AgentGPT

What is AgentGPT?

AgentGPT is a browser-based platform for building and deploying autonomous AI agents without writing code. You type a name and a goal, press deploy, and the agent breaks the goal into smaller tasks, runs them one after another and reports what it found. The hosted app at agentgpt.reworkd.ai offers ready-made templates such as ResearchGPT, TravelGPT and StudyGPT, so first-time users can see the loop in action within seconds.

AgentGPT was built by Reworkd, a startup backed by investors including Paul Graham and Nat Friedman that later moved into AI web data extraction, and its code is published on GitHub under the GPL-3.0 licence. The repository was archived on 28 January 2026 and is now read-only, with more than 36,000 stars and 9,000 forks. When we checked in October 2026, the hosted site still loaded, carried a Beta label and showed a counter of five runs, but reviewers report that it no longer receives updates, security patches or support. Reworkd’s own documentation now points AgentGPT visitors straight to the GitHub page.

Key Features

• Goal-to-task agent loop: AgentGPT turns one plain-English goal into a list of subtasks and executes them in sequence. Each finished task can add new ones, so the agent keeps iterating until it reaches the goal or hits its loop limit.
• Ready-made templates: AgentGPT ships with starter agents including ResearchGPT, TravelGPT and StudyGPT on the hosted homepage. These presets show the expected goal format and give new users a working example before writing their own.
• Named, custom agents: AgentGPT lets you give each agent its own name and goal before deployment. That makes it easy to run separate agents for research, trip planning or study notes and tell their outputs apart.
• Pause and review: AgentGPT includes a pause option so you can inspect the task list before the agent continues, according to HowAIWorks. Stopping early helps you catch a run that has drifted off course before it burns more API calls.
• Optional web search and images: AgentGPT’s self-hosted setup accepts optional Serper keys for web search and Replicate keys for image generation, according to the GitHub README. Without them, the agent relies only on the language model’s own knowledge.
• Self-hostable open-source stack: AgentGPT publishes its full code under GPL-3.0, built on Next.js and TypeScript at the front end with a FastAPI Python back end and MySQL. Developers can run it locally with Docker and their own OpenAI key.

Pros

• No coding needed: AgentGPT runs entirely in the browser, so non-developers can launch an agent by typing a name and a goal. There is nothing to install for the hosted version, which made it a popular first look at autonomous agents.
• Free to try: AgentGPT’s hosted app shows five free runs, and directories describe the free tier as five demo agents per day. That is enough to test a research or planning goal before deciding whether the approach suits you.
• Open-source code: AgentGPT’s GPL-3.0 code base lets you read, modify and self-host the whole agent. With more than 36,000 GitHub stars, it remains a widely studied reference for how browser-based agent loops are built.
• Clear task visibility: AgentGPT shows each generated task and its result in the chat window as the run progresses. Watching the plan unfold makes it a useful teaching tool for students and teams learning how agent loops behave.

Cons

• No longer maintained: AgentGPT’s GitHub repository was archived on 28 January 2026 and is now read-only. Reviewers report no updates, security patches or support for the hosted site, so bugs are unlikely to be fixed.
• Unreliable on long tasks: AgentGPT agents often drift off track or loop without finishing, and reviewers describe the results as demo quality. Prospeo also reports that the loops can underperform simply asking GPT-4 directly.
• Tied to OpenAI models: AgentGPT is built around the OpenAI API through LangChain, and Voiceflow notes it is locked to OpenAI models. Teams wanting Claude, Gemini or local models would need to modify the code themselves.
• Tight free limits: AgentGPT’s free hosted tier allows only five runs, and each agent has a capped number of loops. Self-hosting removes those caps but shifts every API charge to your own OpenAI account.
• No production tooling: AgentGPT lacks a public API, monitoring and evaluation tools, according to HowAIWorks and Voiceflow. That makes it unsuitable for business workflows that need logging, testing or integration with other systems.

Who is Using AgentGPT?

• Researchers: Gathering and summarising background material on a topic.
• Students: Building study plans and revision notes with StudyGPT.
• Travellers: Drafting trip itineraries with the TravelGPT template.
• Developers: Studying or forking an open-source agent architecture.

Pricing

AgentGPT is free to try on the hosted site and free to self-host, with a Pro plan reported at $40 per month (checked October 2026; paid tiers from third-party reviews).

The official site no longer shows a plan page, so we could not confirm whether the Pro upgrade still accepts new payments after the January 2026 archive.

Because the project is unmaintained, avoid committing to a paid plan until you have checked billing and support directly with Reworkd.

What Makes AgentGPT Unique?

An agent you can watch think is what set AgentGPT apart: it shows every self-generated task in a chat-style feed, inside a plain browser tab, with no install. Listed by There’s An AI For That as launching in April 2023, it put autonomous goal-chasing agents in front of non-programmers, and its open-source code became a common starting point for developers building their own. Today its value is mainly educational, since the archived project receives no new models or fixes.

Compatibilities and Integrations

• Platforms: Any modern web browser for the hosted beta; self-hosting needs Docker and Node.js 18 or later.
• AI models: OpenAI models through the OpenAI API and LangChain; you can bring your own key.
• Integrations: Optional Serper web search and Replicate image generation in self-hosted setups.
• Tech stack: Next.js, TypeScript and Tailwind front end, FastAPI back end, MySQL with Prisma.

AgentGPT Tutorials

A first AgentGPT run takes about a minute on the hosted site.

1. Open agentgpt.reworkd.ai in your browser and sign in if prompted.
2. Pick a template such as ResearchGPT, or type your own agent name.
3. Write a specific, measurable goal in the goal box.
4. Press deploy and watch the agent list and complete its tasks.
5. Pause the run if it drifts, then copy any useful results.

How We Rated It

Summary

AgentGPT suits students, hobbyists and developers who want to see how an autonomous agent plans and runs tasks. Its main limitation is that Reworkd archived the project in January 2026, so the hosted beta receives no fixes, new models or support. We would treat it as a learning tool rather than a production assistant. If you try it, write narrow goals with a clear end point, because broad goals tend to send the agent into long, costly loops.