AutoGPT: what the open-source agent platform actually installs, and who should self-host it

What AutoGPT is for, and the two audiences it splits into

AutoGPT is not one program. The repository root holds two directories, autogpt_platform/ and classic/, which correspond to two different generations of the project. The README describes the current product as a platform where you "describe an outcome in plain English or shape every step in the visual builder, then run the agent on demand, on a schedule, or from a trigger." That sentence is the whole pitch, and it also defines the split in its audience.

The first audience wants an outcome and does not want to operate anything. For them the README points at the hosted Platform, where AutoGPT manages "infrastructure, model access, credentials, reliability, and updates." The second audience wants the same builder and runtime but on their own machines. The README is explicit that self-hosting is "the free path" and that you "provide the infrastructure and model API keys, and you maintain the deployment."

What the project is not is a library you import into an existing Python service. Nothing in the README or the top-level layout suggests a pip-installable package with a stable API. The unit of adoption is a deployment, not a dependency. That single fact should decide most adoption questions before you read further.

Four surfaces and the runtime behind them

The README names four surfaces. AutoPilot turns a plain-English conversation into a working agent. Agents is the monitoring view, showing "every agent, run, cost, and action that needs your attention." Marketplace is a library of prebuilt agents you can add and customize. Build is the visual editor where you "drag, connect, branch, and inspect blocks for exact control over every step."

The mechanism underneath is a block graph. The Build surface lets you connect and branch blocks, and the same graph is what runs when an agent is triggered. That is a meaningful design choice: the visual builder is not a code generator that emits a script you then own, it is the runtime's own representation. The practical consequence is that anything you assemble in the builder stays editable there, and anything AutoPilot generates lands in the same structure.

Execution is triggered three ways according to the README: on demand, on a schedule, or from a trigger. The README does not enumerate which triggers exist beyond that, so treat trigger coverage as something to check in the docs rather than assume. The platform also advertises "45+ connected platforms and hundreds of AI models," which is the integration surface agents can act against.

One thing the README states plainly and that is easy to skim past: both the managed Platform and the self-hosted deployment use the same repository and include "Core builder and agent runtime." The difference is operational, not functional.

Installing AutoGPT: the self-host script and the Windows path

The README gives one-line installers for macOS and Linux, and a separate PowerShell command for Windows. On macOS or Linux you download the install script and run it in the same command:

curl -fsSL https://setup.agpt.co/install.sh -o install.sh && bash install.sh

On Windows the README uses PowerShell to fetch and execute a batch file:

powershell -c "iwr https://setup.agpt.co/install.bat -o install.bat; ./install.bat"

Both scripts pull from setup.agpt.co. The README does not document what the scripts install, which ports they expose, or which environment variables they set, so read the script before you run it on a machine you care about. The README's own summary of the self-hosted path is that it requires "Docker and configuration."

After installation, the README points to the self-hosting guide at docs.agpt.co/platform/getting-started rather than reproducing the configuration steps. That guide is where model API keys get wired in, since self-hosting means bringing your own. There is no documented example in the README of a completed configuration file, so budget time for the docs before you expect a first agent run.

If you would rather not run any of this, the alternative in the README is the hosted Platform at platform.agpt.co, which it describes as requiring "No model API keys or infrastructure setup."

The cost model is the real fork in the road

The README devotes a section to why the hosted Platform is paid, and the reasoning is worth reading as a design statement. Every agent run "consumes real model usage, compute, storage, secrets management, and operational support," and the managed Platform covers that while funding development of the open-source project. Self-hosting stays available "without a license fee" for people who provide those resources themselves.

So the comparison table in the README comes down to this. Hosted: public signup, a paid plan plus agent usage, managed setup, built-in model access, AutoGPT-managed updates, data hosted by AutoGPT, plan-dependent support. Self-hosted: clone and install, no license fee but you pay your own infrastructure and model providers, Docker and configuration required, bring your own API keys, you manage updates, your infrastructure holds the data, community support.

