From a 2023 experiment to a platform
AutoGPT started in March 2023 as a standalone script that chained GPT calls into an "autonomous" loop, and it became one of the most-starred projects on GitHub almost overnight. What sits behind those 185,000 stars today is a different thing: a platform for building, deploying, and managing continuous AI agents that automate workflows, with a low-code builder, a server that hosts the agents, and a marketplace of pre-built ones. The original standalone agent still lives in the repository under classic/, but the active work is the platform.
How you actually run it
There are two paths. You can self-host the AutoGPT platform for free, or join the waitlist for a cloud-hosted beta that is still in closed release. Self-hosting is a real setup, not a one-file download. The README lists Docker Engine and Docker Compose, Git, Node.js and npm, and a code editor, on a machine with at least four CPU cores and 8 to 16GB of RAM. There is a one-line install script for macOS, Linux, and Windows (through WSL2) that wires up dependencies and Docker for a local instance. If that stack sounds like more than you want to operate, the project itself points you to the cloud waitlist instead.
Agents are built from blocks
The part worth understanding is the building model. In the frontend's Agent Builder, you assemble an agent by connecting blocks, where each block performs a single action. Workflow management, deployment controls from testing to production, and monitoring all live in that same interface, alongside a library of ready-to-use agents for people who would rather not build from scratch. The server is where those agents run: once deployed, an agent can be triggered by an external source and operate continuously. The README's own examples are concrete. One agent reads trending Reddit topics and produces a short-form video; another transcribes a new YouTube upload, picks the strongest quotes, and drafts a social post. That is the shape of what the platform targets: repeatable, triggered automations rather than a single open-ended "do everything" loop.
The license line most people skip
There is a licensing detail that matters if you plan to build on this commercially. Everything inside the autogpt_platform folder, which is the new platform, is under the Polyform Shield License. That is a source-available license, not an OSI-approved open-source one. The rest of the repository, including the classic standalone Agent, Forge, and the agbenchmark tool, stays under MIT. So "AutoGPT is MIT" is only half true now: the platform you would actually deploy carries Shield terms. Read them before you assume you can fork the platform into a product.
Where it fits, and the caveats
For a small team exploring agent automation, AutoGPT is a reasonable way to see the block-based approach in practice without writing a framework yourself, and the ready-made agents and the agbenchmark give you something to measure against. The honest caveats are the ones any continuous-agent system carries. Self-hosting is ongoing infrastructure work, not a weekend toy. Agents that run on every trigger and call a model each time have a real and variable cost, so monitoring and budget limits are not optional. And the smoothest path, the managed cloud, is still behind a waitlist. Treat AutoGPT as a way to prototype and understand agent workflows now, confirm the license fits your plans, and price the operational overhead before you wire one of these into anything that matters.
Top comments (0)