faces at once?
OpenFace.cmd --face work gives each one its own settings, folder and voice.