Two in five of Rauch's open-source grants now go to agent tooling
The same list also pays the person who builds the proof-of-work gate that keeps AI crawlers off self-hosted wikis, which is the most honest thing about it.
Guillermo Rauch gave away thirty-five thousand dollars of his own money last week, and the interesting part is the spreadsheet rather than the cheque.
On 8 September he announced the second round of his personal open-source grants: $1,000 each to 35 contributors, unconditional, and explicitly not from Vercel. Nobody's rent is being paid here.
But somebody who runs one of the platforms most web software is deployed on sat down and wrote a one-line reason beside each of thirty-five projects. That list is about as direct a readout as you will get of what "foundational" means to that person in September 2026.
Two in five of the citations are about agents
So we read all thirty-five reasons and counted them (the page is a plain table, and the counting took about ten minutes). Fourteen name agents, AI, MCP or running models locally.
pgbot is "Postgres tooling for agents". fff is a "Fast and accurate file search SDK for agents". ast-grep is a "CLI for structural code exploration that agents love". slotstream and colibri are engines for running mixture-of-experts models, the kind that wake only part of themselves for each token, on a Mac and in pure C respectively. Emil Kowalski's grant is cited as "Great design eng AI skills & components", and Rauch explains the theme himself: "Skills are now a valuable form of software."
Stop on that sentence for a moment. A skill is, mechanically, a markdown file with instructions in it.
Eighteen months ago a fund like this would probably have paid for a bundler.
The rest is compilers, and one terminal emulator
The other half of the list is more familiar and, we would argue, more interesting. Kovid Goyal gets one for kitty and calibre, a terminal and an ebook manager, both more than a decade of one person's work. Lars Kappert gets one for knip, which finds the code in your repo that nobody uses. Oliver Medhurst gets one for porffor, an ahead-of-time JavaScript compiler, and Ralph Küpper for PerryTS, a TypeScript compiler written in Rust. Rauch's stated theme there is blunt. "There's no excuse for software to be slow anymore."
And then there is Xe Iaso, cited for "Bot mitigation infrastructure & novel PoW mitigations". Anubis is the proof-of-work gate that people put in front of their git forges and wikis to survive AI crawlers.
Now read the first fourteen again with that one in mind.
The same fund pays for the crawl and for the sandbags
Which is it, then? Is this money backing the agents or the defence against them?
Both, and we would not call that hypocrisy. It reads more like an unusually honest picture of the year. If you fund the tools that let software consume the open web at machine speed, then funding the person who keeps small self-hosted forges standing under that load is the consistent position rather than the contradictory one (if you self-host a wiki, you are probably behind something like Anubis already).
It is also the only grant on the list that exists because of the other fourteen.
As money a gesture, as a signal cheap at the price
The arithmetic below is ours rather than his. Thirty-five thousand dollars is roughly two months of one senior engineer in San Francisco, split across thirty-five people who between them maintain a terminal emulator, two compilers, a Postgres tool and a fleet of MCP servers. As funding it is close to a rounding error.
As a public, itemised statement of what a platform CEO believes is load-bearing, it is worth considerably more than $1,000 a name. Which is presumably why the list sits at a URL instead of arriving quietly by DM.
The obvious objection is that one person's signal is not a market. Fair. Rauch is careful to say these are personal and unaffiliated, which likely cuts both ways (no strings attached, and no continuity either). There is no institution here that has to do this again next year.
He has now done it twice. Our expectation for a third round is that more than half the citations will name agents, AI or MCP. If the share falls instead, something has changed about what this ecosystem thinks it is building, and we would want to know what.
