AutoGPT: An Open Platform for Building and Running AI Agents, Now With a Hosted Option

What AutoGPT Actually Solves Now

The original AutoGPT was a CLI experiment that chained LLM calls until a task was done. The current repository is a different product. It is an open-source platform for building, deploying, and running AI agents that carry out complete workflows. The README frames it as: describe an outcome in plain English or shape every step in a visual builder, then run the agent on demand, on a schedule, or from a trigger. That target audience is broad, but the material makes it specific: people who want agents without writing code, and teams that want to automate multi-step tasks like preparing a daily brief or researching accounts before meetings. The repository is no longer just a script. It is a platform with four surfaces: AutoPilot, Agents, Marketplace, and Build. AutoPilot turns a conversation into a working agent. Agents shows runs, costs, and actions. Marketplace offers proven agents to customize. Build is a visual block editor for exact control. The shift is significant. AutoGPT now competes as a low-code agent platform, not as a research tool.

How the Platform Is Structured: Four Surfaces

The README describes four surfaces that share one platform. AutoPilot is the entry point for non-technical users: you describe the job in plain English, and the system converts the conversation into a working agent. Agents is the management view: it shows every agent, run, cost, and action that needs attention. That cost visibility is a practical feature, because agent runs consume real model usage. Marketplace is the reuse layer: you start from proven agents, add one to your library, and customize it. Build is the control layer: you drag, connect, branch, and inspect blocks. The key architectural point is that these are not separate products. They are views over the same agent runtime. The hosted platform and the self-hosted repository both include the core builder and agent runtime, according to the comparison table. That means the visual builder is not a hosted-only feature. You can run it on your own infrastructure if you want.

Getting It Running: Hosted vs Self-Hosted

There are two paths. The hosted platform is public, paid, and requires no setup. You sign up, and AutoGPT manages infrastructure, model access, credentials, and updates. It supports 45+ connected platforms and hundreds of AI models, with no model API keys needed. The self-hosted path is free of license fees, but you provide the infrastructure and model API keys. The install commands are simple. On macOS and Linux you run: curl -fsSL https://setup.agpt.co/install.sh -o install.sh && bash install.sh. On Windows PowerShell: powershell -c "iwr https://setup.agpt.co/install.bat -o install.bat; ./install.bat". The README notes that self-hosting requires Docker and configuration. The documentation link points to a self-hosting guide at docs.agpt.co/platform/getting-started. The trade-off is clear: the hosted path gets you running immediately, but you pay per agent run. The self-hosted path has no license fee, but you own the operations burden.

The Business Model Behind the Open Source

The README is unusually direct about why the hosted platform is paid. Every agent run consumes real model usage, compute, storage, secrets management, and operational support. The managed platform covers that infrastructure and funds continued development of the open-source project. This is a honest statement. It means the hosted service is not a free tier with limits. It is a usage-based paid service. The pricing page is linked, but the README does not give numbers. For engineers evaluating the project, this matters. The open-source code is free, but the hosted platform is the revenue engine. The self-hosted path remains available without a license fee, but you must provide and operate the resources yourself. The repository itself does not state a license, which is a gap. The README says self-hosting is the free path, but without a license identifier in the repository metadata, you cannot assume the code is open source under a permissive license. That is something to verify before adopting.

What the Repository Does Not Tell You

Several things are missing from the README. The README is marketing-heavy, with quotes from executives and a star count, but it does not describe the agent runtime's internal architecture. There is no mention of how agents handle long-running tasks, error recovery, or state persistence. The visual builder is described in terms of blocks, but not what those blocks are or how they connect to external systems beyond the 45+ connected platforms. The release notes for beta v0.7.2 and v0.7.3 are not detailed in the README. The repository is not archived and has recent pushes, but the platform is still in beta. That means the API and the builder may change. The README also does not discuss security or data handling for the hosted platform, which is a notable omission for an agent that handles credentials and runs workflows. For an engineer, this is a red flag. You would need to dig into the docs or the source to understand how secrets are stored.

Limitations and Wrong Tool Cases

The most obvious limitation is the cost model. Every agent run consumes real model usage, and the hosted platform charges for it. If you have a workflow that runs frequently, the cost can add up. The README does not give pricing, so you cannot estimate it without visiting the pricing page. Another limitation is the lack of offline capability. Self-hosting still requires model API keys, so you cannot run agents without external model access. That makes AutoGPT the wrong tool for air-gapped environments or for teams that need to keep all data on-premises with no external calls. The visual builder is a trade-off. It gives control, but it is not a programming language. If your workflow needs complex conditional logic, custom data transformations, or integration with a proprietary system that is not among the 45+ connected platforms, the builder may be limiting. The README does not say how you extend the platform beyond those integrations. For those cases, a code-first agent framework might be a better fit.

A Real Alternative: Code-First Agent Frameworks

The closest alternative is not another hosted platform. It is a code-first framework like LangChain or a custom Python loop using an LLM API. The difference is in approach. AutoGPT gives you a visual builder and a managed runtime, so you describe the workflow and the system handles orchestration. A code-first framework requires you to write the orchestration yourself: you define the steps, the tool calls, and the error handling in code. That gives you full control and full responsibility. You can debug with a debugger, write unit tests, and integrate with any library. The trade-off is that you must build the agent infrastructure from scratch, or rely on a framework that is not as turnkey as AutoGPT's builder. For a team that already has Python expertise and a complex, non-standard workflow, the code-first path may be more maintainable. For a team that wants to automate a standard business process quickly, AutoGPT's builder is faster to start with.

Maintenance and Upgrade Costs

The material gives some hints about maintenance. The self-hosted path requires you to manage updates and operations. The README says the hosted platform manages updates for you, but the self-hosted path does not. That means you are responsible for pulling new releases, migrating data, and handling breaking changes. The release names include beta, so the platform is not stable. The last push is recent, and there are two beta releases in the same month, which suggests active development but also potential churn. The README does not describe a migration path or a changelog, so you cannot assess the upgrade cost. The license is unknown, which is a legal risk. The README says self-hosting is free of license fees, but without an explicit license in the repository metadata, you cannot confirm that. Before using the code in a commercial product, you should check the repository for a LICENSE file or contact the maintainers.

Editorial conclusion

AutoGPT is for teams that want to build and run AI agents without writing agent orchestration code from scratch, especially if they accept the hosted platform's usage-based pricing or can manage their own Docker deployment. It is not for teams that need fine-grained control over every model call, strict data residency, or a fully offline agent runtime, since self-hosting still requires external model API keys. Before adopting, verify the beta status of the platform, the exact pricing for agent runs, and whether the visual builder's block model can express the workflows you need. The repository is actively developed, but the hosted path is a paid service, so the open-source code is only free if you operate it yourself.

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes