autonomous-loop-studio/Autonomous

### A self-hosted AI agent console in plain PHP.
No Docker. No Node. No build step. No account.
Point it at any model — hosted, or running on your own machine — hand the agents
real tools, and watch the whole loop on a live map. It is one folder of PHP
and static files: drop it on the host you already pay for, double-click, and a
team of agents is running.
🚀 Quick start ↓ · ✨ Features ↓ · 🔒 How the guards work ↓ · 🧩 Architecture ↓
## Run it in five minutes
Windows — double-click start.bat. First run installs Apache + PHP (no MySQL), creates your
access token, starts the server and opens the app. Later runs just start Apache.
Linux / macOS — ./start.sh — or ./start.sh 8090 for another port.
Paste the token it prints into Config → Access token, add one provider key, and you have an
agent. Nothing ships with a key, and a fresh copy refuses to serve (503) until a token
exists — you cannot forget to set one.
Then add a provider — OpenAI, Anthropic, Google, OpenRouter, Groq, DeepSeek, xAI, Mistral,
Cerebras, Together, Perplexity, NVIDIA NIM, or Ollama / LM Studio / vLLM on your own LAN.
## What it is
🤖 Any model, any provider. It detects each provider's API shape and its tool dialect —
JSON, Claude XML, raw tool-call tags — and remembers which one actually worked for that model.
🛠️ Real tools, not chat. Read and write files inside a jail, run allowlisted commands
(argv only — no shell), drive a real Chrome/Edge over DevTools, commit with git, call MCP servers.
🕸️ You can watch it. A live 3D map of agents, providers and traffic; every hop also written
out in words in a run trace; routing charts you draw instead of describing.
🔐 Safe by default. exec, run_code, the browser and the MCP server are off until you
switch them on. Every write asks first. No accounts, no cloud, no telemetry.
## What it is not
Not a hosted service, not a wrapper around one vendor, not a SaaS with an account page. There is
nothing to sign up for, and the only outbound calls are the provider endpoints you configure
yourself.
## A fleet for $0
Free tiers and local models are first-class here, not an afterthought — and paced so one exhausted
key does not stop the fleet.
⏱️ Paced to each provider's real ceiling — the pacer knows the host's limit (NVIDIA NIM
40/min, Groq 30, OpenRouter 20, a default for anything unknown), and 0 = AUTO uses it. Global,
per-provider or per-agent, and an agent can lock its own.
🔑 A 429 is not a stall — a rate-limited key rests for 60 seconds and the next call reaches
for another; add extra keys and the API pool rotates them after N requests.
♾️ Loops that keep going — Always-On re-prompts a standing goal on a cadence, with max
iterations and max minutes caps, per-agent hold/release and one master kill switch.
Free tiers have daily caps, so pacing smooths bursts rather than making them infinite, and "always
on" means the host stays awake — the tab-hidden worker, or a machine that does not sleep. Against
Ollama, LM Studio or vLLM on your own box there is nothing to pace at all.
## A look at it
Glass desktop and Classic (the default) · a routing chart · Brain Lab · the built-in browser.
## Everything else, below
Quick start · Is it safe? · What's inside ·