Skip to content

AutoGPT is the vision of accessible AI for everyone, to use and to build on. Our mission is to provide the tools, so that you can focus on what matters.

10.9 KB
Raw Blame

AutoGPT Platform Release Installer

The release installer runs the published AutoGPT Platform appliance as one Docker container. It does not clone the repository, build AutoGPT, install Docker, elevate privileges, or accept Docker Desktop license terms.

The appliance exposes one loopback-only port. Application state and generated secrets live in a named Docker volume; installer identity and runtime configuration live in a dedicated private state directory.

Prerequisites

Before running the installer:

  • Install and start Docker from the official Docker Engine or Docker Desktop documentation.
  • Select a local Docker endpoint: unix:// on Linux and macOS, npipe:// on Windows. Remote and TCP contexts are rejected.
  • Configure the daemon for Linux containers on amd64 or arm64.
  • Allow about 25 GB of free disk; at least 8 GB RAM is recommended.
  • Have Bash, curl, and sha256sum or shasum available for the bootstrap, or PowerShell and curl.exe on Windows. Docker Compose, Git, Node.js, and npm are not required.

Docker Desktop is a separate product with its own license terms. AutoGPT does not install it or grant a Docker license.

Install

Each command below uses a unique temporary file, executes it only after the HTTPS download succeeds, and removes it afterward. Do not stream a network response directly into a shell.

Linux and macOS

(
  installer="$(mktemp)" &&
  trap 'rm -f "$installer"' EXIT &&
  curl --proto '=https' --proto-redir '=https' --tlsv1.2 -fsSL \
    -o "$installer" https://setup.agpt.co/install.sh &&
  bash "$installer"
)

Windows

& {
  $installer = Join-Path ([IO.Path]::GetTempPath()) "autogpt-install-$([guid]::NewGuid().ToString('N')).ps1"
  try {
    curl.exe --proto '=https' --proto-redir '=https' --tlsv1.2 -fsSL -o $installer https://setup.agpt.co/install.ps1
    if ($LASTEXITCODE -eq 0) {
      powershell.exe -NoProfile -ExecutionPolicy Bypass -File $installer
    }
  } finally {
    Remove-Item -LiteralPath $installer -Force -ErrorAction SilentlyContinue
  }
}

Options

Add a flag after bash "$installer" on Linux and macOS, or after -File $installer on Windows.

Goal Linux/macOS Windows
Current published appliance (no flag) (no flag)
Specific published version --release=vX.Y.Z -Release vX.Y.Z
Appliance release-tag form --release=autogpt-platform-beta-vX.Y.Z -Release autogpt-platform-beta-vX.Y.Z
Hardware and Docker checks only --preflight-only -PreflightOnly
Print image selection only --resolve-only -ResolveOnly
Skip RAM/disk checks --skip-preflight -SkipPreflight
Custom private state directory --dir=PATH -Dir PATH

--skip-preflight does not skip the local-endpoint, Linux-container, or architecture checks.

The default state directory is $XDG_CONFIG_HOME/autogpt, or $HOME/.config/autogpt when XDG_CONFIG_HOME is unset, and %LOCALAPPDATA%\AutoGPT on Windows. A custom directory must be user-owned and private; the installer rejects symlink paths and shared or unsafe locations.

Artifact and Runtime Contract

With no release flag, the installer pulls significantgravitas/autogpt:latest. An explicit vX.Y.Z selects the matching version tag. After pulling, the installer:

  1. Requires the expected OCI title and source repository plus a well-formed 40-hex source revision label.
  2. Requires a Linux image matching the local daemon's amd64 or arm64 architecture.
  3. Resolves exactly one native RepoDigest and runs that immutable digest, never the mutable tag.
  4. Creates an installer-owned, labelled autogpt-platform-data volume.
  5. Starts autogpt with 127.0.0.1:3000, the private environment file, restart and stop policies, log rotation, shared memory, and file-descriptor limits.

The bootstrap authenticates transport with HTTPS, and the installer pins the pulled image by digest after checking its appliance metadata. It does not independently verify a publisher signature or protect against compromise of the image-publishing credentials. Signed release artifacts are a follow-up, not a claim made by this installer.

The installer refuses to adopt an unlabelled container or volume. A rerun is a no-op, or starts a stopped container, only when the installer identity, environment hash, immutable image digest, process, port, mount, privilege, restart, logging, memory, and limit settings all match. Drift fails closed.

If established installer lifecycle state exists but its named volume is missing, the installer refuses to create an empty replacement. An interrupted first pull has no established marker and can be retried safely. Restore a previously established volume or choose a new state directory.

Runtime Configuration

The private autogpt.env initially contains:

AUTOGPT_PUBLIC_URL=http://localhost:3000

Add provider keys and other appliance variables there. The installer binds the environment file's SHA-256 hash to the container label. To apply a deliberate configuration change, stop and remove only the container, then rerun the same installer release:

docker stop --time 360 autogpt
docker rm autogpt

Do not remove autogpt-platform-data; it contains accounts, agents, generated secrets, and application state.

