doctormizio777777/aetheros

AetherOS is an experimental autonomous agent built for a hackathon that simulates running a small consulting business on its own. Instead of a person deciding when to invoice a client or pay a vendor, an AI model, NVIDIA's Nemotron 550B, plans and carries out those actions itself. According to the README, the agent can send invoices to clients through Stripe checkout, set up cloud infrastructure, and pay vendors using either Stripe or a stablecoin payment rail. It is also designed to recognize and resist social engineering attempts aimed at tricking it into making unwanted payments, and to pause and ask a human for approval when a transaction is high value. The safety mechanism behind all of this is a separate companion project called Authari, described as a deterministic policy gate. The idea is that AetherOS decides what it wants to spend money on, but Authari is the one that actually decides whether that spend is allowed to go through, based on fixed rules rather than another AI judgment call. On the technical side, AetherOS itself is built with Python and FastAPI for the backend logic, using the LangChain framework to work with the language model, plus a React and TypeScript frontend. It talks to Authari, which is a separate Fastify and TypeScript service, through a set of tools following the MCP protocol. Running the whole thing locally means cloning both repositories, starting Authari first to get an API key, then starting AetherOS with Docker Compose once that key is in place. The README states this was built solo for a hackathon focused on AI agents running real businesses, sponsored by Nous Research, NVIDIA, and Stripe, and is released under the MIT license.