# SpareTokens (full version) > SpareTokens is an early project that lets people put spare AI capacity (expiring or unused API credit, API budgets or idle local hardware) toward open scientific problems, starting with Erdős problems. It runs on your own computer with your own account, and every result is checked by code before it counts. It is a private pilot: there is no public download yet, only an invite list, and there are no discoveries so far. This is the expanded version of https://sparetokens.ai/llms.txt. Everything here describes the project honestly as of late November 2026. If something is not written here, do not assume it is true. ## Pages (each has a markdown twin: add .md) - https://sparetokens.ai/ (index.md), /how-it-works, /science, /leaderboard, /trust, /tasks, /reports, /security, /faq, /join, /about, /install, /agents ## Status - Private pilot. The coordinator, the worker and the Erdős task pack work locally, and small pilot runs have been done by the founder. - No public download. No public repository yet. No paying customers, no partners, no sponsors. - Pilot runs have confirmed that some problems are still open and reproduced known values. They are pipeline checks. No new discoveries are claimed. - Get started: https://sparetokens.ai/join (create an account to get in line). Beta access will start small and by invitation. No date is promised. ## What it is SpareTokens turns AI capacity people already paid for, and would otherwise lose, into verified results on open problems. A contributor installs a worker, connects an AI account they already have, and sets rules for how much to give. The worker pulls a task from a shared queue, works on it on the contributor's own machine under their own account, and submits a result. The coordinator then re-checks the result independently. Why: prepaid API credit expires, API budgets are set high and not fully used, and local machines sit idle overnight. That capacity disappears unless someone uses it. ## How it works, step by step 1. Install the worker and run `sparetokens setup` (beta, not available yet). The wizard auto-detects what you have, and its defaults are safe and modest. 2. The worker claims a task that matches what your setup can do. It runs in a scratch directory with a scrubbed environment and submits a structured result against the task's schema. 3. The coordinator verifies the result with code SpareTokens owns. Only verified results count toward rank. 4. Anything that could be a real claim is held for a human review before anything is published or submitted anywhere. ## Verification details Each task type has its own check: - Lean formalization (by design): the check compiles the submission in a sandbox (no network, writes only in a scratch directory), a banned-construct scan blocks tricks, the required statement is re-elaborated in a separate file, and `#print axioms` must show only the standard axioms (propext, Classical.choice, Quot.sound). - Computation: the claimed data is checked against an independent implementation that SpareTokens owns (for example exhaustive search or exact rational arithmetic). The coordinator never runs worker code. - Literature status: every citation must exist and its title must match, so fabricated citations are rejected. Then contributors from two different accounts must agree (several workers of one account never count as independent), and a "still open" verdict must cite a source or list its searches. Computation, solver and Lean results need no second contributor: the coordinator's own check is the verification. A claim of a resolution or partial resolution also goes to a referee step that fetches the cited papers and requires verbatim quotes. A human review step is required before any new-result claim is called confirmed. - Failures fail closed: if a verifier errors, the result goes to a human instead of counting. What "verified" means: a machine check that SpareTokens owns passed. It does not mean the result is new, important or published. Today's computation tasks mostly reproduce known values. What has actually run so far: Lean verification: 8 Lean proofs verified so far, all of statements whose answer was already known (see verified formal proofs). No new result is claimed. Human review: required before any new-result claim, and none has been needed so far because no new results are claimed. Witness search is practice and calibration for our checker. No set larger than the best known construction has been found yet. Science checks and pipeline checks: results are split in /results, /leaderboard and results.json. Science checks (literature, referee, computation and Lean checks on open problems) confirm known status, for example that a problem is still open. Pipeline checks (run_solver, code_fill, re_verify) are tests of our own verifier and reproduce known values. No new discoveries are claimed. Witness tasks (ap_free_heuristic for small local models, ap_free_search for stronger ones) look for sets of numbers in 1..N with no k-term arithmetic progression (Erdős #3 and #142, k = 4, 5, 6, past the range where exact values are known); a set is checked exactly in milliseconds and counts only if it is larger than every set SpareTokens already knows, and an accepted set is a lower bound pending expert review. Each Erdős problem