xChk TVP | Trust ◦ Verify ◦ Proof
Explained

A human approval gate for Claude Code and other coding agents

A fail-closed hook in the agent's execution path. Routine work runs; the actions that matter wait for the right person's passkey.

Coding agents run shell commands, edit files, call APIs and send things. Most of what they do is harmless and should not be interrupted. Some of it, an outbound email, an ssh session, a payment, a change to a system setting, should not happen without a person saying yes. A human approval gate is the mechanism that tells those apart at the moment of action and gets the answer from a person before the agent can continue.

How the xChk gate works

The gate is a pre-tool hook. Claude Code, Codex, Cursor, Antigravity, GitHub Copilot CLI and Hermes each call it before a tool runs. The hook sends the proposed action to the agent's registry, which answers from the plate's grants: allow, refuse, or hold. A hold stops the agent and sends the plate owner an ask on their phone, by push, SMS or email. The owner unlocks it with their passkey, reads the exact command, and approves, declines or edits it. Only then does the agent continue, and every step is sealed as an authorization receipt.

The whole first minute, from one install command to the first approval on a phone, is drawn in this diagram, and the approval itself in this one.

Fail closed, or it is not a gate

A gate that passes when the registry is unreachable, or when it cannot parse what the agent asked, is a decoration. The xChk hook refuses in those cases. That is uncomfortable in practice, a network fault stops the agent, and it is the correct discomfort: the alternative is an agent that is unguarded exactly when something is wrong. The same rule covers the agent's own attempts to route around the hook; a command that names the gate's files or its configuration is held or refused.

The right person, not just a person

Most approval schemes ask whoever is sitting at the keyboard, which is often the agent's own operator or a second agent. xChk asks the person the grants name for that area: usually the owner, sometimes a delegate, a finance lead for spend, for example. Their identity is established with a passkey, not a chat reply, and the approval is bound to the exact action and expires. An approval cannot be reused for a different command.

What stays on the machine

The agent's work does not leave the host. What the gate sends is the proposed action and its context; what comes back is a decision. The connectors page and the connections map list exactly what is reported where.

Measured, not claimed

A gate is a security control, so it is tested as one. The bypass and detection benchmark runs before every release: 23 bypass techniques and 79 per-harness cases, with 0 silent passes on the current release. External suites have been reproduced against the gate, AgentDojo 96.2 %, AgentGovBench 47 of 48, and spikee 94 % with no false positives. The numbers, the method and the failures found along the way are on the benchmark page, and the AGSC self-assessment rates the gate against a published control catalog.

Where it fits with the rest

Patents pending. The gate's source is published with each release, signed and manifest-hashed, so a copy that is behind shows up in the owner's console.