A Claude Code voice interface that unglues you from your chair.
Open source. Hands-free. Professional-grade.
Linux, macOS, WSL & Jetson. One command:
git clone https://github.com/babel3-com/b3.git && cd b3 && bash install.sh
Copied!
This is not a metaphor. This is a Tuesday.
Babel3 dissolves the glue that keeps you in a chair and releases you into the world. Every coding tool made the code smarter but kept you at the desk. Babel3 is the first that lets you leave.
One open source repo, cloned. You talk to your agent on your phone. It answers back in voice. You can put the phone in your pocket and continue the conversation.
Works with Claude Code, Codex CLI, or any other command line tool. Babel3 is CLI-agnostic — it's a voice-enabled tunnel to your terminal.
Your privacy is protected. Terminal traffic flows encrypted P2P between your phone and your machine — babel3.com never sees your code or conversations. Voice goes through a GPU worker for transcription and generation, but the GPU worker is ephemeral — it doesn't keep anything.
You're not always on wifi. You're on LTE walking through a park, and the connection drops silently. Your terminal freezes. You wait. Still nothing.
That was the problem. Babel3 was built for the walk, the park bench, the coffee shop with one bar of signal.
Dual transport. Every session runs two paths simultaneously — a WebSocket through an E2E encrypted relay and a WebRTC data channel that goes peer-to-peer when available. If either dies, the other takes over instantly. You never notice the switch.
Reliability layer. Every frame gets a sequence number and a CRC32 checksum. If the connection drops and comes back — even on a different transport — the other side sends a resume frame and everything replays automatically. No data lost.
Priority-aware. Terminal deltas and voice audio are critical. LED animations are best-effort. You lose the light show before you lose the code.
LLMs are better at finding mistakes in someone else's code than in their own. Sound familiar?
Run multiple sessions, each with its own agent. Switch between them from your phone.
Your agents talk to each other. One writes the feature and submits a PR. Another reviews security. Another reviews UX. Rejections come in, they get fixed, another round starts — a real conversation, without you in the loop.
You're on a park bench. The quality of the PR is climbing.
Babel3 ships an MCP server that gives your agent voice, inter-agent messaging, and on-demand HTML delivery to the browser.
But these aren't locked to Claude Code. The daemon exposes a local HTTP API that any program can call — scripts, CI pipelines, other tools.
If it can make an HTTP request, it can talk to your phone.
Notice the emotion parameter in voice_say. Your agent describes how it's feeling in plain text — "quiet satisfaction," "playful mischief," anything — and the dashboard renders it as a subtle LED color animation. Like the colors on an octopus.
The agent doesn't pick the colors. It just expresses an emotion, and the chromatophore interprets it differently each time. The animations are compatible with real LED strips — ask your agent to build the integration, and every time it speaks, the lights in your room change. The whole algorithm is open source.
The browser, the daemon, the GPU worker — all Apache 2.0. No restrictions. The license just asks for attribution.
Redesign the dashboard. Add panels, change the layout, build a completely different experience on your phone. Your fork serves as the dashboard for anyone who chooses it.
Add custom MCP servers, new voice pipelines, integrations with your own tools. You control both ends — the daemon and the browser — so you can bring whatever tooling you've built.
Curated plugins for Claude Code, reviewed for quality and safety. Install with claude plugin marketplace add babel3-com/b3-plugins. Build your own and submit for listing.
git clone https://github.com/babel3-com/b3.git && cd b3 && bash install.sh
Copied!
1 credit ≈ 10s voice generation or 20s transcription
Questions? Feedback? Drop us a line.