Warp Ai

Inspiration

We talk more than we type - Slack replies, emails, quick edits. MacOS dictation stops the second you want the text improved, so you end up cleaning it up in another app. We wanted dictation and AI in the same gesture, at the cursor.

What it does

Hold a key, talk, release. Text lands wherever your cursor is - live dictation, in-place rewrite/translate/explain on any selection, or voice chat with an ephemeral reply overlay. Todoist, Slack, and Bear ship today; Notion, Linear, and Gmail are in beta.

How we built it

A SwiftPM macOS app with a Hono/Prisma/Postgres backend and an Astro site. Streaming transcription via Soniox, AI via Groq with server-held keys. A five-state machine (idle arming listening injecting cooldown) keeps the voice flow valid, and the UI is just an audio-reactive edge glow on the focused window.

Challenges we ran into

Code-signing. Notarizing isn't enough - you have to staple the ticket to both the .app and the .DMG or quarantined downloads still warn users. And text injection into arbitrary apps: every field behaves differently, so that took more iteration than transcription.

Accomplishments that we're proud of

The edge glow. No floating orb or pill - the focused window itself signals it's listening and disappears when you're done. Plus a credential model with nothing sensitive baked into the binary.

What we learned

The hard part isn't the transcription engine. It's making the gesture feel instant and earning the trust of someone letting a tool type for them. Latency, not features, is where that trust lives.

What's next for Warp Ai

Ship the beta integrations to everyone and keep squeezing the gap between releasing the key and seeing the text.

Built With

• claude
• cursor
• groq
• railway
• soniox
• zcode