First Run

After the installer reports a healthy container, open http://localhost:3000. Registration starts open so you can create the intended account. Promote it:

docker exec autogpt autogpt-admin promote [email protected]

Then set this in autogpt.env, recreate the container as described above, and verify registration is closed:

AUTH_ALLOW_NEW_ACCOUNTS=false

Keep the default loopback binding. To serve other machines, place AutoGPT behind a TLS reverse proxy and set AUTOGPT_PUBLIC_URL to the exact public URL.

Upgrades

Before an upgrade, back up autogpt-platform-data as described in Cold backup. Stop and remove only the autogpt container, then rerun the installer with the intended release. The installer will reuse the named volume only when its ownership labels and the private installer identity still match.

Maintainer Release Gates

https://setup.agpt.co/install.sh, https://setup.agpt.co/install.ps1, and https://setup.agpt.co/install.bat redirect to the files of the same name in autogpt_platform/installer/ on master, so the hosted installers change whenever those files change on master. Repository CI cannot prove this public bootstrap handoff, so each release must keep every gate below passing:

  • Only a published, non-prerelease autogpt-platform-beta-vX.Y.Z release publishes images. Its run smoke-tests and scans both architectures, then pushes sha-<commit> and vX.Y.Z and moves :latest forward. Pushes to dev and manual dispatches build, test, and scan without publishing.

  • The release's tag commit contains the full installer, publication workflow, and helper, and its release-triggered run succeeded.

  • Both public tags expose Linux amd64 and arm64 manifests and the expected OCI identity:

    docker buildx imagetools inspect significantgravitas/autogpt:vX.Y.Z
    docker buildx imagetools inspect significantgravitas/autogpt:latest
    
  • Fetch each hosted installer into a new temporary file from an external machine and compare it with the released repository file. On each supported operating system, perform a clean install against the public image, wait for health, verify the loopback endpoint, rerun to prove exact-contract idempotency, and confirm no mutable tag was used to create the container.

Installing from source (Compose)

The scripts below are the other supported path: they clone the repository and start every service with Docker Compose. Use them for development, or for a fully offline install with a local LLM. The appliance described above is the image documented in Run AutoGPT in one Docker container; these scripts do not use it.

Manual Installation

If you prefer, you can manually download and run the installer scripts:

  • Linux/macOS: setup-autogpt.sh
  • Windows: setup-autogpt.bat

These scripts are located in the autogpt_platform/installer/ directory. Run ./setup-autogpt.sh or setup-autogpt.bat from there. Inside a clone, a script uses that checkout; otherwise it clones the repository into a new AutoGPT folder first.

Both create the three .env files if they are missing and generate the secrets .env.default leaves blank (ENCRYPTION_KEY, UNSUBSCRIBE_SECRET_KEY, INTERNAL_API_KEY, BETTER_AUTH_SECRET), the same as make init-env. Running a script again never changes a value that is already set. If you are upgrading an install that ran without its own ENCRYPTION_KEY, read Upgrading: secrets are generated per install first so your connected integrations move to the new key.

Running fully offline with a local LLM (Ollama)

Both installer scripts accept an opt-in flag that installs Ollama, pulls a default chat model, and wires backend/.env so AutoPilot runs without any cloud API keys. This is useful for air-gapped or privacy-sensitive deployments — see Running AutoPilot on a self-hosted LLM for the full reference.

Linux / macOS

cd autogpt_platform/installer
./setup-autogpt.sh --with-ollama
# Optional overrides:
#   --ollama-model=qwen3:14b-q4_K_M
#   --ollama-host=http://gpu-rig.lab:11434   # use an existing Ollama

Windows

cd autogpt_platform\installer
setup-autogpt.bat /with-ollama
REM Optional overrides:
REM   /ollama-model=qwen3:14b-q4_K_M
REM   /ollama-host=http://gpu-rig.lab:11434

The installer:

  1. Installs Ollama (skipped if already present, or if --ollama-host points at an existing one).
  2. Configures OLLAMA_HOST=0.0.0.0:11434 + OLLAMA_CONTEXT_LENGTH=262144 so containers can reach it and so AutoPilot's ~8 k system prompt isn't truncated by Ollama's 4 k default; 262,144 tokens is the default model's native window.
  3. Pulls the chat model (default hf.co/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M).
  4. Appends a marker-bounded block to autogpt_platform/backend/.env with CHAT_USE_LOCAL=true plus the CHAT_BASE_URL / CHAT_API_KEY / CHAT_*_MODEL overrides.

Re-running with --with-ollama is idempotent — the wiring block is rewritten in place.

Troubleshooting

  1. Confirm docker info succeeds against a local Linux-container daemon.
  2. If an image pull fails just after release, wait for public manifest publication and retry; the installer never falls back to a source build.
  3. If the installer reports state or runtime drift, preserve the named volume, inspect the existing container, and follow the explicit upgrade procedure.
  4. Check docker logs --tail 100 autogpt when health checks fail.