A patch
$ git apply autooptm.patch $ python your/entry.py …
A plain diff that depends on nothing of ours. You run it exactly as before.
Make your code
faster
Give it a repo. Get back a patch.
First run free. See your repo's measured speedup, pay only to unlock the patch
before
937.8ms
→
after
249.2ms
same command · same inputs · same output
Measured on public repositories · end to end · output verified
23.41×csvkit
14.65×networkx
9.16×xarray
7.59×CycleGAN / pix2pix
7.48×litgpt
5.24×statsforecast
4.58×coqui TTS
3.35×Real-ESRGAN
3.03×sentence-transformers
2.68×moviepy
2.62×yolov5
2.25×WhisperLive
1.86×PySceneDetect
1.84×GroundingDINO
1.63×RIFE
1.41×nanoGPT
1.35×VibeVoice
1.19×whisperX
1.18×darts
23.41×csvkit
14.65×networkx
9.16×xarray
7.59×CycleGAN / pix2pix
7.48×litgpt
5.24×statsforecast
4.58×coqui TTS
3.35×Real-ESRGAN
3.03×sentence-transformers
2.68×moviepy
2.62×yolov5
2.25×WhisperLive
1.86×PySceneDetect
1.84×GroundingDINO
1.63×RIFE
1.41×nanoGPT
1.35×VibeVoice
1.19×whisperX
1.18×darts
Runs we submitted ourselves on these public repositories: the end-to-end wall-clock speedup of one command, output checked item by item against the original; the same code lands differently on other GPUs and inputs.
A number
3.75×
Your command, your inputs, sampled before and after, median reported.
03 Integrations
For agents and the terminal
no browser needed
Three doors onto the same backend: a CLI, MCP, and an agent skill. Every submission starts with a free estimate — seconds, and none of your code runs. Only continuing takes the 2-credit analysis deposit, credited in full against the patch.
CLI
$ pipx install autooptm $ autooptm login $ autooptm run . --entrypoint infer.py --gpu auto $ autooptm unlock <jobId> $ git apply autooptm.patch
login opens the browser to approve this machine (30-day token; --email for a mailed code on a server); run shows the free estimate, asks, then blocks until the job finishes and prints the measured speedup and the patch price. A failed run charges nothing. A local directory is listed before it is uploaded: weights, data, large files and crowded directories stay out until you say yes to each. Install the CLI with pipx (or uv tool install autooptm): Homebrew's Python on macOS and recent Debian / Ubuntu refuse a bare pip install into the system environment.
runestimatedecidestatusunlockcancelbalancedatasetskeysloginlogoutinstall-skill
MCP
{
"mcpServers": {
"autooptm": { "command": "autooptm-mcp" }
}
}
Run pipx install 'autooptm[mcp]', then paste this into Claude Code, Cursor, or any MCP client. optimize_submit parks at the free estimate and only optimize_decide takes the deposit; unlock_patch will not spend a credit without an explicit user confirmation.
optimize_submitoptimize_estimateoptimize_decideoptimize_statusoptimize_waitoptimize_cancelunlock_patchdownload_patchaccount_balance
Agent skill
$ pipx install autooptm $ autooptm install-skill
autooptm-optimize teaches Claude Code the whole loop: pick the entry script, pick the cheapest GPU the workload fits on, submit, report the number and the price, and never unlock until you say yes. Then just ask it to speed the repo up with AutoOptm.
Self-hosted
coming soon
The loop, running on your own machines: on the GPU you actually deploy on — an H100, a B200, anything our pool does not have — so the number in the report is that GPU's number. Tokens only, no machine line.
04 The alternatives
Every other route hands you a change.
Who proves it got faster?
| AutoOptm | Prompting an LLM yourself | Hiring a performance engineer | |
|---|---|---|---|
| How much of your time it takes | None. Submit and walk away. | All of it. You are the only one who notices when it goes wrong. | Briefing, questions, review |
| Does the output change | No. Anything outside tolerance is reverted. | Nobody checks | Depends on the engineer |
| Who proves it is faster | A frozen harness, both arms, median of repeated samples | Nobody. You build the harness. | The same person who wrote it |
| If it does not get faster | You are not charged | You still paid for the tokens | You still paid for the time |
Optimised and re-measured on the GPU your code actually runs on: the number in the report is that GPU's number, not one carried over from another.
