Murmur

Inspiration

Voice assistants tell you how to do things. We wanted one that actually does them.

What it does

Press a shortcut, speak a command, and Murmur browses real websites, fills forms, and answers live questions on your behalf, then reads back what it did. It connects to 19 services like Gmail, Notion, and GitHub.

How we built it

Electron desktop app with a React frontend and Node.js backend. ElevenLabs handles voice input and output, Gemini figures out what you're asking for, and Browser Use drives a real cloud browser to carry out the action. Auth and storage run on Supabase.

Client (mic + UI)
Voice Layer (ElevenLabs STT/TTS via WebSocket)
Agent Orchestrator (LLM + tool calling)
Tool Layer (Browser Use agents + APIs)
Agent output cleaning (to clean output for user)
State + Memory (session context, preferences)
Response → Voice (ElevenLabs TTS stream)

Challenges we ran into

macOS window management in Electron was painful with a lot of focus and timing bugs. ElevenLabs voice detection doesn't always fire for short phrases, so we built a fallback that manually encodes a WAV file and retries via REST. The browser automation API can return any data type, which required a lot of defensive handling throughout.

Accomplishments that we're proud of

Three independent safety checks so the agent can never take a destructive action. The full voice lifecycle is cancellable at any point. We also run intent classification and context resolution in parallel to reduce latency. All shipped in under 20 hours.

What we learned

Real-time voice is mostly a timing problem. Every layer runs on its own schedule and getting them to work together reliably takes more effort than expected.

What's next for Murmur

We want to build out user onboarding so Murmur can learn your preferences, common workflows, and the services you use most. The more it knows about you, the more useful it becomes over time. We also want to add voice-driven email and calendar drafting, and expand from 19 integrations to the full catalog of 978 services that Browser Use supports.