Maya — a local-first AI OS for developers

One command boots a model gateway, an agent orchestrator, a message bus, and a web dashboard on your Mac. Run local GGUF models with RAM guardrails, drive real Claude Code agents (cloud or local), watch each agent's actual terminal live, and let agents message you proactively — no polling, no schedule hacks.

npm install -g maya-os     # or: npx maya-os serve
maya serve                 # gateway :4100 · dashboard :4200 · bus :4300 · orchestrator :4400

Open http://127.0.0.1:4200 → Chat.

What you get

• Chat with agent teams — create a team (lead + members, each fully configured: model, working directory, tools, auth profile), talk to it in the dashboard, @mention an agent to wake it.
• Real terminals 🖥️ — every standing agent runs interactive Claude Code in a real PTY (not --print). Click the terminal icon on any agent to watch the actual TUI work, live.
• Proactive agents — each agent gets an inbox + outbox file; anything it appends to its outbox posts to the chat as that agent, on its own initiative. curl your agent's server and watch it say hello in your chat.
• Local models with guardrails — pull GGUFs (maya models pull hf:...), served via llama.cpp with RAM-floor monitoring, auto quant-downgrade, hung-process restarts. maya tune <model> benchmarks quant×context for your machine.
• One gateway for every model — OpenAI-compatible /v1/chat/completions + Anthropic-compatible /v1/messages; point any coding agent at ANTHROPIC_BASE_URL=http://127.0.0.1:4100 and it runs on whatever model you route.
• The Loop — a persistent lead plans, spawns workers, verifies against the real files, and reports honestly (it will not claim success it can't verify).
• Teams survive restarts — standing teams re-hydrate on boot; nothing is lost on upgrade.

Requirements

• macOS (Apple Silicon) — other platforms untested
• Node ≥ 22.5
• For local models: brew install llama.cpp
• For agents: the claude CLI (npm i -g @anthropic-ai/claude-code)
• Optional: maya doctor --install-switchyard provisions the Anthropic⇄OpenAI translation sidecar

Commands

| command | what it does | |---|---| | maya serve | boot everything (Ctrl-C to stop; teams survive) | | maya status | what's running | | maya models pull/ls/run | manage local GGUFs | | maya tune <model> | find the fastest quant×context for your RAM | | maya team create/list/drain | standing agent teams | | maya loop "<goal>" | run a verified multi-agent loop | | maya doctor | environment checks |

Security notes

• Everything binds 127.0.0.1. The admin API (/api/*) refuses proxied/tunneled requests — a tunnel can only ever expose the inference surface (/v1/*).
• Auth profiles are read server-side from ~/.cldz/config.json; credentials never reach the browser and never leave your machine.
• Config stores env-var names, never keys.

MIT © himansh