Wispr Flow
A dictation app: hold a key, talk, and cleaned-up text lands in whatever app has the cursor, terminals and agents included. Cole's verdict: good, but thinking of making one.
f. 8rThe DeskCapitula · contents
Wispr Flow is a dictation app. You hold a key, speak, and polished text appears wherever the cursor is: an email, a browser, a terminal, a prompt to an agent. It is on every one of Cole’s Macs, installed from the Brewfile, and Cole’s verdict on it is short.
What it does
Wispr Flow turns speech into text “in any app”, across Mac, Windows, iOS and Android, in more than 100 languages. Homebrew’s listing sums up the difference from plain dictation: “Voice-to-text dictation with AI-powered auto-editing”. The output is cleaned up, not a raw transcript of every “um”. It also learns the names and jargon you use, and keeps a dictionary and snippets, shortcuts that expand into longer text you say often.
The free plan includes 2,000 words a week on the desktop. Pro removes the limit ($15 per user per month ($12 billed annually) (as of , wisprflow.ai)) and adds shared dictionaries and snippets for teams.
Why dictation fits agent work
Talking to coding agents is mostly writing: what you want, what you already tried, what must not change, what done looks like. The better that brief, the better the agent’s work, and most people talk much faster than they type. A long, specific prompt that would take five minutes to type takes one to say.
Claude Code has voice dictation of its own: run /voice, then hold the space bar to talk. Its docs list the catches:
it needs a Claude.ai sign-in, so it is unavailable when Claude Code runs on an API key or a cloud provider like Amazon
Bedrock, and it needs a local microphone, so it does not work in SSH sessions. In Cole’s setup many agents run on
servers, reached over SSH in tmux, and some run through Bedrock with the yuma launcher. A
system-wide dictation app does not care: it types into the focused window on the Mac, and the terminal sends those
keystrokes wherever it is connected.
Tips
- Say the structure. “First, … Second, … Don’t touch the tests.” The cleanup keeps it.
- Read before you send. Dictation is fast, and fixing a misheard file name before an agent acts on it is faster than undoing what it did.
- Teach it your words. Add project names, tool names and teammates’ names to the dictionary early; jargon is what dictation gets wrong most.
- On the phone, Termius has voice typing built into its keyboard panel, for short commands.
Making one
Cole is thinking of building a replacement. The ingredients are well understood: a global hotkey, a speech-to-text model, a language-model pass that cleans up the transcript, and a way to insert text into whichever app is focused, which on a Mac means accessibility permissions. It is the kind of small menu bar app the Mac Apps and Menu Bar Apps page covers.