Skip to content

Voice dictation

Mjolnir can transcribe a spoken prompt locally and insert the result into the composer. Press Ctrl-R to start dictation and press it again to stop. Review or edit the transcript before sending it to the selected agent.

PlatformSupportInstallation note
macOSSupportedThe release installer includes mj-voice-worker; Cargo users install it separately
LinuxSupportedThe release installer includes mj-voice-worker; Cargo users install it separately
WindowsSupportedUse a release archive containing the worker or install both Cargo crates
AndroidNot currently availableAndroid builds omit the desktop voice worker and hide the Ctrl-R shortcut

For a Cargo installation:

Terminal window
cargo install --locked brokk-mjolnir brokk-mj-voice-worker

The worker must be installed beside mj. Developers can point to a separate binary with MJ_VOICE_WORKER.

The voice worker records microphone audio only after Ctrl-R activation. It runs the speech-recognition engine in a separate local process, returns transcribed text to Mjolnir, and does not send the audio to any agent. Only the prompt text you choose to submit enters the agent conversation.

The recognition model is downloaded on first use and currently requires about 0.7 GB of cached assets. The sidecar process keeps a native speech-engine crash from taking down the active Mjolnir session.

Voice input changes how a prompt enters the composer; it does not bypass normal permissions, workspace limits, review, or provider data boundaries.