Early. Building in the open.
Spare AI credit,
put to work on open problems.
Most plans and API budgets end the month with some headroom unused, and it simply expires. SpareTokens is a Folding@home for AI tokens: donate that headroom to a shared queue of open scientific problems, and have every answer checked by a machine.
The problem
Expiring credit is wasted compute.
Plan allowances reset. API budgets get set high and not fully used. Local machines sit idle overnight. None of that capacity is lost to anyone's benefit, it just disappears.
Meanwhile there are well-defined open problems where a tireless, careful reader can still make progress: checking whether a conjecture has quietly been settled in the literature, running an exhaustive search, or formalizing a proof so a proof checker can confirm it.
How it works
Three steps.
- 01
Run the worker
Install the open-source worker on your own machine and point it at your own credentials, with a spend cap you choose.
- 02
It pulls a task
The worker claims a task from the shared queue, works on it in a scratch directory, and submits a structured result.
- 03
We re-verify it
The coordinator never takes the worker's word for it. Proofs are compiled in Lean, computations are re-run independently, and literature claims need agreement and quotes.
Status: the pipeline runs locally and has been piloted. There is no public worker download yet. That is what the waitlist is for.
Trust model
Your key stays on your machine.
Every result is machine-verified.
Keys never leave
The worker runs locally with your credentials. The coordinator only sends tasks and receives results. By construction, your key is never sent to it.
No trust in workers
The coordinator does not run worker code and does not believe worker claims. Verification code that we own checks each result, and anything it cannot check goes to a human.
Open source client
The worker is open source so you can read exactly what runs on your machine and what it sends.
"Verified" means a machine check passed. It does not by itself mean a result is new or important, and a human reviews anything that could be a real claim.
First goal
Unsolved Erdős problems.
erdosproblems.com catalogues hundreds of open questions posed by Paul Erdős. We start there because results can be checked by machine.
| Task | How it is checked |
|---|---|
| Proofs | Compiled in Lean with Mathlib. The statement is re-checked in a separate file and axioms are inspected. |
| Computations | Re-run by our own independent implementation. |
| Literature status | Citations must exist. Two agreeing runs, then a referee that must quote the cited papers verbatim, then a human. |
Problem statements stay on erdosproblems.com. We link to them and credit the maintainers. We have no discoveries to report.
Works with
Bring whatever you have.
Planned support: Anthropic API or Claude subscription, OpenAI API or Codex / ChatGPT subscription, OpenRouter, and local models (Ollama, LM Studio, MLX). Everything runs under your own account and your own terms with each provider. All names belong to their owners, and SpareTokens is independent and not affiliated with or endorsed by any of them.
- Anthropic
- OpenAI
- OpenRouter
- Local models
Leaderboard
Ranked by verified work, never by spend.
Loading leaderboard.
Score is the sum of weights of results that passed a machine check. So far, most verified results confirm that a problem is still open or reproduce already-known values. They are pipeline checks, not discoveries. Spend, where shown, is self-reported by contributors and never affects rank.
FAQ
Questions, answered plainly.
Is this live?
Not for the public yet. The coordinator, worker and Erdős task pack work locally, and small pilots have run. There is no public worker download, and no discoveries so far.
Does my API key get sent to you?
No. The worker runs on your machine with your credentials, and the client code never sends them to the coordinator. The worker is open source, so you can check.
Can I use my subscription plan?
Maybe, and that is your call. Subscription plans (Claude, ChatGPT/Codex and others) have their own terms, which may limit automated or shared use. Everything SpareTokens does runs under your own account, so read your provider's terms and only contribute what they allow. API keys with a spend cap are the clearer route.
How much can I lose?
You set a hard budget per task and per session in the worker. It stops when it hits the cap.
How do you know results are right?
We do not trust the worker. Proofs are compiled in Lean, computations are independently re-run, and literature claims need two agreeing runs, a referee that quotes sources verbatim, and a human. A verified check is not a claim of novelty.
Will you pay contributors or rank by spend?
No. Ranking counts verified results only, and spend never affects rank.
Is it open source?
The worker is Apache-2.0 and will be public. The rest of the repository is private for now and will be opened when it is ready.
What other problems will you take on?
Other packs with checkable answers, such as computational biology or OEIS gap filling, are planned. Each needs its own verifier, and anything a machine cannot check will be labeled as human-reviewed.
Waitlist