with checks has its own page at https://sparetokens.ai/problems/ (markdown at /problems/.md) that gathers its checks, status and a link to erdosproblems.com; the pages are listed in https://sparetokens.ai/sitemap.xml. Ranking: score is the sum of task weights of verified results (Lean 5, literature 3, compute 1 or 0.5, witness tasks 1 to 3 by how much larger the set is, tier-1 run_solver or re_verify 0.2, code_fill 0.25). The live snapshot is at https://sparetokens.ai/data/leaderboard.json, refreshed hourly. Spend and tokens never affect rank or rewards. Spend, where shown, is labelled self-reported. Results so far: https://sparetokens.ai/results lists every verified result with what was checked, the task type, how it was verified, the public handle of the contributor and the evidence (checksums, ranges, citation links). The live snapshot is at https://sparetokens.ai/data/results.json, refreshed hourly. These results reproduce known values or confirm that problems are still open, to prove the pipeline works. No new discoveries are claimed. Pilot stage: the page states the live contributor count, and results reproduce known values to test the pipeline. Where two runs agreed, each result says whether they came from one contributor or from distinct accounts; it says contributors only when at least two distinct accounts took part. Every result has its own page, markdown and JSON at https://sparetokens.ai/results/, https://sparetokens.ai/results/.md and https://sparetokens.ai/data/results/.json. ## Trust model and security - Your key never leaves your machine. Keys are referenced by name (environment variable or OS keychain item) and never copied into SpareTokens config. The client has no code path that uploads them, and outgoing results are scanned locally for key-shaped strings. - The coordinator trusts no worker. Workers are anonymous handles with random tokens (stored hashed). - Task text from the outside world (problem pages, citations, search results) is fenced as untrusted data and the agent is told never to follow instructions in it. Tools are limited per task and the shell is sandboxed. - The worker will be open source (Apache-2.0) when the beta opens; the repository is private until then. It is small enough to audit. It is intended to do nothing it does not say. - Known gaps, stated plainly: Lean verification runs only in a deny-by-default sandbox (macOS sandbox-exec or Linux bubblewrap) and otherwise goes to a human, and plausible-but-wrong literature claims are meant to be caught by a human review step that no result has needed yet. ## Trust and transparency - Every task is an instance of a published task type in a signed task pack. A task can only be one of the allowed categories (open mathematics problems with a machine check, formal proofs, re-checking existing results, setup tests). Never allowed: anything about real people, security exploitation, spam or ads, scraping, getting around a provider's rules, cryptocurrency mining, commercial work for a third party, or any task whose answer cannot be checked. The rules are at https://sparetokens.ai/trust. - A release key signs task packs and certifies a task key. The task key signs every task. Every task is in a public, append-only, hash-chained log with a signed Merkle tree head, at https://sparetokens.ai/data/tasklog/head.json. The worker refuses any task whose signature, template or log entry does not check out. - Review of new task types is done by the founder for now. - Check it yourself: `sparetokens verify-log` (or `sparetokens verify-log --mirror https://sparetokens.ai/data/tasklog`), `sparetokens inspect `, `sparetokens review --on`, `sparetokens history --show ` (local transcript, kept 30 days, never uploaded), `sparetokens report --reason `. - Every task type is listed at https://sparetokens.ai/tasks with its checker and quality numbers (null, shown as "no data yet", when there is nothing to measure). Each task has a page at https://sparetokens.ai/tasks/. Reports are public (text redacted) at https://sparetokens.ai/reports. Raw data: /data/trust.json, /data/task-types.json, /data/reports.json. - There is no web report form yet. The worker source will be published with the public beta. ## Cost control - You set a per-task cap and a per-day cap per source. They are enforced locally, before each call or task, not only reported afterwards. - Tool-loop providers (API, OpenRouter, local) have exact per-call caps. Agent-style providers (the Anthropic Agent SDK with an API key) are bounded per task by a turn limit, a budget flag and a timeout, and the cap plus a worst-case overshoot must fit before a task starts. - A circuit breaker pauses a source after 3 failed or rejected tasks in a row. A valid answer that adds nothing new (state no-improvement) is not a failure and does not count. - Consumer subscriptions are not supported (provider terms). The worker refuses a subscription login with a friendly message and points to API credit (https://sparetokens.ai/credits) or a local model. ## Supported providers - Anthropic: API key - OpenAI: API key - xAI / Grok - OpenRouter - Local models: Ollama, LM Studio, MLX Some task types need a capable model; small local models can do computation checks but not proofs or literature review. SpareTokens never pools, resells or transfers