⌜ Codescribe ⌟
Native macOS dictation and voice conversations with agents.
Download ·
User guide ·
Online guide ·
Releases
Codescribe turns speech into text, displays it in a floating overlay and delivers
it to the destination you select. It also connects existing coding-agent
conversations to numbered channels, with written messages, attachments and
spoken replies. A separate built-in Agent provides its own model, history and
tools.
Start using Codescribe
- Download a package from the website or GitHub Releases, copy the app to
Applications and launch it from Finder or Spotlight.
- Follow Settings → Creator for permissions. Check Audio for the input
device and Dictation → Engine for recognition.
- Hold Fn/Globe, speak a short sentence, then release. Review the result
and the configured paste policy before relying on delivery to another app.
A packaged app does not require Rust or Xcode. Continue with the
Quick Start or
installation guide.
Choose recognition, formatting and destination
| Choice | Available paths |
|---|
| Speech recognition | Apple only (fresh-install default), Local power (local Whisper), Cloud (online STT with audio-egress consent) |
| Text formatting | Off / Correction / Smart automatically; Max is an explicit per-take Agent consultation |
| Destination | Another app under the selected paste policy, the built-in Agent, or an explicitly chosen channel/audience |
Apple only does not require downloading Whisper. For Local power, use
Dictation → Whisper to download and validate a model. A selected mode,
installed model and loaded recognizer are separate states. Developer options
can add recognition observations or acoustic admission; they are not a required
relay through every engine.
Local recognition can run without a cloud AI service once its dependencies and
permissions are ready. Formatting, remote agents, cloud transcription, speech
synthesis and update checks have separate network boundaries. See
Privacy & Security.
Live recognition can revise text. A successful processing status is not proof
that every word was preserved. Review names, numbers and content inferred from
quiet audio; AI formatting does not replace comparison with the original take.
Conversations and channels
Show Agent opens Codescribe's built-in assistant. Configure its model,
workspace and tool permissions in Settings → Agent, and accounts/keys in
Providers.
For an existing external session, connect its Codescribe integration under a
name and channel 1–9. The main roster, compact navigation and floating agent
cards are views of those same conversations. The numbered rail remains an
entry point when the sidebar is narrow. A detached agent retains its roster
entry so you can bring its existing window forward.
0 · All is a shared channel with audience rules. Your inbox (10) collects
messages explicitly addressed to you. Neither means “select every agent”.
Viewing a conversation, choosing written recipients and choosing a microphone
destination are distinct actions.
The composer's recipient menu supports one or several agents. Sync shares
text and attachment drafts across the selected group while retaining private
drafts. Inspect recipients before sending. Published, queued and read/ACK are
different states; an ACK is not completion of the requested work.
See Chat & Overlay for recipient selection, floating
cards, delivery states and attachment previews. Markup for image attachments
is nightly-only. This guide follows development source; your installed public
release can have an earlier interface.
Connect an external agent
Use the matching Codescribe BUS/plugin package and the client's setup flow.
Installation alone does not attach a conversation or prove that a running client
loaded the update. Keep the session's identity, lease, channel and preferences
when reconnecting it.
This is the beginning of the README. Read the rest at the original. Read the full README ↗
