OpenYabby
Voice-driven multi-agent execution. For macOS.
Say "Yabby". Speak a task, approve a plan, and let local agents execute real project work through CLI runners, connectors, review, and QA. MIT-licensed and built in public.
Your Mac, fully
autonomous.
Yabby doesn't just chat. It executes. Terminal, browser, files, apps. Everything.
Voice-First. Always.
Bidirectional audio via WebRTC. Wake word “Yabby” with Silero VAD and optional speaker verification. Natural conversation in any language. No wake button, no typing. Speak and Yabby acts instantly.
Multi-Agent Teams
Lead agents create managers and sub-agents. Auto-orchestration with cascading task queues: parallel within a phase, sequential across phases. Your project team runs itself.
Full System Access
Terminal, AppleScript, Playwright, file system, browser DOM. Agents never say “I can't”. They find a way and execute it.
Persistent Memory
Mem0 extracts facts from conversation every 6 turns. Qdrant vector search + SQLite. Yabby learns who you are, your preferences, your context. Persistent across every session.
Multi-Runner Support
Claude Code (default), OpenAI Codex, Aider, Goose, Cline, or Continue CLI. Each task spawns a CLI process with its own session, system prompt, and full context.
Real-Time Everything
SSE + WebSocket live feed of tool calls, agent progress, and task completions. Redis pub/sub for inter-agent messaging. Watch your AI workforce in real time.
The app is almost
ready.
A mobile companion for Android & iOS is coming later. Drop your email and we’ll let you know the day it’s ready to test.
Voice to execution.
In seconds.
You speak
"Yabby, build me a dark mode portfolio with a blog section"
Natural voice via WebRTC. Wake word "Yabby" activates. Client-side VAD + noise filter ignores background sounds. Optional speaker verification ensures only your voice triggers it.
Yabby assigns a lead
Three tool calls in one response. Yabby creates the project, assigns a single lead agent, and sends the full brief. The lead builds the team autonomously.
Lead builds the team & executes
The lead submits a plan for your approval, recruits agents, and sequences work: design first, then frontend + backend in parallel. Each agent is a CLI process with full Mac access.
Auto-review, QA & delivery
When tasks complete, the orchestrator auto-triggers the lead for review. The lead inspects the sandbox, creates a QA agent, and only marks the project complete when QA passes. You get a voice notification. Zero micromanagement.
Three-tier orchestration
that runs itself.
Agents communicate via Redis pub/sub. Completions cascade up automatically.
Project Director
Full API access. 5-phase workflow: discovery, planning, execution, review, QA. Submits plans for user approval. Reports to you via voice.
CTO
Coordinates dev sub-agents. Auto-triggered review on task completion. 5s debounce.
CDO
Manages design team. Reviews visual output, ensures consistency.
Frontend
Backend
UI/UX
QA
Connected to
everything.
Channels, connectors, and MCP servers. Yabby talks to the tools you already use.
5 Channels
Discord, Slack, Telegram, WhatsApp, Signal. Same tools as voice. Slash commands, group mentions, per-thread agent binding.
37 Connectors
GitHub, Jira, Linear, Notion, Google Drive, Stripe, and more. Built-in or MCP-backed. Project-scoped with encrypted credentials.
MCP Servers
Spawn and manage Model Context Protocol servers. Tool schemas auto-bridged to OpenAI format. Works with any MCP-compatible tool.
Scheduled Tasks
Cron, interval, or manual triggers. The built-in scheduler ticks every 30s, retries on failure, and recovers orphaned runs on startup.
What makes it
actually work.
7 design decisions that define how Yabby operates.
Voice + action, same response
Speaks and calls tools simultaneously. No "let me do that for you". It's already done.
Full Mac access per agent
Bash, AppleScript, GUI control. Write files, run servers, take screenshots, open apps.
Agents recruit agents
The lead builds its own team via API calls. No human in the loop for team formation.
Plan review & approval
Leads submit plans for your approval before execution. Approve, revise, or cancel via voice or UI modal.
Auto-triggered reviews
Sub-agent finishes → orchestrator re-spawns the lead to inspect and decide. 5s debounce, 10s delayed re-check.
Mandatory QA
No project ships without a QA pass. Corrections loop until clean. Hallucination detector catches false claims.
Live voice notifications
Agents push milestones to SSE + WebSocket. Yabby relays them as speech. Also forwarded to WhatsApp.
Running in
two steps.
git clone https://github.com/OpenYabby/OpenYabby && cd OpenYabby && npm i && npm run dev
Clone & configure
git clone https://github.com/OpenYabby/OpenYabby.git
cd OpenYabby
npm install
cp .env.example .env
# Add your API keysLaunch
npm run dev
# → http://localhost:3000
# Say "Yabby" to start talkingRequires Node.js, PostgreSQL, Redis, Claude CLI, and an OpenAI API key.
Serious infrastructure.
Named after the Australian crayfish.
A yabby (Cherax destructor) is a small freshwater crayfish native to Australia. Tough, resourceful, and surprisingly effective. They thrive in creeks, dams, and rivers across the continent.
Like its namesake, OpenYabby is compact but powerful. Small claws, big results. It burrows into your system, finds what it needs, and gets the job done. Autonomously, relentlessly, without fuss.
Ready to command
your AI task force?
Open source. Self-hosted. Your machine, your data, your rules.