0.1.0 · Apple Silicon · MIT

PSST

Talk with your agent.

PSST is dictation and speech for coding agents: Claude Code, Codex, Grok, Copilot in VS Code, Antigravity and Pi.

It types what you say into the agent's prompt, and the agent speaks short replies back while it works.

Install Source on GitHub

00 / Dictate

Hold Space to dictate.

A soft note says PSST is listening. Each phrase is typed when you pause, and the rest when you let go.

PSST never presses Enter, so you review before you send. The key, and holding versus tapping, are settings.

01 / Speak

The agent talks back.

It speaks first in every turn, a sentence or two, and once more at the end when it began by saying what it would do.

Each voice has an accent, and the agent takes a persona to suit it: dry British for bf_emma, upbeat can-do American for af_heart, faintly unimpressed French for ff_siwis, and so on for each of the 8 accents. Models drift now and then, and nothing enforces it.

It speaks in your language when PSST speaks it, and in English otherwise.

02 / Local

Nothing you say leaves your Mac.

Everything runs locally: Whisper listens and Kokoro speaks on your Mac. The only network traffic is model downloads.

The local server only answers your own processes. It binds to 127.0.0.1 and checks that the process on the other end of each connection is yours.

03 / Install

Build it from source.

PSST needs a Mac with Apple Silicon and macOS 14 (Sonoma) or later. 0.1.0 is a source release. Signed binaries and a Homebrew formula come later.

To build it you need Rust stable and cmake (brew install cmake):

git clone https://github.com/lineofflight/psstcd psst./scripts/install.sh

The script builds and signs PSST, installs it to ~/.local/bin/psst and starts it. The first start downloads about 1 GB of speech models, and psst start says how far it has got:

PSST is starting: downloading the whisper model, 90 of 547 MB
...
PSST is starting: loading the models
PSST running (pid 4242).

macOS then asks for Accessibility, and for the Microphone at your first dictation. Allow both.

04 / Agents

Connect your agent.

PSST serves one MCP server, at http://127.0.0.1:7778/mcp. Register it with your agent once PSST is running. For Claude Code, Codex, Grok, Copilot in VS Code and Pi, one command is the whole setup:

claude mcp add -s user --transport http PSST http://127.0.0.1:7778/mcpcodex mcp add PSST --url http://127.0.0.1:7778/mcpgrok mcp add PSST http://127.0.0.1:7778/mcpcode --add-mcp '{"name":"PSST","type":"http","url":"http://127.0.0.1:7778/mcp"}'pi mcp add PSST --url http://127.0.0.1:7778/mcp

Antigravity takes a small plugin, from your PSST checkout:

agy plugin install ./integrations/agyagy plugin enable psst

Say hi in a new session, and the agent speaks first.

Not yet in 0.1.0: