Banks are now blocking AI agents outright, because they cannot tell an authorized one from a rogue one. UBAG gives every agent a verifiable identity at your web edge, so you can say yes to the agents you trust and put every action on a signed record.
AI agents now arrive at your site and act on it, sometimes carrying a customer's own login. Most sites can see only two choices: allow everything, or block every agent and lose the ones that help.
UBAG sits at the door of your site and routes every arrival into one of three lanes, by what it can prove, not by a guess about who it is. MCP standardized how agents talk to tools; it left this gap at the web edge, agent identity when no human is in the loop.
Pass a hard proof-of-human check and get the normal HTML site, exactly as today. Real visitors walk straight in.
Declare an identity and get a clean, structured JSON-LD version of the site built for machines, instead of scraping the HTML meant for people.
No declared identity and no pass on the human check means no entry. Undeclared automation is turned away at the gate.
An AI agent is told to book a class on a real booking site. Run it with the site undefended, then with UBAG on. The agent keeps what it is allowed to do and loses what it was never granted, like cancelling a booking it does not own. This is the live web layer, running now.
Built by Dixit Algorizmi under ubag.ai. One brain, many plugs.
Deterministic authorization for what an agent is allowed to do against your systems and MCP tools. Holds credentials, reads the whole plan, blocks what was never allowed.
ubag.ai →Verifiable agent identity, routing, and accountability at the web edge, plus the free agent-readiness scanner above. Open reference implementation.
github →