VoxBee turns your voice into text and text back into speech. As of v0.9, your coding agent can do both too. VoxBee now speaks the Model Context Protocol (MCP), which means agents like Claude Code, Codex, and Cursor can call VoxBee's voice features as native tools — transcribe an audio file, read text aloud, or stop playback — the same way they call read_file or run_command.
Tell your agent "summarize this PDF and read me the summary" and it does, out loud. No copy-paste, no separate app to babysit.
One Line to Connect
Open the new API tab in VoxBee. It gives you a ready-to-paste connect command for each agent with your token already filled in. For Claude Code, that's a single terminal command:
claude mcp add --scope user --transport http voxbee http://127.0.0.1:5111/mcp \
--header "Authorization: Bearer YOUR_TOKEN"The --scope user flag registers VoxBee once for every project, so you don't have to re-add it per directory. Codex and Cursor get their own copy-paste config snippets in the same tab — a TOML block for ~/.codex/config.toml and a JSON block for ~/.cursor/mcp.json.
The Tools Your Agent Gets
Once connected, your agent sees a small, focused toolset:
transcribe— point it at an audio file on disk and get the text back. Runs on whatever STT model you have installed in VoxBee.speak— hand it text and a voice, and VoxBee reads it aloud through the floating playback dock.stop— stop whatever is currently playing.list_modelsandlist_voices— discovery tools so the agent can pick the right model or voice before it acts.
Agent Audio Routes Through the Dock
When your agent plays audio, it doesn't hijack your speakers with no off switch. Agent-played audio routes through VoxBee's floating playback dock, so you can pause or stop it from any page in the app — exactly like audio you started yourself. The agent gets a voice; you keep the remote.
Prompts to Try
These are the kinds of things that just work once VoxBee is connected:
- "Read README.md aloud."
- "Summarize this PDF and read me the summary."
- "Transcribe voice-memo.m4a and save the text next to it."
- "Read this PR description out loud while I look at the diff."
It All Runs On-Device
The MCP server is a thin protocol adapter over the same engines that power VoxBee's app. Transcription uses your local Whisper or Parakeet model; the System voice synthesizes speech on-device with no API key. Nothing leaves your Mac unless you've explicitly configured a cloud STT or TTS provider with your own key — and even then, there's no VoxBee middleman and no per-minute billing.
The server binds to loopback (127.0.0.1) by default and every request is authenticated with a bearer token, so a random process on your network can't call your microphone or speakers.
MCP Is Half the Story
The MCP endpoint sits on top of VoxBee's new local API server — an OpenAI-compatible HTTP server that exposes the same transcription and text-to-speech over plain REST. If you'd rather call VoxBee from a script or any tool built for the OpenAI SDK, that's the post to read next.
Download VoxBee and try it free for 14 days. The System voice and on-device transcription work with no API key — connect your agent and start talking to it.