Skip to main content
The Voice Commands plugin lets you define commands like “post to Slack…” or “add a reminder…”. When your dictation matches a trigger, Freestyle runs the matching action instead of typing the text.
“post to Slack that I’ll be five minutes late” → posts “I’ll be five minutes late” to your Slack webhook “add a reminder to buy milk” → runs your Add to Reminders macOS Shortcut with “buy milk”

​
Install

Or install from Settings → Plugins inside the app.

​
How it works

Detection runs on the server’s afterTranscribe hook, in three stages:
  1. Prefilter — a cheap, deterministic phrase match gates everything. If your speech contains no trigger phrase, nothing else runs and the transcript is dictated normally. No LLM call, no added latency.
  2. Agent — when a trigger matches, a multi-step tool-calling agent (using the LLM you’ve already configured for cleanup) decides whether it’s really a command and extracts the payload — the part of your speech that isn’t the trigger words.
  3. Action — the command fires and the utterance is consumed (via api.control.consume()): the rest of the pipeline is skipped and no text is delivered to the focused app.
If no LLM is available, detection falls back to a deterministic match (the longest matching trigger wins) so commands still work — just without the smarter intent check and payload extraction.
The plugin reuses your configured model and keys through the SDK’s api.llm capability, so there’s no separate provider or key to set up. Signed-in Freestyle Cloud users get this too — the agent runs on Freestyle Cloud’s managed LLM, with no local model required.

​
Action types

Script actions run arbitrary shell commands with your dictation as input. Only use commands you trust, and be mindful that spoken text is substituted into the command string.

​
UI page

The plugin adds a Voice Commands page to the app sidebar (shown once the plugin is enabled). From there you can:
  • Add a command with a name, one or more trigger phrases, a description, and an action.
  • Edit or delete existing commands, and enable/disable them without deleting.
The description you give each command matters: it’s used as the tool description the agent reads to decide when to run the command and what payload to extract. A clear description like “Posts the dictated message to the team Slack channel” works better than a terse one. Every change is saved through the window.freestyle host bridge to the plugin’s own CRUD endpoint and persisted with plugin storage, so it survives restarts and takes effect on the next dictation.

​
Build from source

This runs pkgroll --minify (server-side plugin code, bundling the AI SDK) followed by vite build (UI page).

​
Source

plugins/voice-commands/ in the monorepo.