About
Give your agents the world. Keep the receipts.
Webfold is the runtime that lets AI agents read the web and documents, act in a browser, and prove what they did. One engine, written in Rust, open under Apache 2.0.
Why it exists.
An agent is only as good as what it can read and what it can do. Today that means a hosted scraper for pages, a fleet of Chrome for anything with scripts, a separate parser for documents, and a redaction tool bolted on afterwards. Each one is slow, heavy or closed, and none of them can show you afterwards what the agent actually saw.
Webfold replaces that stack with one small program. It converts a page in microseconds, runs a page's scripts in its own engine in a twenty-fourth of Chrome's memory, reads Office files, PDFs and LaTeX in process, and keeps the exact bytes it fetched so the result can be checked later. It runs on your own machine, in your own cloud, or as a hosted API.
How we build.
Measured in the open →
Every speed and accuracy figure on this site has a benchmark page with the protocol, the machine and the limits.
Proved where it counts →
Core kernels of the engine are extracted from the Rust source and proved in Lean. The proofs ship with the code.
Open licence
Apache 2.0, across the engine and the libraries under it. Use it, ship it, audit it.
Where it is going.
Reading and acting are the start. The same libraries that make Webfold fast (the memory layer, the model runtime, the admission guard and the evidence format) are the parts of a full agent runtime: one that is as accurate as the incumbents, faster, lighter, and able to account for every step it takes. That is what we are building.
The work started in June 2026. The research behind it is published by Ptyktos Research, with the papers, proofs and benchmark sources alongside.
Get in touch.
Product questions, API access and commercial licensing: [email protected]. Vulnerabilities: [email protected]. Privacy requests: [email protected].