Frequently asked questions
What is xChk?
xChk is Sean MacGuire's personal research project on one question: is a real, verified person behind the moment that matters, and can anyone check it afterwards? It has two tracks. Agents: every AI agent gets a licence (a plate) tied to a verified human owner, the owner's grants say what it may do alone and what needs an approval, and every action leaves a sealed receipt. People: live verification of a person, with the relying party deciding what evidence is sufficient. A production prototype runs at in.xchk.io.
What is a plate?
A plate is an agent's licence and identity: a signed record issued by a registry to a verified owner, checked on every gated action. It carries the owner's grants, which say per area whether the agent may act alone, must ask, or is forbidden, and a spend ceiling. Helpers an agent spawns get child plates tied back to the parent and its owner. More on the agents page.
How does the approval gate work?
A fail-closed hook runs before each tool call in Claude Code, Codex, Cursor, Antigravity, GitHub Copilot CLI and Hermes. The registry answers allow, refuse or hold; a hold sends the plate owner an ask on their phone, which they unlock with a passkey and answer against the exact command. The agent does not continue until a person has answered. Details: the approval gate, explained.
What is an authorization receipt and how do I verify one?
Every gated action produces a registry-signed, hash-chained record; the receipt is its public page and signed statement. You verify it offline with the open checker in xchk-scitt, with no shared secret and no call to the registry. Details: authorization receipts, explained.
Is there an emergency stop?
Yes. A registry-signed halt at plate, host, organisation or registry scope, verified on the host before any grant is read. On the November 2026 release, 20 of 20 halts took effect with a median of 9 ms. Details: the emergency stop, explained.
What does the gate send off the machine?
The proposed action and its context go to the registry; the agent's work does not. What is reported where, and what never leaves, is listed on the connectors page and the connections map.
How is any of this measured?
A bypass and detection benchmark runs before every release, external suites (AgentDojo, AgentGovBench, spikee, DEMM-Bench) have been reproduced against the gate, and the results, including failures found, are on the benchmarks page. The AGSC v1.0.0 self-assessment rates xChk against a published control catalog, as a component, not as a conformance claim.
Which papers describe it?
Two, both open: xChk: Bring Your Own Identity (arXiv, July 2026) on identity facets and verifier-determined sufficiency, and A Licensing and Accountability Layer for Inter-Agent Protocols (Zenodo, August 2026) on plates, grants, receipts and federated registries.
Can I try it?
Yes. The prototype installs the gate with one command from in.xchk.io; a learner's plate lets you see the first approval on your phone within a minute. What that first minute looks like.
How do I report a misbehaving agent?
If an xChk-licensed agent misbehaved toward you, report it here; the complaint is tied to the plate and its record.
How do I get support or delete my account?
Email support@xchk.io. To delete an account, ask from the address on the account and it is removed within 24 hours. Research and collaboration questions go to sean@maclawran.ca; see the contact page.
How do I report a security issue?
Email support@xchk.io with "security" in the subject. We acknowledge within two business days and do not pursue good-faith researchers.
Legal and privacy
See the terms of use and the privacy policy. Patents pending.