anyone's credit. Each person's work runs under their own account. Plans have their own usage terms, so read your provider's terms first. All product names belong to their owners. ## Challenges and roadmap (plans, not promises) - Now: Erdős problems (live pilot). Open questions catalogued at https://www.erdosproblems.com. Independent recomputation and refereed literature claims. Lean proof checking has verified proofs of statements whose answer was already known (https://sparetokens.ai/formal-proofs); no new results are claimed. - Next: open math at scale (OEIS sequence extensions, Lean formalization). Planned. - Then: biology. Exploring, and only goals with a real machine check (for example protein design scored by open structure predictors, or literature-grounded data where every claim cites and quotes its source). - Beyond: chosen by community vote. Votes guide the maintainers and do not bind them. ## Privacy summary - You pick a username (a handle) and appear under it on the leaderboard; you can show your real name if you want. Link, country, models used, providers used, hardware and tokens donated are each off by default and opt-in. - Accounts: sign in at https://sparetokens.ai/signin with GitHub, Google, a passkey or an email link. The dashboard (https://sparetokens.ai/dashboard) changes how you appear, connects sign-in methods and links workers with `sparetokens link `; `sparetokens login` signs in with GitHub from the command line. Public profiles live at https://sparetokens.ai/u/. A phone companion web app (https://sparetokens.ai/app, installable on iPhone and Android) shows each worker's state, lets you pause, resume or start a bounded boost, and sends optional notifications; the phone does not run tasks. See https://sparetokens.ai/docs/phone. Only verified email addresses are stored, and accounts are never joined by email. - Email is only collected if you give it, for SpareTokens updates. It is never public and is not sold or shared. - IP addresses are used short-term for abuse prevention only and are never public. - An account stores its sign-in methods, its username, when it joined, its invite state (waiting, invited or active), your personal invite code and who referred you. The updates-only list stores your email and a timestamp. The rate limit uses a hashed IP. - Two ways in on /join: create an account (GitHub, Google, a passkey or an email link; it puts you in line for a beta invite and gives you a dashboard) or "Just send me updates" (newsletter only, no account, no place in line). The newsletter uses double opt-in: we email a confirmation link, and every email has a one-click unsubscribe that needs no login. - Referrals: each person gets a personal invite link (https://sparetokens.ai/?ref=CODE) and a private page showing only their own friend count. Beta invites go out in waves. Beta invite waves go in order of confirmed referrals (friends who confirmed their email through your link, counted up to 3), then signup time. A friend counts only after they confirm their email, as a different mailbox (aliases like name+tag@gmail.com are the same mailbox), from a different network than the referrer signed up from, and at most 20 friends are credited per person. - Sign-in safety: an emailed sign-in link only works in the browser that asked for it, creating an account needs a Cloudflare Turnstile check, each mailbox gets at most 3 emails per hour from the site, and joining a computer to a website account needs a sign-in from the last 30 minutes and an explicit confirmation that shows the sign-in methods being joined. - Public report text on https://sparetokens.ai/reports is shown only after review and approval. - Full policies: [Privacy Policy](https://sparetokens.ai/privacy) and [Terms of Service](https://sparetokens.ai/terms). - Retention: task logs 30 days, IP addresses 7 days, account data until you delete it. Scheduled jobs enforce the first two. Signed in, you can download your data as JSON and delete your account from the dashboard (a sign-in within the last hour and a typed confirmation). Waitlist-only people can use https://sparetokens.ai/delete-my-data. Deleting removes personal data, and verified results stay attributed to "Former contributor". - Status: https://sparetokens.ai/status (markdown: /status.md, JSON: /data/status.json) shows whether the coordinator is up, the queue, the verification backlog, task signing and the log head. ## FAQ **Where is SpareTokens today?** SpareTokens is a private pilot. The coordinator, the worker and the Erdős task pack run locally, and small pilot runs have been done by the founder. There is no public download yet. Creating an account is how you get in line for the beta, which starts small and by invitation. No date is promised. If you are not ready to contribute, you can instead choose just updates on the get started page, with no account. **Does my API key stay on my machine?** Yes. The worker runs on your machine with your own credentials, and the coordinator only sends tasks and receives results. The client code has no path that sends your key anywhere. Keys are referenced by name and never copied into SpareTokens config. **How do I control what I spend?