RTX 4090RTX 5090
RTX 4090 and RTX 5090 today, with more GPUs on the way.
05 Who this is for
The same speedup,
four different ledgers
No procurement, no workflow change: one person and a repo URL is all it takes.
ML engineer at a tech company
½the compute, same quality
The production model retrains every week on a script nobody dares touch. Halve the step time and the same budget runs twice the experiments, better model, same bill. The hardest line in a performance review is the GPU spend you saved, with a number on it. Promotions follow numbers.
Grad student
2×experiments per week
The lab queue is always full and ideas outnumber GPU hours. Twice the step speed on the same GPU is a second GPU out of thin air: one more ablation before the deadline, one more column in the paper. It was never ideas you were short of, it was machine time.
Indie developer
2×the QPS, same GPU
Inference runs on GPUs you rent yourself, so latency and throughput are the margin. The same 4090 carrying twice the requests halves unit cost on the spot, that is where pricing room comes from.
Competition grinder
+3more models per night
A competition hands everyone the same deadline, not the same compute. While others train three models a night, you train six; leaderboard places are climbed one slot at a time.
06 Your code
Your code is deleted when the run ends
No repository access, and nothing of yours left behind.
Never leaves private storage
Your uploaded .zip stays in private object storage, used only to run your job, reproduce problems and improve the method — never published, never used to train third-party models.
Runs only in a throwaway sandbox
Your code executes only inside an isolated sandbox created for that one run and destroyed with it. Our own dispatcher never runs customer code: pip install on an unfamiliar repository is already arbitrary code execution, and that boundary protects both of us.
No repository access
We will never ask for an access token or a GitHub App install. Public repo: paste the URL. Private code: upload a .zip — you choose which files leave your machine instead of handing over an account.
Results are yours alone
Reports, logs and patches sit in private storage behind an identity check — holding the link is not enough. We do not publish your code or your numbers, and we do not use them as a case study without your explicit say-so.
The full terms are in the privacy policy.
07 Pricing
Metered on what it actually costs
Quoted first, charged after
Priced on the confirmed speedup, not a flat subscription. A free estimate comes first: a static scan that takes seconds, in which none of your code runs, so you see a projected speedup before deciding. Continuing takes the 2-credit analysis deposit, returned when no speedup is found or when the failure is ours: infrastructure failure, a timeout, a redeploy interrupting the run, the environment failing to build, an entry point whose running code we cannot locate, or a program too large for our biggest GPU — and credited in full against the patch; the run bills on actual usage. Skip the estimate and the deposit is taken at submit; cancelling within 5 minutes still returns it.
Trial pack
$3
10 credits · $0.30 / credit
- The best rate we offer, once per account
- ≈ 2–3 full runs (a run takes about 2–5 credits); credits valid 24 months
- Private code welcome (zip upload)
Usage pack
$5 +
$0.77 / credit · +30% in beta · pick an amount
- Credits valid 24 months
- $5 / $10 / $20 / $50 / $100 or a custom amount
- Private repos, results kept private
- A run takes about 2–5 credits ≈ $1.5–4, depending on code size, complexity, what the run finds and the speedup
Pro +50% credits · beta
$28 / mo
42 credits · $0.67 / credit
- 50% more credits per dollar during the beta
- ≈ 10–15 full runs per month
- Private repos · 5 concurrent
Max +60% credits · beta
$99 / mo
158.4 credits · $0.63 / credit
- ≈ 40–60 full runs per month
- Private repos · priority queue · 5 concurrent
Max+ +70% · beta
$199 / mo
338.3 credits · $0.59 / credit
- ≈ 85–130 full runs per month
- Top queue priority · 5 concurrent
- For heavy individual use: big repos, many runs in flight
During the beta top-ups get +30% and subscriptions +50% and up (Pro +50%, Max +60%, Max+ +70%). Subscription credits reset monthly; usage-pack credits stay valid for 24 months. Full rate card on the pricing page.
Find out what it gets on your repo
The estimate is free and lands in seconds. You decide whether to spend the two credits once you have seen what it projects.