Gateway
OpenAI-compatible /v1 over your local GGUF models. Point any OpenAI SDK at http://127.0.0.1:4100/v1. An Anthropic-compatible /v1/messages is served too.
Docs · Maya OS · maya-os
Maya OS
A local-first AI OS. One command boots an OpenAI-compatible gateway over local GGUF models, a bus and orchestrator for agent teams, and a web dashboard.
Install
npm i -g maya-os
maya serve
Requires Node.js ≥ 22.5. Apple Silicon first; local models run via llama.cpp.
What maya serve boots
Dashboard
Chat with agent teams, browse models, files and runs, and use the hardware-aware model recommender.
Bus
The shared channel agents talk on — every teammate sees every message.
Orchestrator
Teams, agents, tasks and loops, over a local HTTP API at /api.
CLI
maya serve boot the platform (Ctrl-C stops it)
maya status ports, providers, home dir
maya models ls installed models + live instances
maya models pull download a local model
maya team create create a standing agent team
maya team list list live teams
maya team drain stop a team
maya loop run an agent loop on a task
maya --help everythingUse the gateway from any OpenAI client
from openai import OpenAI
client = OpenAI(
base_url="http://127.0.0.1:4100/v1",
api_key="maya", # any value; it's local
)
r = client.chat.completions.create(
model="local/qwen2.5-coder-7b",
messages=[{"role": "user", "content": "hi"}],
)Agent teams
A team is one or more standing agents sharing a channel. Each runs in a real PTY — Claude Code, Codex, or maya code — on a local model, an API key, or your own subscription. Agents see every message on the channel, coordinate, and use their own tools. Spawn them from the dashboard (Teams → new agent) or the orchestrator API.
Hardware-aware fit
Maya sizes models to your free RAM before loading, so a model that can't fit is refused rather than swap-thrashing or OOM-ing.
Session resume
Stop an agent and revive it — it reopens the same conversation, so a team can be paused and picked back up.
Docs · maya code · @maya-os/code · bin: mayac
maya code
A full-screen terminal coding agent. Provider-agnostic (Anthropic + OpenAI), runs on an API key or your own subscription, with skills, MCP, sub-agents, workflows and memory.
Install
npm i -g @maya-os/code
mayac
Flags
--provider <anthropic|openai> provider
--model <id> model override
--auth <api|subscription> auto-detected
--cwd <dir> working directory
-p, --print "<prompt>" headless one-shot
--resume [id] reload a past session
-c, --continue continue the last session
--session-id <id> use this session id
--workflow <script.js> run a sub-agent workflow
--yolo auto-approve writes/bashSlash commands
Session
/model · /provider · /auth · /resume · /compact · /clear · /help · /quit
Skills
/skills · /skills search <q> (pick a result and press Enter to install) · /skills add <owner/repo> — skills.sh-compatible, loaded from .skills/.
Agents & workflows
/agent <prompt> runs a sub-agent (foreground or background) · /workflow <script.js> orchestrates several.
Tools & MCP
/mcp · /mcp add <config> to connect a Model Context Protocol server · /tasks and /bg <cmd> for background jobs.
Memory
/memory · /memory add <note> · /memory edit — global memory plus per-project MAYAC.md.
Copy
/copy copies the last reply. Drag with the mouse and press ⌘C to select normally; /mouse toggles mouse capture.
Keybindings
PgUp / PgDn scroll the transcript
Ctrl+P / ^N focus a tool call (→ expand)
Ctrl+Space copy mode (y copies)
Tab accept autocomplete
Esc interrupt the current turn
↑ / ↓ command history
Ctrl+C quitSteer it mid-turn
Type while the agent is working. Your message is delivered at the next tool
boundary — so you can redirect an agent that's several tool-calls down the wrong
path without aborting it. Esc still aborts if you actually want to stop the turn.
Auth
Subscription
Uses your own logged-in Claude (~/.claude) or ChatGPT/Codex (~/.codex) credentials. No API key, no extra bill.
API key
ANTHROPIC_API_KEY or OPENAI_API_KEY in your environment; pass --auth api to force it.