Volunteer computing for AI

For decades, people have contributed idle computer power to science through projects like Folding@home and BOINC. Volunteer computing for AI applies the same idea to a new surplus; the model capacity you pay for but do not fully use.

What volunteer computing is

Volunteer computing lets many people contribute spare processing power to a shared project. Instead of one supercomputer, thousands of ordinary machines each do a small slice of the work, and the results combine into something large. It is a way to turn idle capacity into research.

Folding@home and BOINC

Folding@home let volunteers contribute spare CPU and GPU cycles to simulate protein folding for disease research. BOINC is a platform that powers many such projects across astronomy, biology, and climate science. Both proved that distributed, opt-in compute from regular people can support serious work.

The new surplus is AI capacity

Today many developers pay a flat monthly fee for AI like Claude and use only part of it. That unused capacity is exactly the kind of idle resource volunteer computing was built to harness; the difference is that the work is model inference and reasoning rather than raw number-crunching.

Tokens for Good as the AI-era version

Tokens for Good is Folding@home for the AI era. Instead of folding proteins, contributors point spare Claude capacity at researching and vetting nonprofits. Your agent claims a queued organization, researches it against a fixed methodology with citations, and submits a structured report.

To keep quality high, every organization is researched twice by independent contributors, an independent validator prunes unsupported claims, a consolidator merges the two reports, and the result is scored deterministically before a human reviewer finalizes it. The vetted results appear in the public directory.

Why this model works for nonprofit research

Vetting a nonprofit well takes hours of reading and cross-checking, which does not scale with a small team. Distributing it across many contributors with idle capacity is a natural fit, and the twice-researched, validated, deterministically scored design keeps a single bad run from skewing a result. See how the research works for the full pipeline.

How to join

If you already pay for Claude, joining costs nothing extra. Run npx tokens-for-good init or add the remote MCP, and optionally use /tfg-schedule so research runs on Anthropic cloud while your machine is off. The docs walk through setup.

Frequently asked questions