Deploy AutoGPT v0.6.59: Autonomous AI Agent Setup Guide

Here’s the unpopular take: most people Googling “how to deploy an autonomous AI agent” with AutoGPT don’t actually need AutoGPT. They need a workflow tool with an LLM block. The reason the original AutoGPT got famous in 2023 – agents looping endlessly until they solved a goal – is exactly why it failed in production. The team knew this. So they quietly killed it and shipped something completely different under the same name.

What you’re installing today is the AutoGPT Platform, a block-based agent builder closer to n8n than to a 2023 CLI demon. The classic CLI version still exists in the repo, but it’s a museum piece. This guide deploys the current platform – autogpt-platform-beta-v0.6.59, the latest release as of May 2026 – and covers the two install failures that account for most of the open GitHub issues.

Why self-host instead of waiting for the cloud beta

There’s a cloud-hosted beta with a waitlist for people who don’t want to deal with the technical setup. If you can wait, take it. Self-hosting makes sense in exactly three scenarios: you want custom Python blocks that touch internal systems, you’re testing before recommending it to a team, or you need agents that run continuously against your own infrastructure.

The repo is split into two licenses – and this matters before you touch a single file. Code inside the autogpt_platform folder uses the Polyform Shield License, per the repo README’s license section. Everything outside it – including the original standalone AutoGPT Agent, Forge, and agbenchmark – stays under MIT. Fork heavily or ship commercially? Read both before you start.

System requirements (the real numbers)

The platform runs a backend, a frontend, a Postgres via Supabase, Redis, and message queues – all in Docker. Old single-binary expectations don’t apply.

Those numbers are for an idle platform with no agents running (via community documentation, as of 2025 – verify against current docs). Once workflows start, vector databases and agent state typically need 4-16 GB RAM for moderate workloads, and a single continuous agent can consume 2-4 CPU cores. Plan accordingly if you’re on a small VPS.

Install AutoGPT v0.6.59 (the recommended path)

The official setup script handles dependency checks, Docker plumbing, and the initial build. Use it unless you have a reason not to.

macOS / Linux – from the official AutoGPT docs:

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

Windows: Check the official docs linked above for the current Windows installer – the path has changed across releases and the docs are the authoritative reference.

The manual route – worth knowing because the script can fail in interesting ways:

git clone https://github.com/Significant-Gravitas/AutoGPT.git
cd AutoGPT/autogpt_platform
cp .env.default .env
docker compose up -d --build

That cp step copies .env.default to .env inside autogpt_platform/ – that’s your config file. Edit it before composing up if you want custom environment variables baked in. The initial build pulls Supabase (Postgres + auth), Redis, the backend API, and the Next.js frontend. Plan for several minutes depending on your connection and whether the image layers are already cached.

Pro tip: Before docker compose up, regenerate the encryption key. Open autogpt_platform/backend/.env and replace the default. Generate a fresh one with: python -c "from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())". The default key is publicly visible in the repo – leaving it means anyone reading the source could decrypt your stored credentials.

First-time configuration and verification

Once the containers are healthy, open http://localhost:3000 – that’s where the official docs point you to confirm the server is running. You’ll hit a Supabase-backed signup screen. Create a local account. Emails aren’t actually sent unless you configure SMTP, which catches people the first time they try “forgot password.”

Quick sanity check from the command line:

docker compose ps

Every container should show healthy or running. If supabase-db shows anything else, skip straight to the next section.

To plug in a language model: builder UI → Credentials. The platform comes pre-integrated with OpenAI, Anthropic, Groq, and Llama (via llamafile). Paste the API key for whichever provider you want, then pick the model from the dropdown inside a block.

The two install failures that cause most GitHub issues

This is the section other tutorials skip.

1. Windows + Hyper-V = dead Supabase

Hyper-V causes supabase-db to be marked unhealthy – this is the single most reported install bug, confirmed across GitHub issues #9190 and #9846. The official docs are explicit: Docker on Windows must use the WSL 2 backend, not Hyper-V. Fix:

1. Open Docker Desktop → Settings → General
2. Check “Use the WSL 2 based engine”
3. Restart Docker
4. Run docker compose down -v then docker compose up -d --build again

Switching engines may erase existing containers and build history. Back up anything you care about first.

2. ARM / Apple Silicon with 16K page size

On recent Linux ARM hosts (or some Asahi setups), Supabase Postgres can refuse to start because it expects 4K memory pages. Check first:

getconf PAGESIZE

16384 means 16K – wrong. 4096 means 4K – correct. If you’re hitting 16K, the AutoGPT docs reference supabase/supabase issue #33816 for the full context. The workaround is OS-dependent (kernel parameter or alternate kernel image). Stock M1/M2/M3 macOS with Docker Desktop emulates 4K, so this trap usually hits cloud ARM VMs, not laptops.

Other gotchas worth knowing

• Port 3000 already in use – kill the conflicting process or change the frontend port mapping in docker-compose.yml.
• “Module not found” on Classic AutoGPT – you’re in the wrong folder. The Classic agent lives in classic/original_autogpt/; the Platform lives in autogpt_platform/. Different products, different setup paths.
• Tutorials linking to Torantulino/Auto-GPT.git – old org name. The canonical repo is Significant-Gravitas/AutoGPT.

Upgrading and uninstalling

To upgrade to a new release, pull the tag and rebuild:

cd AutoGPT
git fetch --tags
git checkout autogpt-platform-beta-v0.6.59
cd autogpt_platform
docker compose up -d --build

Before pasting that tag blindly, check the releases page for whatever is current – the team ships frequently. Read the changelog for breaking schema changes before upgrading a production instance.

To remove everything, including volumes:

cd AutoGPT/autogpt_platform
docker compose down -v
docker system prune -af
cd ../.. && rm -rf AutoGPT

The -v flag wipes the Supabase Postgres volume – agents, credentials, and run history are gone. If you want a clean reinstall but need to keep agent definitions, export them from the UI first.

The honest next step isn’t reading more tutorials. It’s running the install script, breaking something, and reading docker compose logs. Open the official AutoGPT docs in a second tab and start the clone. You’ll know within 20 minutes whether your machine is a good host or not.

FAQ

Is AutoGPT free?

The platform is open source and free to self-host. LLM API calls are not – you’ll pay OpenAI, Anthropic, or whoever you point it at.

Can I run AutoGPT without Docker?

For the Classic CLI version – yes. That one needs Python 3.10+, a clone of the repo, and a pip install inside classic/original_autogpt/. For the new Platform? No. It ships as a multi-service Docker Compose stack (backend, frontend, Postgres via Supabase, Redis, queues) and isn’t designed to run bare-metal. People who try usually give up around the Supabase auth setup. Use Docker.

What’s the difference between AutoGPT Classic and the AutoGPT Platform?

Classic is the original 2023 CLI agent – give it goals, it loops, it sometimes finishes. The Platform is a block-based visual builder for repeatable agent workflows, closer to a low-code automation tool with LLM nodes. They share a GitHub repo but are fundamentally different products under different licenses (MIT vs. Polyform Shield). Most modern tutorials, this one included, target the Platform.