Talk to Claude Code by voice: OpenWhisp is now an MCP server

If you use a coding agent, you know the rhythm: it works for a while, then stops to ask a question — “deploy to staging or production?”, “which of these three approaches do you prefer?” — and you type a paragraph back into a chat box. OpenWhisp’s new Agent Bridge closes that loop with your voice: the agent asks out loud, you answer out loud, and it keeps working. No other dictation app ships this.

What the Agent Bridge is

OpenWhisp now doubles as a local Model Context Protocol server — the standard way tools plug into agents like Claude Code, Cursor, Hermes, and OpenClaw. Register it once and your agent gets three new tools:

• openwhisp_dictate — the agent asks you a question by voice. OpenWhisp’s overlay opens with the prompt, you speak, and the transcript returns to the agent.
• openwhisp_refine — the agent rewrites text using the same on-device AI model as your refine hotkey. No other dictation app exposes this to agents at all.
• openwhisp_history — the agent reads your recent dictations (text, timestamp, target app), stored only on your Mac.

Set it up in one command

Turn the bridge on in Settings → Agent Bridge (it’s off by default — nothing listens until you enable it), then let OpenWhisp wire itself up:

openwhisp setup claude-code

That does two things: registers OpenWhisp as an MCP server (via claude mcp add), and appends a standing instruction to your ~/.claude/CLAUDE.md telling the agent to ask questions through openwhisp_dictate instead of plain text. The instruction line matters — agents only reach for tools they’re told to prefer. Both steps are idempotent, so re-running is safe, and --print previews the changes without writing anything.

openwhisp setup cursor merges the server into your project’s .cursor/mcp.json without disturbing other servers you have; Hermes and OpenClaw get printed instructions. The first time an agent connects, OpenWhisp asks you to approve it — once, always, or only while the app runs.

Example 1: answer your agent without touching the keyboard

You ask Claude Code to ship a release. Mid-task it needs a decision, so instead of parking the question in the terminal and waiting for you to notice, it calls openwhisp_dictate:

● openwhisp_dictate("Tests pass. Deploy to staging or production?")
  ↳ the OpenWhisp overlay opens — you say:
    "production, and tag it v1.4"
  ↳ you go quiet — the answer is sent on its own
● Deploying to production, tagging v1.4…

You were reading a doc in another window; you answered in three seconds without switching apps or pressing a single key — when you stop talking, OpenWhisp detects the silence and ends the turn for you (a natural pause between words won’t cut you off; toggleable in Settings). Prefer an explicit finish? openwhisp dictate --stop from any shell sends what was captured, and --cancel discards it. This is the difference between an agent that waits for typed input and one you can talk to.

Example 2: brain-dump a bug report, let the agent file it

Describing a bug is faster out loud than in prose. Tell your agent “ask me about the bug, then file a GitHub issue” and it will interview you by voice — what happened, how to reproduce it, what you expected — then turn your rambling answers into a clean, structured issue. You talk for ninety seconds; it writes the report.

Example 3: your dictation AI, in any shell pipeline

The bridge also ships an agent-callable CLI, which means the refine step works anywhere you can pipe text — no agent required:

# make the clipboard formal, in place
pbpaste | openwhisp refine -i "make it formal" | pbcopy

# tighten a commit message before you commit
git log -1 --format=%B | openwhisp refine -i "tighten this up"

The rewrite runs on the same on-device model as OpenWhisp’s refine hotkey — your prompts, your machine, nothing sent anywhere.

Example 4: turn your dictation history into a standup

Everything you dictate is already in OpenWhisp’s local, searchable history. With openwhisp_history, your agent can use it: “read my dictations from this morning and draft a standup update” — and the notes you spoke into tickets, commit messages, and chats all day become a summary you didn’t have to write.

Built so agents can’t go rogue

Handing agents a microphone and an AI model deserves paranoia, so the bridge is engineered around consent:

• Off by default. No socket exists until you enable the bridge in Settings.
• You approve every client. Each agent is approved the first time it connects, and you can revoke any of them in Settings.
• Local, signed connections only. The transport is a private UNIX socket on your Mac — not a TCP port — and by default only OpenWhisp’s own code-signed CLI may connect, so browser pages and random processes can’t reach it.
• The cloud gate. If your AI provider is OpenAI, agent-initiated refinement is blocked unless you explicitly allow it — a prompt-injected agent can’t exfiltrate text through your API key. Local models are unaffected.
• The human always wins the mic. Pressing your dictation hotkey during an agent session cancels it — the agent gets nothing — and starts your own. Agent microphone use always shows the overlay, and password fields are never dictated into.

The full design is documented in docs/AGENT_BRIDGE.md — and because OpenWhisp is open source, you can read the implementation rather than take the security story on faith.