Compute

Compute is in private beta. It is open to the orgs we have let in, and the commands, flags and prices on these pages can change before it opens to everyone. Request access and we will write back.

The div CLI

div drives the same control plane the dashboard does, with an API key instead of a session. It is one static binary, written in Rust, and it is written for a pipe as much as for a terminal: an agent can drive it without parsing anything but exit codes.

div login
div run --gpu h100 train.py::run --max-spend 5.00

The first line signs this machine in through the browser. The second ships the directory you are in to a volume, places a run on one H100, stops it when it has cost five dollars, and prints the run’s log until it ends. div then exits with the run’s own exit code.

The commands

Command	What it does
div login	signs this machine in through the browser and saves the key it is given
div whoami	which org, key and control plane this machine uses, and where the key was found
div run	places a run, ships the working directory to it, and follows its log
div logs	prints a run’s log, or follows it until the run stops
div result	prints what a run declared about itself, or one value in it
div eval	measures a model version, and ranks it against the others
div model	names weights on a volume as a version to measure, and lists the registry
div runs	lists the org’s runs, and what the unfinished ones cost per hour
div stop	stops a run and prints what it spent
div storage	puts files on a volume and gets them back
div secrets	sets the values every run has in its environment
div papers	searches arXiv, or reads papers whole by id — no key needed
div catalog	what --gpu takes and what an hour of each costs
div limits	what the org holds against what its plan allows, and its spend against its usage limit
div dashboard	opens the dashboard in the browser

Everything else the control plane can do — endpoints, billing, keys with other scopes — is in the dashboard until it earns a command here.

An agent with an MCP client can skip the binary entirely: the MCP server is the same operations, as tools.

A first run

Sign in once per machine:

div login

It prints a code like QZHR-LVKG and opens the dashboard with it. Approve it there if the codes match. The key is saved to ~/.config/divergent/config.json, readable only by you.

Then, from the directory your code is in:

cd ~/very-cool-classifier
div run --gpu h100 train.py::run --max-spend 5.00

new project very-cool-classifier
14 files, 38.2 KB → very-cool-classifier-code/code/9f31c2ab
run-k4t9  H100  $4.95/h  $5.00 buys 1h 0m
...the run's own output...
run-k4t9 finished in 12m 40s and spent $1.05

The first line before the log is the one worth reading: what the run costs an hour, and how long the ceiling buys at that rate. Everything above the run’s output is ours and goes to stderr; the output itself goes to stdout untouched.