Skip to content

Use AutoLab from Claude Code & Codex

If you work with a coding agent — Claude Code or Codex — you can teach it to drive AutoLab. autolab install drops a skill into the agent's config so it knows how to use this CLI: how to set up an AutoResearch project properly, queue experiments, attach execution nodes, and steer the agent — without you reciting the commands each time.

Install the skill

autolab install

With no argument it asks which tool(s) to set up. Or name them directly:

autolab install claude        # just Claude Code
autolab install codex         # just Codex
autolab install both          # both

That writes the skill — a lean SKILL.md plus references/ task playbooks the agent reads on demand (Codex also gets a small agents/openai.yaml metadata file) — into the tool's skills directory:

Tool Destination
Claude Code ~/.claude/skills/autolab/
Codex $CODEX_HOME/skills/autolab/ (defaults to ~/.codex/skills/autolab/)

Picking it up

Claude Code loads the skill on your next session. Codex needs a restart to discover a newly installed skill.

What the skill teaches

It turns the agent into an AutoResearch copilot that runs locally with you and drives AutoLab through the CLI. The same operating guide is used for both tools:

  • Orient first. On activation the agent checks for an existing .autolab project (here or in a parent) and, if found, summarizes where it stands — objective, agent state, nodes, recent experiments, metric trend — before doing anything else. (Working on a project that isn't checked out here? It lists your projects and clones one.)
  • Set a project up well. If there's no project yet, it inspects the repo and proposes the target, metric, and constraints (rather than asking blindly), then pins down a specific objective, a comparability rule, explicit guardrails (what would count as "cheating"), run/setup commands, a stop policy, and a cost cap.
  • Attach compute. Nothing runs without an execution node, so it helps you connect one — this machine, or a remote GPU box over SSH.
  • Act as the copilot. Status and summaries, explaining results and failures from logs, building plots and reports locally, and writing new experiments that get submitted to AutoLab for compute — plus steering the agent (start/pause, autogen). It writes and submits experiments rather than running training on your machine, and keeps an eye on spend against --max-cost.
  • Stay safe. It spends real money, so it sets --max-cost; treats tokens as secrets; and drives the workspace only through autolab verbs.

Depth lives in references/ playbooks (projects, experiments, compute, agent) the agent opens only when it needs the specifics — the exhaustive flag list always comes from autolab <command> --help, so nothing drifts.

Once installed, just ask your agent in plain language — "set up an AutoLab project to minimize val loss on this repo and attach my GPU box" or "summarize what the agent did overnight and queue an experiment trying a cosine LR schedule" — and it will use the right autolab commands.

Safe by default

autolab install never clobbers your edits. An existing file that differs is reported as a conflict and left untouched; re-run with --force to overwrite it. A file that's already identical is left alone (already up to date).

autolab install both --dry-run   # show exactly what would change — write nothing
autolab install claude --force   # overwrite an out-of-date skill

If you've customized the installed SKILL.md, install won't overwrite your version unless you pass --force.

Prerequisites

The agent still needs the CLI itself installed and signed in:

autolab --version     # the CLI is on PATH (see Install & setup)
autolab whoami        # signed in to the right host

On a headless box, the agent authenticates with autolab login --token or the AUTOLAB_TOKEN environment variable — see Tokens & API keys.

See every install flag in the CLI reference, and the model the skill teaches in Core concepts.