Wireface Desktop
Rooms
Several faces, each with its own agent, sharing one microphone and one conversation. One agent leads, the others add a sentence each, and the lead wraps up.
On this page
Start a room
Run each face with the same room name:
Wireface.cmd --face claude --room team
Wireface.cmd --face codex --room team
Wireface.cmd --face opencode --room teamWhen installed, use Wireface.exe in the install folder the same way (see
Several faces). Or type a Room name in each face's Agent tab, under
Room (room.name in profile.json). A face that joined with --room
remembers a name typed there for next time.
The first face to join starts a small room process that the others find. It closes itself when the last face leaves. The Room group lists who's In the room.
One microphone
The room owns the microphone and the speech recognition, so only you are heard, and only once. The agents never hear
each other's voices: what one agent says reaches the others as text, labelled with its name ([Claude]: ...),
in the next prompt they get. Faces in a room always use the classic voice.
How a round works
- You ask something.
- The first agent with something to say takes the floor and leads. The others keep working but hold their answers.
- When the lead has finished speaking, each other agent in turn adds something new in a sentence or two, or says
nothing (it's told to reply
PASS, and a pass is never spoken). - The lead gets everyone's comments and gives the conclusion.
For example: you ask, Claude answers first, Codex adds a point, Claude concludes. That's the whole round: one pass over the agents, with a time limit on every turn, and nothing an agent says can start another round, so they can't loop. An agent that's slow (a rate-limited model, say) can hold the round up for up to 90 seconds before it's skipped.
Choosing who answers
- "Codex, what do you think?" goes to Codex first: it leads, and the others comment after.
- "Everyone, ..." goes to all of them.
- Typing into one face's message box goes to that face's agent first.
- While anyone is speaking, saying an agent's name, or "stop", "wait" or "quiet", stops everyone. "Codex, ..." then makes Codex lead.
- A bare "Codex", or "Codex, stop", stops everything, and your next sentence goes to Codex.
Every agent works on every question you don't address by name.
Room settings
On each face's Agent tab, under Room:
| Setting | In profile.json | What it does |
|---|---|---|
| Room name | room.name | The room this face joins. Empty (null): it works alone. Letters, digits, ., _ and -, up to 40. |
| Also called | room.aliases | Other ways of saying this agent's name, separated by commas, for when speech recognition mishears it. Up to 10. |
| Who leads | room.lead | Whoever answers first (first, the default), or one agent that always leads (its name). The lead answers first and has the last word. |
Permission questions
Permission questions are spoken in turn like anything else, but the Allow and Deny buttons show at once. Your "yes" or "no" answers the agent that asked.
Things to know
- Interrupting by voice relies on telling your voice from the agents' speech coming through the speakers: the room compares what the microphone hears with the words being spoken, and only reacts to a name or stop word you say more often than they do. It works with speakers; headphones make it more reliable.
- Speech recognition can mishear a name. Add spellings under Also called.
- Every agent works on every question you don't address by name, so give each one its own working folder (or Codex the read-only mode) to keep them from editing the same files, and expect each question to cost tokens for all of them.
- Comments are prompted not to change files, but agents decide that for themselves.