Skip to content

Install & setup

Prerequisites

  • git on your PATH — the CLI shells out to git to snapshot and push your code. git ≥ 2.38 is recommended (the control node needs it for clean experiment merges).
  • Python 3.11+ — for the CLI itself. The installer bootstraps uv to manage it.
  • Your project's own toolchain (PyTorch, CUDA, uv/pip, …) on whatever machine runs the experiments. The CLI is dependency-light and stays out of the way.

Install

curl -fsSL https://app.autolab.ai/install.sh | sh

The installer bootstraps uv (a single static binary — no system Python needed) and installs the slim autolab CLI as an isolated uv tool. Verify:

autolab --version
autolab --help

autolab: command not found?

The CLI installs to uv's tool bin. If your shell can't find it yet, add it to your PATH:

export PATH="$HOME/.local/bin:$PATH"

Self-hosted Autolab instance

The installer pulls the CLI wheel from the Autolab host that serves it. If your organization runs its own Autolab, install from that host and point the CLI at it with AUTOLAB_URL:

curl -fsSL https://autolab.your-company.com/install.sh | AUTOLAB_URL=https://autolab.your-company.com sh

Coding agents (Claude Code & Codex)

Driving AutoLab from a coding agent? Install the skill so it knows the CLI and how to set up and run research projects:

autolab install        # asks which tool; or: autolab install claude | codex | both

See Use from Claude Code & Codex.

Sign in

autolab login          # opens a browser, stores a token for this host
autolab whoami         # show the active host and signed-in user

Tokens are stored per host in ~/.config/autolab/credentials.json (owner-only, mode 0600), so logins to different hosts never collide. To sign in without a browser — on a remote or CI box — pass a token:

autolab login --token "$AUTOLAB_TOKEN"

See Tokens & API keys for minting and revoking tokens.

Drive AutoLab from your coding agent

Using Claude Code or Codex? Run autolab install to drop a skill into it so it can set up projects and run experiments for you. See Use from Claude Code & Codex.

Choosing a control node

By default the CLI talks to https://app.autolab.ai. Override the host when you need to — for example, to point at your organization's self-hosted Autolab:

autolab --url https://autolab.your-company.com status   # one-off
export AUTOLAB_URL=https://autolab.your-company.com      # whole shell

Or define named profiles in ~/.config/autolab/config.toml:

default_profile = "prod"

[profiles.prod]
url = "https://app.autolab.ai"

[profiles.selfhosted]
url = "https://autolab.your-company.com"
autolab --profile selfhosted status

Resolution order is --urlAUTOLAB_URL → active profile → https://app.autolab.ai. Full details in Configuration & environment.

Update & uninstall

autolab update             # update to the build your control node serves
autolab update --check     # just compare versions (exit 1 when out of sync)
uv tool uninstall autolab  # remove the CLI

Re-running the install one-liner does the same as autolab update. A machine running a node keeps the old code until the daemon restarts: autolab serve stop && autolab serve --project <owner>/<slug>.


Next: the Quickstart, or Core concepts to understand the model before you dive in.