My spare Claude tokens research charities while I sleep
The strange part is that I wake up, see the notification, and feel like I did a small good thing, even though I'm not entirely sure I did anything at all.
Gardeners have a warning about August. Lock your car during zucchini season, they say, or you’ll come back from church to find a back seat full of squash. One zucchini plant produces more than any family can reasonably eat, so for a few weeks every summer, an entire subculture of otherwise law-abiding Americans turns into low-grade produce smugglers, leaving grocery bags on porches, ringing doorbells and speed-walking away.
That might be the purest picture of generosity there is. Nobody organizes a zucchini drive or sets up a donation portal for surplus squash. The abundance just finds its way to whoever happens to be nearby, because handing it over is easier than watching it rot on the counter.
When sharing is frictionless, we share, every time. Add a form, an account, a phone call, a decision between forty organizations you know nothing about, and the impulse quietly dies in the driveway.
My leftovers rot at midnight
I pay for a Claude Max subscription. Every cycle, the meter resets whether I touched it or not, and most cycles I don’t get anywhere near the cap. There’s real capacity sitting in that account, capacity I already paid for, and when the window closes it evaporates like produce I grew and never picked.
That kind of waste is easy to ignore, the same way the empty seat in your car on the morning commute is easy to ignore. The cap always felt like a ceiling to stay under, and I’d spent exactly zero minutes thinking about the space between the ceiling and my head.
Then my friend Hank Taylor showed me what he’d been building, and now I can’t stop seeing the zucchini.
Hank’s porch
Hank has a friend, Todd Manwaring, a philanthropist who has made it his life’s mission to figure out which charities actually do what they say they do, the ones that can produce evidence, real measured outcomes, proof that the work changes lives, rather than just the best photos and the saddest commercials. That vetting work lives in a social impact directory run by Fierce Philanthropy, and it currently holds 1,417 organizations.
1,078 of them are still waiting for someone to do the research. And the 1,417 is just the seed list, because Hank’s goal is one million organizations in that directory.
The obvious move at that scale is the one every institution makes. Raise money, hire researchers, buy compute, burn through a grant. Hank went the other way. He looked at all of us sitting on subscription allowances we never max out and built Tokens for Good, an MCP server that lets your AI agent pick up one of those 1,078 organizations and do the research itself, on your account, with the tokens you were never going to use anyway.
He told me the inspiration came from Folding@home, the project that borrowed millions of idle PlayStations and screensavers to fold proteins for disease research, and Tokens for Good runs on that same instinct, except the leftover it borrows is your subscription allowance instead of your CPU.
What the robots actually do
When my agent claims an organization, it works through Todd’s social impact evaluation framework. It digs through the org’s annual reports and impact pages, pulls their Form 990s from ProPublica’s Nonprofit Explorer, checks GiveWell and Charity Navigator, and searches the academic literature for randomized controlled trials and independent evaluations. Then it fills out an evidence table where every single row has to be a verbatim quote from a real page at a real URL. Paraphrasing gets flagged. Made-up citations get the whole report rejected.
My favorite detail is that the agent doesn’t score the charity. The score is computed by plain old deterministic code from whatever evidence made it into the table. If an organization has no proof of outcomes, the rows sit empty, and empty rows are considered the honest answer. Two independent reports get produced for every org and consolidated, and a human reviews everything before any of it is finalized. The AI does the legwork while the code and the humans do the judging.
The charities with receipts rise. The ones without receipts get something arguably more valuable, a precise map of exactly what evidence they’d need to gather to earn trust. Either way, the person deciding where to give their money gets to see through the marketing.
Set it and forget it
My setup took one command, npx tokens-for-good init. I had to update Claude Code first, which I’d been putting off anyway, so the whole experience mostly served as an intervention for my update hygiene. I picked the daily cadence, closed the terminal, and forgot the project existed.
Now every morning my phone buzzes with a little notification that research got done overnight, and half the time it’s about an organization I had no idea existed. That’s turned into my favorite side effect of the whole thing. So far my agent has drafted reports on the Green Belt Movement, the reforestation effort Wangari Maathai won a Nobel Prize for, along with Conservation International, Ecotrust, Forest Trends, and GiveDirectly’s historic programs. Overnight it claimed Agri Evolve, a B Corp working with fifteen thousand smallholder coffee farmers in Uganda’s Rwenzori Mountains, which is a sentence I could not have produced last week for any amount of money. The tokens do the research, and I get a slow drip of an education in who’s actually out there doing the work.
Hank, for the record, has pushed 336 organizations through that same pipeline, and that means exactly what it does on my machine: his agents did the reading. That’s the whole point of building it. A person vetting charities by hand might get through a dozen a year, and one guy with a pipeline has already cleared nearly a quarter of the current queue. Against a goal of a million, even Hank’s pace needs a few thousand more porches. I am number three on the leaderboard with one, which tells you both how early this is and how little effort my position required.
The strange part is that I wake up, see the notification, and feel like I did a small good thing, even though I’m not entirely sure I did anything at all. My laptop did it. The best I can honestly claim is that I didn’t stop it. But the feeling is real, and I’ve started to think the feeling is the entire point. Once the friction hit zero, the giving started happening on its own, like squash appearing on a porch. Apparently the generosity was sitting in the account the whole time, right next to the tokens.
The pattern is bigger than the project
Tokens for Good is one instance of a shape that barely existed a year ago: take a real problem that needs enormous amounts of careful grunt work, wrap it in an MCP server with a rigid methodology and human review, and let strangers point their idle agents at it.
Once you see that shape, you start seeing openings everywhere. Somebody’s leftover tokens could be triaging issues on the open-source projects they depend on. They could be transcribing handwritten archives for a historical society, or running accessibility audits on small-town government websites that will never have a budget for one. All that’s missing, in most of these cases, is somebody to build the porch where the zucchini gets dropped off.
So that’s my actual question for you, beyond this one project. Where’s the friction in your own giving? What’s the thing you’d contribute to constantly if contributing took one command and zero further thought? Because the tooling to delete that friction is sitting on your machine right now.
Fair warning, you may not have a spare token to your name this week, what with the Fable deadline tomorrow and every developer on earth currently feeding it their entire backlog like it’s the last helicopter out. That’s fine. After Fable goes API-only, your meter will reset, the leftovers will pile back up, and they’ll rot at midnight like they always do.
Or you could run npx tokens-for-good init, go to sleep, and wake up to a note that your computer spent the night checking whether a charity’s numbers hold up.
Somewhere out there, a porch is waiting for your zucchini.
If any of this resonates, come say hi: @davekiss