Authority AI Editorial
AutoGPT: the repo that started the autonomous agent wave
AutoGPT is the project that put "autonomous agent" into everyday use among people who run AI tools. It launched in March 2023 as a single script that gave a large language model a goal, let it write its own next step, run that step, look at the result, and decide what to do next without a human approving each move. That loop, plan, act, observe, repeat, is the origin point for most of the agent frameworks that followed. The repository has since grown into a full platform, and the reader this article is for is someone who runs coding agents day to day and wants to know what AutoGPT actually is now, not just what it was at launch.
What it is today
The GitHub organization Significant-Gravitas still owns the AutoGPT repository, and the original single-agent script now lives under classic/ in that same repo. The active project is AutoGPT Platform: a visual agent builder, a marketplace of prebuilt agents, a run dashboard, and a chat-based agent creator called AutoPilot. You describe an outcome in plain language or wire the steps yourself in a drag-and-drop canvas, then run the result on demand, on a schedule, or from a trigger such as an incoming email or webhook. The platform connects to more than 45 external services, including Gmail, Slack, GitHub, Notion, and Stripe, so an agent can read a calendar, post to a channel, or update a spreadsheet as one of its steps.
How you use it
There are two paths into AutoGPT. The hosted Platform at platform.agpt.co is a paid, managed service: sign up, connect the accounts you want the agent to touch, and run agents without operating any infrastructure yourself. The self-hosted path is free of license fees. You run Docker on Linux or macOS, either through the published single-container installer or by cloning the repository and running the setup scripts, and you supply your own API keys for whichever model provider you use. Self-hosting can run fully offline against a local model through Ollama with a --with-ollama flag on the installer. Either way you end up at the same builder and dashboard once the service is running; the difference is who operates the infrastructure and pays the model bill.
What to know before adopting it
Licensing is split by directory. The autogpt_platform/ code, which is the current product, ships under the Polyform Shield license: free for personal and internal business use, but you cannot resell it as a competing hosted service. Everything under classic/, including the original agent script and the Forge framework for building your own agent on top of it, is MIT licensed. The stack is Python on the backend with a web frontend, self-hosting needs Docker plus roughly 25 GB of free disk space and 8 GB of RAM, and the hosted tier bills by agent usage on top of a paid plan. If your interest is specifically the 2023-era single-file autonomous loop rather than the current builder and marketplace, go straight to classic/ and treat the rest of the repository as a different, newer product built by the same team.
Where it sits next to similar tools
AutoGPT was the first widely used implementation of the goal-driven agent loop, and it is the reason terms like "autonomous agent" and "agent loop" are now common shorthand. The tools that came after it split into two camps: graph and workflow orchestrators such as LangGraph that give a developer explicit control over each transition, and role-based multi-agent frameworks such as CrewAI that assign named agents to parts of a task. AutoGPT's current platform sits closer to a no-code or low-code automation tool, comparable to something like n8n with agent steps built in, than to a library you import into your own Python codebase. If you are choosing a framework to embed inside code you write and maintain yourself, LangGraph or CrewAI is the more direct comparison; if you want a hosted or self-hosted product with a builder UI and a marketplace, AutoGPT's current form is closer to what you are looking for. People assembling their own agent tooling from scratch, rather than adopting a platform, often work from a fixed reference set of instructions instead: the AI Ultimate Skill Collection is 3,029 ready-made skills in 12 packs that work with Claude Code and other agents that use skills.