AutoGPT now ships as Otto plus a team of Experts, hosted or in one Docker container

Otto coordinates named Experts instead of running one open-ended agent

AutoGPT now presents itself as a team rather than a single autonomous agent. You talk to Otto, labelled Head of AI, and Otto brings in specialists, hands over the work and keeps you up to date. You can also talk to a specialist directly when you want to get into details.

The ready-made Experts each have a narrow job. Maria, an SEO Content Manager, turns a target keyword into a content brief and an article draft. Max, an Account Executive, researches prospects, drafts outreach and prepares the next conversation. Frankie, an Executive Assistant, prepares meeting context, email drafts and a digest for the day. Each Expert's role is shaped by your instructions, context, skills and connected apps, and you can create your own specialist instead of starting from one of these.

The demo run prepares work for review and sends nothing

The project's demo is one launch brief given to Otto. Maria prepares the article, Max researches prospects and Frankie prepares the launch review. The clip is a 15-second highlight of a real run, edited and sped up, and in that run the team prepares the work for review without sending or publishing it.

That matches the intended loop: choose an Expert, brief it on the job and connect the apps it needs, then review the output and adjust the instructions. Routines can be activated for jobs that keep coming back. The consequence for anyone evaluating it is that what you see demonstrated is drafts and research, not unattended outbound action, so a human review step is part of the workflow. The controls on offer are choosing which accounts each Expert can access, setting spending controls, and following progress in each Expert's activity.

The hosted Platform bills for a plan and for usage

The hosted Platform is the route that avoids managing infrastructure or supplying model API keys. Model access is built in, AutoGPT manages setup and operations, and data and infrastructure are hosted by AutoGPT. A trial is available through sign-up at platform.agpt.co.

It is a paid service: cost is a paid plan plus usage, with plans, trial terms and usage pricing published at agpt.co/pricing. Two groups should not start here. Anyone who cannot place business accounts and working data on a third-party host has to self-host instead, and anyone expecting AutoGPT to be free in its hosted form will find that it is not.

The single-container image listens only on 127.0.0.1

The quickest self-host path is the significantgravitas/autogpt image on Docker Hub:

docker run -d \
  --name autogpt \
  --restart unless-stopped \
  --shm-size 2g \
  --ulimit nofile=65536:65536 \
  --log-driver json-file \
  --log-opt max-size=50m \
  --log-opt max-file=5 \
  -p 127.0.0.1:3000:3000 \
  -e AUTOGPT_PUBLIC_URL=http://localhost:3000 \
  -v autogpt-data:/data \
  significantgravitas/autogpt:latest

Each flag carries a decision. The port mapping 127.0.0.1:3000:3000 publishes the web UI on loopback only, so with this exact command the instance is reachable from the host machine and not from other machines on the network. AUTOGPT_PUBLIC_URL is set to http://localhost:3000 to match. All state lives in the named volume autogpt-data, mounted at /data, which makes that volume the thing to back up. Shared memory is raised to 2g, the open-file limit to 65536, and container logs are rotated at 50m per file with five files kept. The image tag is latest, so the version you get depends on when you pull. Give the first boot a few minutes before opening http://localhost:3000.

Your first account is an ordinary user until you promote it

After the container starts you create an account in the browser, and it does not become the administrator on its own. You promote it from the host:

docker exec autogpt autogpt-admin promote [email protected]

The container name autogpt in that command comes from the --name flag in the run command, so if you rename the container you have to change it here too. Closing signup is not part of the run command; it is one of the topics covered in the single-container guide at agpt.co/docs/platform/self-hosting/single-container, together with requirements, ports, model providers, backups and upgrades. Treat closing signup as a step to finish before you widen the port binding beyond localhost.

Self-hosting moves model credentials, upgrades and backups onto you

Running AutoGPT on your own infrastructure means you provide model credentials, configure the deployment, and manage its updates and operation. Model access becomes bring your own credentials or a local model, and cost becomes your infrastructure plus your model usage. Nothing about upgrades or backups is handled for you, which is the main practical difference from the hosted Platform beyond price.

There are three self-host routes. The single container shown above is the shortest. An installer for Linux, macOS and Windows sets up the same image with one command; its instructions are at agpt.co/docs/platform/self-hosting/installer, and that page is the place to start on Windows. To run each service in its own container with Docker Compose, or to develop from source, the manual self-hosting guide at agpt.co/docs/platform/self-hosting/getting-started covers the setup. The more services you split out, the more of the operation you own.

Workflows, the marketplace and an MCP server sit under the Experts

Experts can use skills and workflows to do their jobs, and the workflow layer is where you get control over individual steps. The visual builder lets you connect blocks, branch logic, define repeatable steps and inspect how work runs. The marketplace on platform.agpt.co offers workflows you can adapt instead of building from scratch. Integrations connect the tools your team already uses, with a list at agpt.co/docs/integrations.

For developers, an API and an MCP server let you run and build workflows from your own code or from an AI coding agent; the API guide is at agpt.co/docs/platform/api-and-integrations/api-guide. This is the part to look at if the Expert abstraction is too coarse for a job and you need deterministic, inspectable steps.

Two code trees and a beta-tagged release line

The repository root holds autogpt_platform/, which contains the Platform source and its development setup, next to a separate classic/ directory, a docs/ directory, and agent instruction files such as AGENTS.md and CLAUDE.md. The default branch is master.

The project is moving quickly. The last push landed on 2026-11-02, and recent releases are tagged autogpt-platform-beta: v0.9.0 on 2026-10-23 and v0.10.0 on 2026-10-27, four days apart. The beta label and that pace mean a self-hosted instance tracking latest will change underneath you often. GitHub does not identify a standard license for the repository, and the README points to its license section for usage terms, so read LICENSE before building a product on the code.

Editorial conclusion

AutoGPT suits teams that want role-based agents to draft content, research prospects and prepare assistant work which a person reviews before anything goes out, either on the paid hosted Platform or in a self-hosted container with their own model credentials. It does not suit anyone looking for a free hosted service or a stable, non-beta release line. Before relying on it, read the LICENSE file and the single-container guide's material on closing signup, backups and upgrades.

Frequently asked questions

Official sources

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