SpareTokens for teams
Put the AI budget your company or lab already has toward verified open science. Your keys stay on your machines, and every result is checked before it counts.
From your budget to the public log
How a capped pilot works, step by step.
Your cap
Set a per-task and per-day spend cap for each source. The worker enforces it on your machine.
Your machines
Workers run on computers you control, with your own keys.
Signed task
Tasks come from a public, signed list, and the worker checks each one.
Verifier
Code we own re-checks every result before it counts.
Public log
Tasks are written to an append-only public log you can inspect.
Where things stand. SpareTokens is a private pilot with no customers yet. There is no public worker download, and no discoveries have been claimed. This page says what a pilot looks like today and which team features are planned and not built.
Why teams use SpareTokens
Spend the AI budget you already have
Prepaid credit expires and budgets are set high and not fully used. SpareTokens points the unused part at open problems, under a cap you choose.
Verified open science
Every result is re-checked by code we own before it counts. The first problems are Erdős problems. We reproduce known values and confirm known status today.
Something you can show
A team leaderboard and an automated impact report are planned and not built yet. For a pilot we expect to write a short summary by hand, from the public log and your members’ consent.
How it is safe
- Keys never leave your machines
- The worker runs on your computer with your own key, read by name from the environment. SpareTokens sends tasks and receives results. No SpareTokens tool or endpoint accepts a key, and results are scanned locally for key-shaped strings before they leave.
- Hard caps
- A per-task and per-day cap, plus any ceiling you set, is checked on the machine before each task. When a cap is reached the worker stops. A source pauses after three failures in a row.
- Signed tasks and a public log
- Every task is signed and written to a public, append-only log. The worker verifies the signature, the pack release, the template and the log before it runs anything, and refuses on any failure. Your reviewer can read a task in plain words first with
sparetokens inspect. - Narrow by design
- A task can only be an instance of a small number of published task types, each with a fixed prompt, a fixed tool set, and a machine check. Shell work runs in an operating system sandbox with no network.
- Provider terms respected
- Only API keys and local models are used. Consumer subscriptions are not supported. We recommend a dedicated project key with its own spend limit. See project and service-account keys.
The details are on the security page, the trust page and the security summary for reviewers.
How a pilot works
You email us
Your organization, who owns the pilot, your providers and the monthly spend you would cap.
We agree scope and caps
One to five machines or keys, the caps, the task types and the hours. Your security reviewer gets the summary first.
You set up the first machine
A dedicated project key, the worker, a giving rule and a preview. About 30 minutes, following the team guide.
You see a first verified result
Then you decide whether to roll out to more machines, pause or stop.
We write it up with you
A short summary: tasks run, results verified, reported spend and anything that went wrong. Nothing is published without asking you.
What we need from you
- One person who owns the pilot, and one reviewer for security.
- A project or service-account key with a spend limit, and a cap you are comfortable with.
- One to five machines, or local models on hardware you own.
- Your consent for how people appear. The default is a username, never a real name.
What you get
- A verified result from your own team, and a record of what ran and what it cost.
- Access to the public task log and the signed task, so your reviewers can check the claims themselves.
- A written summary at the end, and a direct line to us.
- Planned, not built yet: a team page, a team leaderboard, organization controls and an automated impact report.
Pricing: There is no price list yet. A pilot costs you the API spend you cap, and we agree anything else in writing before it applies.
Questions from procurement and security reviewers
Data handling: what data do you receive from us?
The results of tasks, the small files a task produces, and usage numbers the worker reports about itself (tokens and spend, labelled self-reported). For people who sign in, a username and the sign-in identifier, and a verified email address. Never your API keys, and never your local transcripts, which stay on the machine for 30 days. Full list: privacy policy.
Where is the data stored, and for how long?
Waitlist records are kept in Cloudflare KV, and the coordinator keeps results and accounts in its own database. Retention periods per category are in the privacy policy. You can ask for a copy or for deletion by writing to hello@sparetokens.ai.
Do you use our data to train models?
No. The privacy policy says we do not use your data to train AI models. We do not sell personal information.
Compliance: which certifications do you hold?
None. SpareTokens is a private pilot run by one person, and we do not hold SOC 2, ISO 27001 or similar certifications. We say so plainly so you can decide. What we have instead is a design you can check: signed tasks, a public log, a sandbox, caps on your own machines, and a published list of known limits.
Has the security been reviewed by outsiders?
A separate reviewer attacked the design and code before the trust layer shipped, and the fixes are listed on the trust page. An external security review of the worker sandbox is planned before the public beta and has not happened yet.
Terms: what applies to a pilot?
The terms of service apply. For automated use of your API key, your provider’s terms apply too, so read them and your own policy first. Consumer subscriptions are not supported. We do not yet have a data processing agreement template, and we will discuss one with you for a pilot.
Can we run it inside our network?
The worker only needs to reach your provider and the SpareTokens coordinator over HTTPS. Network hosts a task may use are limited to none, except for literature tasks that need web access. Ask us if your network policy needs a specific allow-list.
Can we stop at any time?
Yes. sparetokens pause stops giving at once, sparetokens uninstall removes the service and local data, and revoking the project key at your provider ends it from your side.
Who builds and runs this?
SpareTokens is built by Paul Ferguson. See the about page. Security reports go to hello@sparetokens.ai, and the standard file is at /.well-known/security.txt.
Talk to us
Tell us about your team and what you would like to try. We reply with a plain plan and the answers your reviewer needs.