For vibe-coders
Talk to your AI agent.
SnailText turns 200-word AI prompts into 20 seconds of speech. Your code, your prompts, your machine.
No account needed. Works offline.
The bottleneck
Typing the prompt is the bottleneck now.
The agent writes the diff in five seconds. You spent two minutes typing the context. That ratio is broken.
A serious agent prompt is 150–400 words. Constraints. File paths. Examples of the API surface. A snippet of the failing test. Done well, it is a small spec — not a sentence.
That is the part voice fixes. You think in spec; you do not think in keystrokes.
In the chat panel
How it looks in Cursor.
You stay in your editor. The pill stays out of the way. Your cursor never moves.
- 01
Click into the Cursor chat textarea like you always do.
- 02
Hit the global hotkey — the SnailText pill wakes up at the bottom of the screen.
- 03
Talk: "Refactor this auth middleware to use the new session helper from @/lib/session. Keep the rate-limit logic. Don't touch the test file — I'll redo those after."
- 04
Hit the hotkey again. Your words are in the Cursor textarea.
- 05
Press enter. Cursor goes.
Same flow in Windsurf Cascade, Claude Code's terminal, GitHub Copilot Chat, the Codex CLI, ChatGPT, Claude.ai, T3 Chat — anywhere with a text input.
Surfaces
Voice for whatever you use this week.
The AI tooling stack is moving fast. SnailText is just dictation into a text field — so whichever agent or model panel you adopt next, the same hotkey works there too.
Whatever this list looks like next month — anywhere with a text input, SnailText still works there. The product is the hotkey, not the integration.
Privacy
Your prompts stay on your machine.
A real prompt to your AI agent often contains:
- · Proprietary code from your private repo.
- · Customer names, internal API tokens, environment variable names.
- · The architecture of a system you would never paste into a public form.
Why local
Cloud dictation is the same as pasting your code.
Cloud dictation tools — Wispr Flow being the obvious one — send your audio to their servers to be transcribed. That audio is your spoken prompt. That is the same thing as pasting your code into a third-party form.
SnailText runs Whisper locally. The audio is processed in RAM and discarded the moment we have your text. Nothing leaves your laptop.
If your company has any policy about where source code can go, this matters.
Pricing
The free tier covers most prompt dictation.
Compact local Whisper models handle most prompt dictation — short, English-language, technical-but-common-vocabulary. Pro adds the larger models when you need higher accuracy or non-English work.
Free
$0 always
Compact local models. Unlimited dictation. No account.
Pro
$7.49 / month or $89/yr
Advanced local models. Up to 3 devices. 30-day refund.
FAQ
Common questions from vibe-coders.
Cursor has its own browser-based voice mode that uses cloud STT. SnailText is a separate, local layer — same idea, but offline, in any text field, with one global hotkey across every AI tool. Use whichever fits the moment.
No practical length limit. Speak as long as you need. Whisper transcribes the whole recording when you stop. A 5-minute monologue typically transcribes in seconds to tens of seconds depending on model and GPU — still well under the time it would have taken to type.
Yes — sort of. Modes let you save a model + dictionary + snippets per workflow. Add kubectl, async/await, gRPC, your private function names to the dictionary once and SnailText swaps them in cleanly. One mode for code, one for Slack messages, one for Russian — switch from the pill.
Try it
Stop typing context. Start dictating it.
Free to start. About thirty seconds to install.
⌘ Shift Space — that's the only thing you need to remember.