The honest reading is that self-hosting is not the cheap option, it is the option where you own the bill and the blast radius. Model usage is the dominant cost either way, and it does not disappear when you self-host, it just moves to your provider account. What you actually save is the management fee; what you actually pay is the operator time.

One gap: the README does not state which model providers are supported for self-hosted deployments, only that you bring your own keys. Check the docs before assuming your provider of choice works.

Where AutoGPT is the wrong tool

If you need deterministic, auditable execution of a fixed procedure, a block graph that can branch and call language models is the wrong shape. The README's own framing, "AI agents that finish the work," implies judgment calls at run time. Anything that must produce byte-identical output from identical input belongs in ordinary code.

The second mismatch is operational. The README does not document rollback, version pinning, or how to move an agent definition between a self-hosted deployment and the hosted Platform. The repository does carry release tags such as autogpt-platform-beta-v0.7.3, and the word beta in that tag is a fair signal about the stability you should expect. If your change-management process requires a documented downgrade path, that path is not in the README.

The third is cost control. Agents that run on a schedule against hundreds of models can spend money while nobody is watching. The README mentions the Agents surface shows cost per run, which helps after the fact, but neither the README nor the comparison table describes spend caps or budget alerts. Verify that before you point an agent at a paid model key.

Finally, the README does not describe a sandboxing model for the actions agents take against connected platforms. With 45+ integrations, that is the question a security review will ask first, and the README does not answer it.

AutoGPT against a plain LLM chat interface

The most common comparison is AutoGPT against a chat assistant, and the README's own framing makes the difference clear rather than marketing it. A chat interface returns text. AutoGPT "builds the agent, runs it, and reports back," and the Agents surface tracks runs, costs and actions rather than messages.

The deeper difference is state. A chat session is a transcript. An AutoGPT agent is a graph of blocks with triggers attached, which can be invoked on a schedule without a human in the loop. That is what makes it useful for recurring work such as a daily brief assembled from internal and external signals, one of the examples the README lists.

The cost of that capability is everything described above: a deployment to run, credentials to manage, and a bill that scales with usage rather than with seats. If your actual need is occasional question answering, a chat interface is cheaper and has no operational surface at all. AutoGPT earns its keep when the work repeats and the steps are stable enough to encode once.

Maintenance, licensing and what to verify

The repository is not archived, and the last push was on 2026-08-28, which coincides with the autogpt-platform-beta-v0.7.3 release. Releases arrive on a roughly weekly cadence in the recent list, with autogpt-platform-beta-v0.7.2 on 2026-08-21 and a rolling preview-seed-fixture tag on 2026-08-06. A self-hoster inherits that cadence as upgrade work, because the README states that updates and operations are "Managed by you" on the self-hosted path.

Licensing is the part to check yourself. The repository root contains a LICENSE file, but the README does not state which licence it contains, and the hosted Platform is a separate paid service with its own terms at agpt.co/pricing. Do not assume the licence that covers the code also covers the managed service, or the reverse. Read the LICENSE file directly, and if you plan to redistribute or embed the platform commercially, get your own advice rather than inferring from the README's "no license fee" phrasing, which describes self-hosting cost, not licensing terms.

The classic/ directory at the repository root is a reminder that this project has been through at least one major rearchitecture. If you find older setup instructions elsewhere that reference classic, they describe a different codebase from the one the current README documents.

Editorial conclusion

Adopt the managed Platform if you want agents running this week and can accept usage-based billing plus AutoGPT holding your credentials and data. Self-host if infrastructure control or data residency outweighs the operational work, and you have someone who can run Docker and supply model API keys. Do not adopt either if you need a documented rollback or version-pinning story, because the README does not cover either. Before committing, open the self-hosting guide at docs.agpt.co/platform/getting-started, read the LICENSE file at the repository root, and confirm the current pricing page, since the README states the hosted Platform is paid and usage-based.

Frequently asked questions

Official sources

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