Follow work away from the terminal
Hear task starts, milestones, questions, errors, and completion summaries without reading every line.
Spoken output for Claude Code
Claude Code can listen. Now it can talk back.
Claude Code includes voice dictation, but its coding responses remain text. TalkToCursor adds spoken updates in a managed, ElevenLabs, or local voice.
you: Fix the failing authentication test.
speak: I found the session regression and I’m applying the fix.
COMPLETE THE LOOP
Keep native dictation. Add selectable speech.
TalkToCursor does not replace Claude Code. It gives Claude a local MCP tool for spoken task feedback while the normal terminal workflow continues.
Pick the speech provider
Use managed TalkToCursor Cloud, bring an ElevenLabs key, or keep speech local with Voicebox.
Control narration with CLAUDE.md
Tell Claude exactly when to speak and keep code, file paths, logs, and stack traces out of audio.
CLAUDE CODE SETUP
Connect the existing package.
No Claude-specific fork is required. Register TalkToCursor as a user-scoped stdio MCP server to make it available across projects.
Add TalkToCursor to Claude Code
The --scope user option makes the voice tool available in all of your Claude Code projects.
claude mcp add --scope user talktocursor \
-- npx -y talktocursorPrefer a single project? Omit --scope user. Claude Code uses local project scope by default.
Choose your speech provider
Open the shared settings app, then select TalkToCursor Cloud, ElevenLabs, or Voicebox.
npx -y --package talktocursor talktocursor-settingsTell Claude when to speak
Add this to your global ~/.claude/CLAUDE.md or a project’s CLAUDE.md.
## Voice feedback
Use the `speak` tool at task start, at important
milestones, when asking a question, and at completion.
Keep spoken updates concise. Do not read code, paths,
logs, or stack traces aloud.Confirm the connection
Restart Claude Code, list its MCP servers, then ask for a short spoken test.
claude mcp list
# In Claude Code:
Use the speak tool to say:
"Claude Code voice is connected."
WHAT EACH PART DOES
Native input, TalkToCursor output.
| Need | Claude Code | TalkToCursor |
|---|---|---|
| Voice input | Native dictation where available | Optional external dictation remains Cursor-focused |
| Spoken coding responses | Text output in the terminal | Concise agent-initiated TTS through MCP |
| Voice choice | Not applicable to text responses | Cloud, ElevenLabs, or local Voicebox |
| When the agent speaks | Controlled by your instructions | The speak tool runs when Claude calls it |
DETAILS
Before you connect.
Does Claude Code already support voice?
Claude Code offers voice dictation where available, but not native spoken output for coding responses. TalkToCursor adds text-to-speech through MCP.
Does this work with Claude on the web or mobile?
This setup is for Claude Code, which can launch local stdio MCP servers. Claude’s consumer web and mobile voice experience is separate.
Does TalkToCursor speak every response automatically?
The agent calls the speak tool according to your instructions. The recommended rule focuses on useful milestones instead of reading long terminal output aloud.
Give Claude Code a voice you want to hear.
Connect one MCP package and keep the terminal workflow you already use.