Actions & automation

OpenWhisper can do more with your words than paste them. An action records, then sends what you said somewhere else: a new Claude or ChatGPT chat, your coding agent, a running Claude Code session, a webhook, a program, or just your clipboard. Your own tools can also follow what OpenWhisper does, start recordings, and adjust dictated text before it lands.

Everything here runs on this computer and listens only locally. Nothing is reachable from the internet unless a webhook or program you set up sends it there.

Create an action#

Open Settings → Automate → Actions, add one from a preset, and give it a shortcut. Press the shortcut anywhere: OpenWhisper records (press again to stop), then hands your words to the action instead of pasting them.

Try it runs an action with sample words, without saving anything to History. Every real run is saved in History as Action · name, with your original words alongside.

Shape the words and the answer#

Each action chooses how your words are prepared: Raw (dictionary and snippets only), Cleaned up with your normal AI cleanup, with a cleanup profile, or cleaned up and then rewritten by a transform such as Prompt engineer.

Actions that answer (an agent's Quick answer, a webhook, a program) can Paste the answer into the app you started in, show it in the Scratchpad or a notice near the pointer, or copy it.

Set up the agent hotkey#

The Agent hotkey card at the top of the Actions page finds Claude Code, Codex, or OpenCode on this computer and sets up one shortcut for it. Speak a request and it opens a session with your words, in the agent's app when that's set up or a new terminal otherwise, or asks for a quick answer. It can include your selected text and the app's name, and start in a working folder; Create a new folder automatically gives each request its own empty folder under ~/OpenWhisper Agent Tasks.

Route by voice#

Turn on Route by voice and start a request with an action's name: “Claude, what rhymes with orange,” “todo: buy milk.” Use the Route by voice shortcut, or turn on Also listen to normal dictations so any dictation that starts with an action's name is routed. Anything that names no action is pasted as usual.

Routing uses Fast judgments (set up in Settings → Meeting Mode → Fast judgments) and gives up after 1.5 seconds, pasting your words instead. A short description on each action (“questions for Claude,” “my shopping list”) makes routing more reliable.

Hooks: webhooks, programs, and text filters#

Settings → Automate → Hooks connects OpenWhisper to your own tools:

• Webhooks receive the events you choose (finished dictations by default) as JSON, signed with Standard Webhooks headers and retried on network errors. Send test posts a test event, and the page lists recent deliveries.
• Programs receive each event as JSON on standard input, one at a time.
• Text filters can change dictated words before or after AI cleanup, via a web address or a program. A filter that's slow or fails keeps your words unchanged.

Hooks are created only in Settings; the API and agents can't add or change them.

Live API and MCP#

Turn on Settings → Automate → API & MCP to serve a local API and an MCP server on 127.0.0.1:8767, behind an access token shown on that page.

• Follow live events (on): stream recordings, finished dictations, actions, and meetings as Server-Sent Events, with resume after a disconnect.
• Run actions, Start recordings, Change dictated text, and Send words to my coding agents: each off until you allow it, and checked on every request.

Agents connected over MCP can search your history, transcripts, and meeting insights, see whether OpenWhisper is recording, and, only with those permissions, run actions or start dictation. MCP Events (preview, off by default) lets an agent subscribe to events delivered as signed webhooks.

Dictate into a running Claude Code session#

A Claude Code channel (research preview) sends your words into a Claude Code session that's already running:

1. Turn on API & MCP. To let Claude answer near your pointer, also allow Run actions.
2. Copy the .mcp.json entry from API & MCP → Claude Code channel into the project you'll work in.
3. In that project, start Claude Code with the command shown on that page.
4. Add a Tell my Claude Code session action and give it a shortcut.

Claude is told which messages you dictated and which came from another program. The channel needs a claude.ai login or Console API key and doesn't work with claude -p.

Reference#

The automation reference documents every event, payload, API route, and signature check, and MCP setup covers connecting Claude, ChatGPT, and other MCP clients.