CozyLabs gives six AI agents names, personalities, and a pixel-art world you can watch in real time. Delegate in plain English — they handle everything else.
works with
LM Studio
Ollama
Claude Code
OpenClaw
Eight capabilities, all confirmed in the codebase. No marketing math.
A pixel-art isometric office built on Phaser 3. Each agent has a desk. Speech bubbles show what they're working on — live, always.
Sage, Pixel, Byte, Luna, Pip, Nova. Each with a role, a personality, and a model you configure. Not interchangeable slots — a real team.
A Kanban board your agents actually use. Backlog → In Progress → Review → Done. Priorities, attachments, blocking dependencies, full audit trail.
Talk to your manager. DM an agent. Watch them chat with each other in the project channel. @mentions, #tickets, /file autocomplete.
Outputs as a portfolio, not a chat log. Filter by type. Review with agents. Consolidate into a branch. Push to git. One button.
Nightly code reviews. Weekly research. Daily reports. Set it once, cron-based, per-agent, with auto-run when you don't need approval.
Pose a question. Get proposals with design specs. Let agents vote. Promote the winner directly into a Kanban ticket. Brainstorming, made actionable.
Claude, Codex, or any OpenAI-compatible endpoint. Different model per agent. Sage on Claude, Byte on Ollama — a settings change, not a code change.
They already know what to do. They just need a goal.
Calm, structured, quietly competent. Takes your messy idea and turns it into a crisp Kanban board without asking too many questions.
Creative, slightly chaotic, opinionated about design. Constantly referencing hex codes and debating spacing.
Fast, nocturnal, slightly smug about performance. Does not talk much but ships constantly.
Detail-oriented, thorough, quietly indispensable. Will find the one obscure thread that solves your problem.
Polite, methodical, relentless. Pip is never satisfied and that's exactly the point.
Unshakable, efficient, never panics. Nova has seen worse.
Every number confirmed in the codebase. No marketing math.
Electron Renderer React 19.2 · Phaser 3.90 · Framer Motion 12 · Tailwind v4 ↕ contextBridge IPC (typed · sandbox · contextIsolation · main-frame guard) Electron Main Node.js · native menus · system tray · per-profile gateway probe ↕ HTTP/SSE (loopback only · ports 8642–8670 · one per profile) Hermes Agent Nous Research runtime · OpenAI-compat API · ~/.hermes/profiles/
Don't want to manage hermes processes by hand? Toggle Run via Docker in Settings → Advanced and CozyLabs writes a per-team hermes-compose.yml + runs docker compose up -d. One nousresearch/hermes-agent:latest container per profile, each bound to its own loopback port. Hermes does the heavy lifting; CozyLabs is the wrapper.
One wants to stop babysitting terminals. The other wants to stop building orchestration from scratch.
Building with Claude, Codex, or local models. Outgrown the terminal. Wants multiple agents in parallel without building orchestration from scratch.
Wants to delegate complex work to AI. Knows what to build. Has zero interest in prompts, terminals, or infrastructure.
Strong positioning means knowing what you're not. CozyLabs has a specific point of view.
Another UI over a single language model. The chat exists to communicate with your team, not replace it.
Your Hermes profiles live on your machine. Your conversations live in your filesystem. There is no CozyLabs cloud.
You don't write prompts. You tell Sage what to build. Personas and tool policies are pre-built.
The agents have names. They have desks. Their personalities were written intentionally into the codebase.
Phaser 3 renders your team as pixel-art sprites at named desks. No other AI orchestration tool does this.
Sage, Pixel, Byte, Luna, Pip, Nova have roles, tool sets, and personalities that survive across sessions.
Powered by Nous Research's Hermes Agent runtime — the same isolated profiles, durable memory, and OpenAI-compatible gateway you'd get from the CLI, wrapped in a Mac app. One download, no Docker required, no account.
Outputs structured into projects, filterable by type, reviewable, and pushable to git from the app.
Sage on Claude, Byte on Ollama. Different models for different jobs — a settings change, not code.
Six agents. One world. Zero terminals.
Coming Soon — macOS Release Imminent