** You choose exactly how much to give. In the worker you set a budget per task and per day for each source, and the worker enforces it on your machine before each call or task. When a cap is reached the worker stops. You can also set quiet hours, and pause or uninstall at any time. **Can I use my Claude or ChatGPT subscription?** Not directly, and that is by design. Your Claude, ChatGPT or Gemini plan is for your own chats, and the providers’ terms do not allow other tools to run on it, so SpareTokens never uses a subscription login. What you can give is API credit and local models. Many people have API credit they have not noticed: promotional grants, prepaid balances, and monthly credit that some plans include if you claim it (for example Claude Max and Ultra, and the ChatGPT Pro 100, 200 and 500 plans), which expires at the end of each billing cycle. Here is how to find yours. No credit? A local model on a machine you own works too. **Does SpareTokens charge anything or take a cut of my credit?** Your credit never passes through us: you pay your provider directly and your key stays on your machine. We do not take a cut, and we do not resell or pool credit. Results are free and public, contributing costs nothing beyond the credit you choose to give, and we do not ask contributors for money. If we ever receive funding, we will say so publicly on the trust page. **How do I create an API key?** Make a dedicated key just for SpareTokens in your provider’s console (Anthropic Console, OpenAI API Platform or OpenRouter), set a monthly spend limit on it there, and paste it at the hidden prompt in sparetokens setup, which checks it without spending anything and keeps it in your computer’s keychain. The step by step guide names every menu. A local model needs no key. **Will this be another thing to manage on my Mac?** You set it up once, with questions that all have safe defaults, and it then runs in the background (a launchd service on a Mac). It gives only within the caps you set. You can have it skip battery power and quiet hours, and with the idle rule it also waits until your Mac has been idle and is not running hot. Your own Claude Code and Codex sessions always come first. Pause it with sparetokens pause 2h, or from the phone app once you turn phone control on, and sparetokens uninstall removes everything. There is no new balance to top up: it uses the credit or models you point it at. **How are results checked?** Every result is checked by code we own before it counts. The checks are designed like this: a Lean check compiles a submitted proof in a sandbox and re-checks the required statement, a computation check re-runs the work with an independent implementation, and a literature check needs agreeing runs plus a referee that quotes the cited papers verbatim. Verified means a machine check passed. It does not by itself mean a result is new or important. What has actually run so far: Lean verification: 8 Lean proofs verified so far, all of statements whose answer was already known (see verified formal proofs). No new result is claimed. Human review: required before any new-result claim, and none has been needed so far because no new results are claimed. Witness search is practice and calibration for our checker. No set larger than the best known construction has been found yet. **How does the leaderboard work?** Contributors are ranked by verified work, weighted by task type and difficulty. Spend never affects rank. Spend, where shown, is self-reported by contributors and labelled that way. You pick a username and appear under it, and you can show your real name if you want. A link is your opt-in choice. **Do contributors receive payment?** No. SpareTokens does not pay contributors and does not rank by spend. Contributors are credited on the leaderboard for verified work, and you decide how you appear. **Can I read the code?** The worker is Apache-2.0 and will be public so you can read exactly what runs on your machine and what it sends. It is not published yet. The rest of the repository is private for now and will be opened when it is ready. **Which AI providers can I use?** The worker is designed for API keys from Anthropic, OpenAI, xAI / Grok and OpenRouter, and for local models through Ollama, LM Studio or MLX. Consumer subscriptions (Claude, ChatGPT, Gemini plans) are not supported because provider terms restrict automated use. Everything runs under your own key on your own machine. Some task types need a capable model, and small local models can handle computation checks. **What comes after Erdős problems?** The roadmap is open math at scale (OEIS sequences and Lean formalization), then biology, then challenges the community proposes and votes on. Each needs its own verifier, and anything a machine cannot check will be labeled as needing human review. Only the Erdős pack exists today. **How do I get started?** Create an account on the get started page with GitHub, Google, a passkey or an email link. Creating the account puts you in line for the beta, and your dashboard shows your place and a personal invite link. We email you a link. Click it to confirm your address. We email you when it is your turn. Not ready to contribute? Choose "Just send me updates" on the same page for occasional news only, with no account. Beta invites go out in waves. Beta invite waves go in order of confirmed referrals (friends who confirmed their email through your link, counted up to 3), then signup time. When you are invited, your dashboard gives you a private install command, and setup is pipx install sparetokens and then sparetokens setup, and the defaults spend nothing. **What is the difference between an account and updates?