Speak rough.
Ship sharp.

Hold a hotkey. Talk for ten seconds. Prompty rewrites your transcript into a structured engineering prompt — file paths, scope, acceptance criteria, edge cases — and drops it into Claude Code, Cursor, or Continue.

Install in 30 seconds Read the code

MIT-licensed. macOS-first (Linux + Windows in v0.4). Privacy-posture-by-default — no audio retained, no transcript stored, no telemetry on by default.

What it does, in three bullets

Voice → structured prompt. Whisper transcribes your rambling description; Claude (or your own LLM) rewrites it into a tight engineering ask.
Context-aware. Reads your active file path, git HEAD, and the last few messages in your coding session — your prompt mentions the right files without you spelling them out.
Lives where you code. An MCP server that Claude Code / Cursor / Continue can call as a tool, plus a daemon that types the rewrite at your cursor anywhere on macOS.

Privacy posture

We took the privacy bit seriously because voice input has obvious failure modes if we hadn't:

Audio never persists. Voice lives in process memory only — opt-in disk buffer is AES-256-GCM encrypted with a per-session ephemeral key, auto-deleted within 60s of transcription.
Transcripts aren't logged unless you set PROMPTY_TRANSCRIPT_LOG=true.
BYO-key by default. The OSS path uses your own Anthropic / OpenAI keys (or skips them entirely via the host driver if you've got a Claude subscription). The optional Cloud tier proxies through us with a managed key — we see metadata only, never your transcripts or rewrites.
Egress allowlist. The daemon fails closed on network calls to any host you haven't configured. No background telemetry.
Open source. Every line of code is on GitHub. Read the threat model to see what we did + didn't promise.

Install

Three install paths. Pick whichever you've already got tooling for:

npm (any platform with Node 20+)

npx prompty-mcp init

Homebrew (macOS)

brew install prompty-dev/tap/prompty
prompty init

From source (developers)

git clone https://github.com/maksodf/prompty
cd prompty && pnpm install && pnpm build
node packages/mcp-server/dist/prompty.bundle.js init

After install, run prompty doctor to verify the setup. The default host driver works with Claude Pro / Max / Team / Cursor / Continue subscriptions — no API key required. Switch drivers via prompty config set llm.driver anthropic (or openai, deepseek, whisper_local, etc.).

Cloud tier (invited beta)

Want managed Anthropic + OpenAI keys behind one bill? Email hello@prompty.khalo.org for a tester key, then prompty cloud login --key pck_…. The dashboard shows usage + quota. Billing is intentionally not enabled yet — we're validating functionality with invited testers first.

FAQ

How is this different from Wispr Flow / SuperWhisper?

Wispr / SuperWhisper transcribe + grammar-fix. Prompty does that, but ALSO reads your coding agent's current context (active file, git HEAD, recent messages) and rewrites the transcript into a structured prompt with file paths + acceptance criteria. Same hotkey ergonomics; different output category.

What's "MCP"?

Model Context Protocol — a Claude / Cursor / Continue standard for tools the host agent can call. Prompty ships as an MCP tool named rewrite; ask your coding agent to "use prompty to rewrite this voice transcript" and it just works.

Do I need an API key?

No, if you've got a Claude Pro / Max / Team / Enterprise subscription, Cursor's plan, or Continue's bundled LLM — the default host driver returns a "rewrite-this" prompt that your coding agent's own LLM handles. Zero extra billing, no key to manage. You only need a separate Anthropic / OpenAI key if you choose those drivers explicitly.

Does it work on Linux / Windows?

The MCP server + most of the daemon path work on Linux today (audio capture is macOS-only until v0.4). Windows is mostly there but the native hotkey path needs additional polish — track it on the issue tracker.

Where's the demo GIF?

Coming soon — the maintainer needs to sit down with screen recording. In the meantime, see docs/demo.md in the repo for an ASCII walkthrough.

Speak rough.Ship sharp.