autooptm

Measured GPU speedups for your training and inference code, from the terminal, an agent, or Python. The estimate is free; continuing takes a 2-credit analysis deposit, credited in full against the patch; the patch is what costs credits.

pipx install autooptm          # a CLI tool: pipx or `uv tool install autooptm`, not a bare pip
                               # (Homebrew's Python on macOS and recent Debian/Ubuntu refuse it)
autooptm login                 # opens the browser to approve this machine (30-day token)
autooptm run https://github.com/you/repo --entrypoint train.py
autooptm wait <jobId>          # resume waiting after a dropped connection or --no-wait
autooptm unlock <jobId>        # asks before spending credits
git apply autooptm.patch

A local directory (autooptm run .) is packed on your machine, and you see the list before it leaves: .git, caches and virtualenvs are dropped without asking; weight files, data files, files over 8 MB and directories with over 200 files are each shown and asked about (they stay out unless you say yes, or answer d to send a directory up as the private dataset instead, unpacked at the same path in the sandbox); then the packing list is printed and upload? [Y/n] waits. --dry-run only shows the list, --include PATH / --exclude PATH answer from the command line, and --yes takes the defaults without asking (off a terminal, run will not upload without it).

run first gets the free estimate and shows it (expected speedup, the range, a lossless line, and the reviewer's note if something in the submission would stop the run); it then asks before taking the analysis deposit. --yes continues without asking, --no-estimate skips the estimate (the deposit is taken at submit). A job left at the estimate can be answered later:

autooptm estimate <jobId>            # the numbers again
autooptm decide <jobId> continue     # takes the deposit, credited against the unlock
autooptm decide <jobId> stop         # free

autooptm login --email you@example.com is the no-browser door (a mailed code) for servers and CI shells; --no-browser prints the approval link instead of opening it.

Python:

from autooptm import AutoOptm
ao = AutoOptm()
job = ao.submit(".", entrypoint="bench.py", workload="inference")
done = ao.wait(job["jobId"])
print(done["speedup"], done["unlock_credits"])

MCP (Claude Code, Cursor, any MCP client), pipx install 'autooptm[mcp]':

{ "mcpServers": { "autooptm": { "command": "autooptm-mcp" } } }

Tools: optimize_submit, optimize_status, optimize_wait, unlock_patch (requires explicit user confirmation), download_patch, account_balance.

API keys (CI and scripts)

autooptm login saves a 30-day token that slides on use, which is the right credential for a laptop and the wrong one for a build machine: it expires while the pipeline is idle, it needs a mailbox to renew, and it can spend credits. A key fixes all three.

autooptm login                      # a key is minted from a session
autooptm keys create "CI pipeline"  # printed once, never again
autooptm keys list
autooptm keys revoke aok_1a2b3c4d5e6f

Put it in CI as AUTOOPTM_TOKEN and it works everywhere a login token does.

Keys carry scopes, and the default is read,submit. unlock is opt-in, so a build machine can submit work and read the result without being able to spend a credit, whatever the build does. Add --scopes read,submit,unlock only if you mean it, and --expires-in-days N to make it die on its own.