Agent Speak and the Claude Code plugin
Connect Agent Speak to Claude Code, reply through the HUD, and optionally hear short summaries spoken on your PC.
Connect Agent Speak
Agent Speak is an experimental Claude Code integration that shows Claude responses and questions in Whisperstream's HUD. You can dictate replies and optionally hear Claude's short summaries through local speech.
Agent Speak connects to Claude Code on your PC. Claude Code can send messages to its configured model service, and Lockdown Mode does not disable Agent Speak.
Install Claude Code and open it at least once, then:
- Open Experimental in Whisperstream.
- Find Agent Speak and select Configure.
- Under Connect Claude, select Connect.
- If Whisperstream says the plugin was installed or updated, run
/reload-pluginsin Claude Code. - Start or continue a Claude Code session.
The Agent Speak status is Ready, Off, or Attention. Follow any message shown below the status. Select Check for updates to refresh Whisperstream's plugin, or Disconnect to remove its connection without uninstalling Claude Code.
View responses and dictate replies
When connected, the HUD shows the current Claude response or structured question. Use the View full response shortcut to expand it. The default is Ctrl+Alt+\, and you can replace or clear it in Agent Speak settings.
Reply window controls how long your normal dictation shortcut replies to a just-finished Claude response. Set it from 1 to 15 seconds, or choose Off when you only want to view responses.
During the reply window, start dictation with your usual shortcut. Your dictated text appears in the HUD:
- With Auto-submit replies off, review the draft and press Enter to send it.
- With Auto-submit replies on, Whisperstream sends it when dictation finishes.
The same setting applies when Claude asks a question. Press Escape to close an active response or defer a question without answering it. Collapsing the expanded HUD returns it to the compact view without disconnecting Agent Speak.
Turn on local speech
Under Agent Speak > Configure > Speech, turn on Read Claude's replies aloud. Whisperstream asks before downloading the selected voice model. Speech is generated locally and reads a short summary rather than every word in the full response.
The first time you enable speech, Whisperstream recommends a model based on your computer's memory:
| Voice model | Choose it for |
|---|---|
| Higher quality | More natural speech on a computer with enough memory. |
| Lower memory | Faster startup and lower memory use. |
You can change Voice model later. Whisperstream may warn that Higher quality could strain the computer or require Lower memory when available memory is too limited.
After the download, choose a Voice, select Listen to sample, and adjust Speed and Volume. To change the playback destination, see Audio.
Choose which sessions speak
Use Sessions to control where spoken summaries are requested:
| Choice | Behavior |
|---|---|
| When I ask | The default. Run /whisperstream:speak in each Claude session you want to hear. |
| Every session | Requests a spoken summary from every connected session. |
Use these Claude Code commands with Agent Speak:
| Command | Result |
|---|---|
/whisperstream:speak | Turns on speech for the current session when using When I ask. |
/whisperstream:speak Project name | Turns on speech and gives the current session a name. |
/whisperstream:name Project name | Names or renames the current session. |
/whisperstream:name | Clears the explicit session name. |
/reload-plugins | Reloads Claude Code plugins after Whisperstream installs or updates its plugin. |
A named session displays that name in the HUD. Its spoken summaries also begin with the name, which helps distinguish responses when several Claude sessions are active.