Alex SaintBack to @alexsssaint

GlassTerm

macOS·Apple Silicon·voice first

You talk. The agents type.

Transparent terminal panes over your wallpaper, each one a real login shell running an agent CLI in one of your projects. You never type into a pane. You say what you want, and Molly writes it in.

  • Freeno price, no plan
  • Open sourcesoon
  • Mac App Storesoon
GlassTerm running: a radio card on the left, an agent pane on the right with a Claude Code session in it, the Settings window open on API Keys, and the voice bar at the bottom reading Listening.
GlassTerm on a real desktop: a pane running an agent in ~/alexsssaint, the radio card, the key sheet that says where keys live, and the bar at the bottom listening. Everything is glass, so the wallpaper stays visible through the work.

Instead of a screen full of terminal tabs

What GlassTerm is

GlassTerm is a window made of glass panes. Every pane is a real login shell, started in one of your projects, running one of six agent CLIs: claude, codex, grok, kimi, qwen or hermes. Several agents work side by side, on camera, which is what it was built for.

Speech to text runs on this Mac. Holding the fn key streams your voice into a local Parakeet server on MLX, and the text goes to Molly, the single brain. She decides what to do with it, then writes into the pane herself through her send_to_terminal tool.

There is no agent-to-agent bus. Molly is the only hub. Panes share state one way, through a read-only roster and a task board, so nothing an agent says can steer another one behind your back.

It replaces a screen full of terminal tabs, and the hand that has to drive each one.

Capabilities

What it does

  • Push to talk

    Hold the fn key through a compiled Swift key server, or hold the orb in the bar. Click it instead and it stays open, hands free.

  • Six agent CLIs

    claude, codex, grok, kimi, qwen, hermes. Claude panes authenticate through your Claude subscription, never through an API key.

  • A brain you pick

    Molly runs on Claude CLI drivers (fable, opus, sonnet, haiku), OpenAI, xAI, Moonshot, DashScope or Nous. Haiku is the default. Grok Voice is the realtime option, where the model is its own voice.

  • Ten voices

    Five from xAI (Eve, Ara, Rex, Sal, Leo) and five from OpenAI (nova, shimmer, coral, sage, onyx), streamed as PCM. A Swift AVSpeech server and the macOS say command are the gated fallbacks.

  • Spoken while it is written

    Rendering is speculative: sentences are synthesised while the model is still writing the next one, so the reply starts before the answer is finished.

  • A local fast path

    Simple commands are matched by a regex in the intent module and never reach a model at all. No call, no cost, no wait.

Where your material goes

Everything that leaves, named

On this Mac

2 crossings

  • Your voice
  • Parakeet speech to text (MLX)
  • Every pane and every shell
  • The task board and the roster
  • Molly’s sessions and memory

The line

Off this Mac

  • The transcribed text

    goes to the model you picked in Settings, on your own key

    Always

  • Molly’s reply text

    goes to xAI or OpenAI, to be turned into a voice

    Only if you switch it on

Audio is never uploaded anywhere. Text is, to the providers you chose and paid for. Pick a local voice and the second crossing closes.

How it works

The pipeline, as built

What happens when you speak

  1. fn key

    held, or the orb in the bar

  2. Parakeet

    speech to text, local, on MLX

  3. routeMessage

    every utterance goes through here

Fast path

A local regex answers it

No model call, no cost, no wait.

Everything else

Molly decides

One brain, one hub. She calls her tools herself.

send_to_terminal

She types into the pane. You never do.

streamed PCM

She talks back while she is still writing.

Panes never speak to each other. They share one read-only roster and one task board, and that is the whole of it.

Requirements

What it runs on

SpecDetail
RequiresmacOS on Apple Silicon. The speech server runs on MLX.
Also needsNode 18 or newer, and the Xcode command line tools for node-pty.
Speech to textparakeet-tdt-0.6b-v2, started locally when the app is ready.
State~/Library/Application Support/glassterm: settings, projects, the board, Molly’s memory.
Sidecars~/.glassterm: the speech server, the key server, the voice server, the roster, the timing log.
KeysOne-line files under ~/.claude, entered through Settings. Never sent anywhere but their own provider.

Honest

Not there yet

  • Speech to text is English only.
  • Dark mode only.
  • No first-run onboarding yet.
  • The build is development-signed, so today it runs on my Mac and not on yours. That is the thing the App Store submission has to fix.

Status

Open source, and on the store

GlassTerm is being opened up. The repository is private today, so there is nothing to clone yet and no public build to download. When it opens it will be at github.com/AlexSSSaint/glassterm, under a licence that lets you read it, build it and keep it.

The Mac App Store version is coming after that, so it installs and updates like anything else on your Mac, without a certificate dance. No date, because a date I have not earned is just a way of being wrong later.

Want it the day it ships?

← All three apps

Surface 02 · Abdomen · sellme.lol ↗