A voice-first assistant that actually does the work on your Mac.

Asha is the voice layer for the coding agent you already pay for — keep your OpenCode plan and key, talk instead of type, hear the diff back.

Asha doesn't replace your terminal agent — it talks to it: you speak a coding task, Asha delegates to the agent running on your own key (including your OpenCode key), and reports the result back by voice.
Speech is on-device (Moonshine STT + Kokoro TTS in-process — no cloud speech bill), and Asha carries whole-day context your terminal never sees: mail, calendar, screen, camera, memory.
MIT-licensed and open source, same as OpenCode — one Mac app to install, no terminal required, no new subscription required if you BYOK.

MIT-licensed · Bring your own model key · macOS, Apple Silicon

How it works

You talk. Asha runs the errand.

  1. Speak or type

    Push to talk, or type instead. Asha transcribes on the spot and shows you what it heard — the speech-to-text runs on your Mac.

  2. Asha plans

    A supervisor breaks the request into steps and briefs a small team of specialist agents with exactly what each one should do.

  3. The team delivers

    Agents work in parallel, then report back in one place. Asha reviews the result and tells you what it did — out loud.

Measured, not promised: 14/14 tasks on our internal benchmark, using roughly 3–4× fewer tokens than the same-model baseline. Method and logs are in the repository.

Why Asha

Honest differences, not marketing ones.

  1. Voice-first

    Built around talking, not typing — push-to-talk with on-device transcription.

  2. Speech stays on your Mac

    Transcription and speech synthesis run locally. The voice loop itself doesn't need the cloud.

  3. Your keys stay yours

    Bring your own model key. It lives on your Mac, and calls go straight to the provider you chose — no Asha server in the middle.

  4. An agent team, not a chatbot

    A supervisor delegates to specialists and reviews the work before anything ships.

  5. Connected where it counts

    Gmail and Google Calendar flows run against your own Google project; Figma files are inspected over the official REST API with your own token.

  6. It makes media too

    Images with your own Leonardo key (with a free fallback), plus an optional local video add-on you run yourself.

Zero switching cost

Keep your existing plan. Add a voice.

Asha works with the agent you already pay for. Bring your own key — an existing OpenCode (or OpenRouter) key works — and nothing you rely on today has to change.

  1. Install the app

    Download Asha.app, drag it to Applications, open it. No terminal, no curl | sh, no dev paths.

  2. Paste the key you already have

    Settings → agent key. OpenCode or OpenRouter both work; the key is stored in a private file on your Mac — never logged, never echoed.

  3. Talk to your agent

    Pick your project, speak a task, watch the steps stream in, and hear a short spoken summary of the diff. Say “undo” to revert.

Open source · BYOK

MIT-licensed. No lock-in, no new subscription.

Asha is free, open-source software. There are no plans, prices or seats on this page — download it, bring the model key you already own, and keep the coding agent you already pay for. An existing DeepSeek, OpenRouter or OpenCode key all work.

  1. Voice-first

    Push-to-talk with on-device transcription. Speak a task, hear a short spoken result.

  2. Delegates to your agent

    Asha hands coding work to the agent you already run and reviews the diff it gets back.

  3. Sees screen and camera

    Local screen reading, plus camera input when you want it to look at something real.

  4. Connectors and memory

    Gmail, Google Calendar and Figma, plus whole-day context your terminal agent never sees.

Nothing to give up: your existing install, plan, key, terminal habits and repo stay as they are. Asha is MIT-licensed, so there is no new subscription — you pay your model provider directly, if you pay anyone at all.

Download

Install it and talk.

Asha for macOS

Version 0.1.0 · Apple Silicon · ~900 MB

Our everyday build. Everything is bundled, including the voice models, so it works offline once installed.

Download the .dmg

First launch: this build isn't code-signed yet, so macOS will warn you. Open the .dmg, right-click the Asha app, choose Open, then confirm.

Asha for Windows

Not published yet

We'll publish a Windows build once it has been tested on real hardware. No placeholder download, and no "coming soon" button that goes nowhere.

Pending hardware testing

What happens on first launch

  1. Grant microphone access when macOS asks.
  2. Add one model key in Settings — that's your bring-your-own-key setup.
  3. Talk to it. Or type, if you'd rather.

FAQ

Questions developers actually ask.

Is Asha free?

Yes. Asha is free, open-source software under the MIT licence. There are no paid Asha plans and no Asha account — you only pay third parties directly if you use them, such as your model provider or image provider.

Do I need an API key?

No Asha account is needed. Bring your own key from a model provider (BYOK): paste it into Settings and it stays on your machine. DeepSeek, OpenRouter and OpenCode keys all work.

What runs locally?

Speech-to-text and text-to-speech run on your Mac, and screen reading is local too. That's what lets the voice loop stay responsive without shipping your audio to a cloud service.

Is my data sent anywhere?

Only what you choose to send. With your own key, requests go straight from your Mac to the provider you configured, under their terms. We don't operate an Asha account or an Asha analytics service, and we don't receive your conversations. This website has no analytics or trackers.

What are the requirements?

An Apple Silicon Mac (the current build is arm64), a microphone, and one model key. An Intel build can be produced but hasn't been tested, so we don't recommend it yet. The packaged app doesn't declare a minimum macOS version — if it installs and boots for you, it works. Tell us if it doesn't.

How do I set up a connector?

Gmail and Google Calendar connect through your own Google project, so the setup is yours to own. Google's public-app verification is still in review, which means these flows currently suit personal and test use. Figma is simpler: paste a personal access token and Asha inspects files over Figma's REST API.

How is this different from a chatbot?

A chatbot answers. Asha acts and reports back spoken. It plans a task, delegates pieces to specialist agents, reviews the result, and can touch real things — mail, calendar, files, and generated media. The conversation is the interface; the work is the point.