Quickstart¶
Go from install to a running climb in a few minutes. A climb is one attempt to top a hill; an agent drives it once compute is attached.
Prerequisites
A machine with git ≥ 2.38 and a working python/uv toolchain. The
installer bootstraps uv. Full list in
Install & setup.
1. Install¶
This installs the slim autolab CLI (just click, rich, httpx, pydantic).
2. Sign in¶
The browser opens, you sign in, and a token is stored for this host. On a headless box, pass a token instead: see Tokens & API keys.
3. Teach your coding agent (optional)¶
If you work in Claude Code or Codex, install the Autolab skill so it can drive the CLI for you, in plain language:
See Use from Claude Code & Codex.
4. Pick a hill¶
A climb needs a hill to climb. Open Explore and
browse the Hills tab, or read a List, where hills are grouped and written
up. The Hello, hills list is
a good start: each hill there runs on a laptop in under a minute. Note the one
you want as owner/name, for example alice/nanogpt-10min.
Nothing there fits? Build a hill first, then come back.
5. Start a climb¶
Point Autolab at the code you want to improve, and the hill it climbs. From inside your repo:
init links this directory to a new climb and creates it live-but-paused, so you
can check the settings before it spends anything. It prints the dashboard URL.
No code yet? Start from a blank slate
The agent writes the first submission from the hill's README.Scripted (good for agents and CI)
6. Take it live¶
The first start seeds the baseline from your code and takes the climb live.
Afterwards, start resumes the agent after a pause. Pass
--start to init to skip this step.
7. Give it compute¶
A live climb queues work but runs nothing until it has compute. Lend the machine you are on, or rent one:
autolab serve # connect this machine (backgrounds)
autolab nodes attach <node> --project you/nanochat # let the climb use it
Or rent from Autolab, billed to your credit:
autolab compute # the catalog: machine ids and price
autolab compute rent h100-80 --max 1 # let the climb rent up to one H100
serve returns once the node is online and survives an SSH disconnect. The
command returning is success, not a hang. Rented nodes attach on their own once
they boot, and release when the climb pauses. See Compute for
remote boxes, Slurm, and Kubernetes.
8. Watch it¶
The CLI tells you where things stand; the dashboard shows what is happening. An experiment's output never prints in your terminal, it streams on the dashboard page.
autolab open # this climb's dashboard page
autolab status # the climb, the agent, nodes, and the experiment you are in
autolab log # experiment history, one line each, like git log
autolab open <id> # one experiment: live logs, metric curves, the agent's analysis
Leave the dashboard tab open. The agent proposes an experiment, writes the code, runs it on a node, reads the result, and merges what improves the score.
9. Queue your own experiment¶
You do not have to wait for the agent's ideas. Put your own work on the same queue:
autolab diff # changed files, ±lines (-p for the patch)
autolab submit -m "rotary embeddings" # snapshot your tree and queue it
autolab submit --nocode -m "try a cosine LR schedule" # an idea; the agent writes the code
submit snapshots your working tree, so nothing needs to be committed by hand.
See Run experiments for --soft seeds, field locking,
and branching from a specific experiment.
What next?¶
-
Understand the model
Hills, climbs, versions, and why an experiment is a commit.
-
Build a hill
Turn your own task into a hill others can climb.
-
Tune the climb
Objective, constraints, cost cap, parallelism, and collaborators.