Voice orders: hold a key, say what you want, T3 Code does it

[leorivastech]

Hold a key, speak, let go: T3 Code carries out short spoken orders. It does not talk back and it does not decide anything. It listens and takes the action, the same one you would take with a shortcut.

This is built and working. It runs in the app installed on my machine and I use it every day. It is not a sketch looking for someone to write it.

This is the push-to-talk idea from #7417, and it goes a step further: it covers the actions around the message, not just the typing.

What you can say

• Threads: "three", "open the second one", "next", "stop", "close it"
• Messages: "write: check the login test" leaves text in the composer, "send it" sends it, "tell it to run the tests" does both
• Model and effort: "switch to Opus", "switch to GPT", "set effort to high"
• Scrolling: "up", "down a lot", "go to the top", "to the bottom"
• Screenshots: "take a screenshot", "paste it" attaches it to the draft
• Chained: "open the second one and tell it to continue"

English and Spanish, no language setting. A toast shows what was understood before it acts.

Why

I talk to my agents all day, and most of what I do between messages is small: change thread, scroll, switch model, send. Dictation on the iPhone exists. On desktop there is nothing, and push-to-talk is the most-upvoted voice idea here (#7417). People there also asked for dictation in their own language and not only English; this works in Spanish and English without a setting.

How it works

• Settings → Integrations → Voice: one key (Groq or OpenAI), kept with the environment's other secrets. Removing it turns voice off.
• Key to hold: per device, Right Ctrl by default. A tap is ignored and a chord stays a normal shortcut. It only listens while T3 Code is in front.
• A bare number, "up" and "down" run on local rules, with no model call.
• Every order uses an existing command, so it needs its shortcut bound in Keybindings.

What it will not do

No order approves a request, answers a pending question or changes an access mode. Those stay in the app.

Privacy

Each order sends the recording and its text to the provider the key belongs to. T3 Code keeps neither.

A fully local option

Hearing and understanding both go through one server service that talks to an OpenAI-compatible endpoint. So everything can be local: point it at a local transcription server for the hearing and a local model for the understanding, with no key and nothing leaving the machine. The numbers, "up" and "down" already need no model at all. I have not built the base-URL setting yet, but I think it is a small change and I am glad to do it if you want it.

How it fits the codebase

• apps/server/src/voice: the interpreter and the rules, with tests
• apps/web/src/voice: hold-to-talk and carrying out the orders
• Contracts for the voice settings and one RPC method, authorized like the rest
• Screenshots need the desktop app

Happy to split it, trim it, or change the shape if you would rather have it differently.