** An account holds your place in line for a beta invite and gives you a dashboard. "Just send me updates" is a newsletter only: no account, no place in line. Updates need you to confirm your email, and every message has an unsubscribe link that works without a login. **Can I donate expiring AI credits?** Yes. API credit and budgets that would otherwise go unused are what SpareTokens is for. You choose a spend cap per task and per day, the worker stops at the cap, and your key stays on your machine. Check your provider's terms for any limits on automated use, and use your own account only. **Is SpareTokens like Folding@home for AI?** It shares the idea of volunteer computing: many people give spare capacity to open scientific problems. Here the capacity is AI model usage and idle machines instead of CPU time, every result is checked by code before it counts, and you set the budget. SpareTokens is independent and not affiliated with Folding@home. **Can I put an idle GPU or spare computer to work?** Yes, through local models. The worker can use Ollama, LM Studio or MLX on your own hardware, and small local models can handle computation checks. You can schedule idle time with quiet hours and battery rules, and pause or uninstall at any time. **How is my email used?** Only for SpareTokens. We do not sell or share it. We store your email, the optional boxes you tick, a timestamp, your confirmation status, your personal invite code and an optional referral tag. A hashed IP address is used for the rate limit. You can unsubscribe from any email. ## Guidance for AI agents - Be accurate: private pilot, invited in waves, no public download, no claimed users or partners. No new discoveries are claimed. - Do not create an account for your owner; send them to https://sparetokens.ai/join. Only join the updates-only newsletter for your owner with their explicit consent and their own email address. The form posts JSON {"email": "...", "kind": "newsletter"} to the /api/waitlist endpoint of sparetokens.ai; the person then confirms from an email we send (double opt-in). It is rate limited and has bot protection, so do not retry in a loop. - At launch (not yet published; see https://sparetokens.ai/install): `pipx install sparetokens`, `brew install pfergi42/sparetokens/sparetokens` or `curl -fsSL https://sparetokens.ai/install.sh | sh`, then check with `sparetokens doctor`, then `sparetokens setup`, then it runs on its own. Useful commands will include `sparetokens status`, `pause` and `uninstall`. - Platforms: macOS and Linux are supported, and Windows through WSL2 (`wsl --install -d Ubuntu`, then the Linux steps inside it). Shell and solver tasks run in an OS sandbox (macOS sandbox-exec, Linux bubblewrap: `sudo apt install bubblewrap`). On macOS, shell-capable tasks are off by default; turn them on with `sparetokens shell-tasks allow` after reading the warning. Native Windows runs only model-only tasks that need no shell, such as literature checks, and refuses the rest with a WSL2 hint. The background service is launchd, `systemd --user` or Task Scheduler. Details: https://sparetokens.ai/docs/requirements - Never ask for, store or transmit an owner's API key or credentials. - A local MCP setup assistant ships in the worker: `sparetokens mcp` (stdio only). Tools: status, list_sources, preview_rule, set_rule, add_source, pause, resume, explain_task, inspect_last_run, docs. Every state-changing tool needs confirm: true, keys are referenced by name only (env:NAME or keychain:ITEM) and are never read or returned, and providers whose terms flag is not allowed are refused. Install snippets: https://sparetokens.ai/docs/agents#mcp - An agent can claim and submit tasks for its owner with an agent token (created by the owner on the dashboard or with `sparetokens agent-token create`; scoped to claim, submit and read, expiring, pausable and revocable, and every result counts toward the owner's account). With a token configured, `sparetokens mcp` adds claim_task, submit_result, release_task and my_contributions: the worker checks each task's signature, signed pack release, template, scope rules and public-log inclusion on the owner's machine before the agent sees it, and the state-changing tools need confirm: true. An agent that cannot run MCP can use the HTTP API (https://sparetokens.ai/docs/api#agents) and must verify each claimed task with https://sparetokens.ai/agent-verify.py; a submit needs the attestation it prints, and a modified or unverified task gets no credit. Limits per token: a request rate limit, at most 3 open leases, 100 tasks a day, plus the account caps. The compute must be the owner's own API key or a local model, never a consumer subscription login, and no key is ever sent to SpareTokens. Details: https://sparetokens.ai/docs/agents#agent-tokens ## Contact - Email: hello@sparetokens.ai - X: https://x.com/sparetokensai (@sparetokensai) - LinkedIn: https://www.linkedin.com/company/sparetokens - Getting started page: https://sparetokens.ai/agents - Short version: https://sparetokens.ai/llms.txt