Agent Speak for Codex and Claude Code

Connect Agent Speak#

Agent Speak shows Codex and Claude Code replies in Whisperstream's HUD. You can read a short written summary, hear it through local speech, and dictate a reply after an agent finishes.

Agent Speak connects to coding agents on your computer. While Lockdown Mode is effective, Whisperstream suspends Agent Speak processing. This does not stop a connected agent from using its configured model service independently.

Install Codex or Claude Code and open it at least once. Then:

1. Open Agents in Whisperstream and select Enable Agents or Continue setup.
2. Under Connect your coding agents, select Connect for Codex, Claude Code, or both.
3. For Codex CLI, run /hooks and trust Whisperstream. In the Codex app, open Settings > Hooks > Whisperstream and select Trust all. For Claude Code, run /reload-plugins if Whisperstream says its plugin was installed or updated.
4. Choose Show summaries, Read replies aloud, or both. Setup starts with both selected and requires at least one. If you choose speech, select a local voice model and voice, then adjust Speed. Continue starts the model download if needed.
5. Set the shortcut that opens the full response, then follow the final on-screen test with your connected agent.

The test checks your selected response mode and voice replies. Claude Code also presents a practice structured question. Codex does not provide structured questions to Agent Speak.

After you select Connect, Agent Speak stays on while Codex hook trust is pending. The Codex row changes to Connected automatically after trust is confirmed. After setup, use General, Voice, and Connections in Agents to change settings. To repair a connection, open Connections, select Disconnect, then Connect again. Follow any status instruction shown below the agent.

Pause Agent Speak#

To stop Agent Speak for a while without disconnecting anything, open Agents > General and turn off the Agents switch. Summaries, speech, and replies stop for every coding agent, and Connections shows each agent as Paused. Your connections, Codex hook trust, and voice and summary settings are kept. Turn the switch back on to resume with the same agents.

Pausing does not stop Codex or Claude Code themselves. To remove one agent's connection instead, use Disconnect in Connections.

View responses and dictate replies#

Turn on Show summaries to display the agent's short written summary when it finishes. This is independent of Voice, so you can read summaries without hearing them. New installations start with both settings off; completing setup saves your selections.

Use View full response to expand the complete reply, even with automatic summaries off. The default shortcut is Ctrl+Alt+backtick, and you can replace or clear it in Agent Speak settings. Existing saved shortcuts are not changed automatically.

Reply window controls how long your normal dictation shortcut answers the response that just finished. Choose 1 to 30 seconds, or Off to disable temporary replies. Setup selects 20 seconds with written summaries or 6 seconds with speech alone. In General, enabling summaries changes a 6-second window to 20 seconds; disabling them changes 20 back to 6. Other values stay unchanged.

During that window, start dictation with your usual Push to Talk or Toggle shortcut. Your draft appears in the HUD:

• With Send replies automatically off, review the draft and press Enter to send it. This is the default.
• With Send replies automatically on, Whisperstream sends it when dictation finishes.

Claude Code can also send structured questions to the HUD. Expand the question, then press the displayed number for your answer. Codex structured questions remain in Codex. Press Escape to close the response or defer a Claude Code question without answering it.

For replies after this temporary window closes, see Coding agent sessions.

Configure local speech#

In Agents > Voice, turn on Voice to hear replies through local speech. Use What to read to choose Spoken summary, Full response when short, or Full response. Code blocks and raw links are skipped.

Choose a Voice model:

Lower memory is the default on Windows, and Automatic is the default on Mac. Choosing Automatic or Higher quality downloads Kokoro from Hugging Face; it has no automatic backup source. Each voice model keeps its own voice choice.

Change Voice model, Voice, Speed, and Volume here. Voice previews are available in the picker. For the playback destination, see Audio.

A voice model stays loaded between replies and unloads after 30 minutes without speech to free memory. To change this, expand Advanced in Agents > Voice and set Unload voice model when idle to Never, 5 min, 10 min, 15 min, 30 min, or 1 hour. After the model unloads, the next spoken reply may take a moment to start.

Use Speak in to choose when speech is requested:

In Claude Code, /whisperstream:speak Project name also names the session. Use /whisperstream:name Project name to name or rename it, or /whisperstream:name to clear that name. Named sessions show their names in the HUD, and their spoken summaries begin with the name.