microclaw-streams 0.1.2

Push-to-talk voice conversations powered by Whisper and Claude Code

MicroClaw Streams

Voice calls with Claude – talk it out. Voice in, voice out. Push-to-talk voice conversations powered by Whisper and Claude Code.

Whisper runs entirely on your machine — no audio is sent to the cloud. Only the transcribed text is sent to Claude.

Just three dependencies: numpy, sounddevice, and openai-whisper.

Setup

pip install microclaw-streams

Requires macOS (uses the built-in say command) and Claude Code to be installed and authenticated.

Usage

microclaw-streams
microclaw-streams --model turbo
microclaw-streams --resume <session-id>
microclaw-streams -l sv -e high

Options

Flag	Description	Default
-m, --model	Whisper model size (tiny, base, small, medium, large, turbo)	base
-l, --language	Transcription language (e.g. en, sv, de)	auto
-e, --effort	Claude effort level (low, medium, high, max)	low
-r, --resume	Resume a previous Claude Code session by its session ID	—
--fp16	Use half-precision inference (requires CUDA GPU)	off

Any additional flags you pass are forwarded directly to the claude CLI. For example:

microclaw-streams --max-turns 5 --permission-mode plan --model turbo

Here --model turbo is consumed by MicroClaw (Whisper model), while --max-turns 5 and --permission-mode plan are passed through to Claude Code. See claude --help for all available options.

Controls

Key	Action
Enter	Start/stop recording
A	Record with auto-approve (allows edits, writes, bash)
W	Record with web search enabled
T	Type a message instead of speaking
E	Cycle effort level
L	Cycle transcription language
Space	Interrupt speech output

How it works

1. Record — Press Enter to start recording, press Enter again to stop
2. Transcribe — Whisper transcribes your speech locally
3. Respond — Claude Code processes your message and streams a response
4. Speak — Voice output is spoken via macOS say command

Requirements

• macOS (uses say for text-to-speech)
• Python 3.10+
• Claude Code CLI

Example Conversation

Loading Whisper 'turbo' model locally (no audio leaves your machine)...
Ready!

ENTER=record  A=auto-approve  T=manually type input  E=effort [low]  L=lang [auto]
Recording ... press ENTER to stop.
Transcribing...
You: Hey what's up how come we have a bunch of html files in this repo?

          Voice: Let me take a quick look at what HTML files are in the repo.
          [Glob **/*.html]
          [Glob result]
          index.html
          demo.html

          Voice: Actually, there are only two HTML files — an **index.html** and a **demo.html**. Not really a bunch. Want me to look into what they're for?
          [5.5s | 4 in | 123 out | 10254 cached | $0.0729]


ENTER=record  A=auto-approve  T=manually type input  E=effort [low]  L=lang [auto]
Recording ... press ENTER to stop.
Transcribing...
You: Yup, for sure.

          Voice: Sure, let me check them out.
          [Read /home/user/index.html]
          [Read result]
          Claude requested permissions to read from /home/user/index.html, but you haven't granted it yet.

          [Read /home/user/demo.html]
          [Read result]
          Claude requested permissions to read from /home/user/demo.html, but you haven't granted it yet.

          Voice: Hmm, looks like I need permission to read those files. Could you grant me access?
          [6.4s | 4 in | 147 out | 20759 cached | $0.0158]


ENTER=record  A=auto-approve  T=manually type input  E=effort [low]  L=lang [auto]
Recording (auto-approve ON) ... press ENTER to stop.
Transcribing...
You: Here you go.