INTYGA lets your service verify a human’s signed approval for an exact payment, permission change or production release before it acts. These engineering notes are for developers and security teams building authorization for AI agents and sensitive actions — parameter binding, passkey-signed approvals, offline verification and tamper-evident audit trails.
Keep people in control of production changes: INTYGA binds human approval to the exact action your service executes. This post proposes how that composes with OpenShell's runtime controls and Sentry; it is not a shipped or tested integration.
An approval step can be mandatory while another credential bypasses it. Test whether every route to a high-risk action enforces the same human decision.
An agent can use credentials and sessions you already have. Map its access to production, distinguish authority from influence, and close the paths around your approval controls.
A deploy gate written as a job inside the repository it protects can be skipped by anyone who can edit and dispatch its workflows, whenever the edited workflow can still reach a deployment credential without an approval it does not control. The remedy is a separate enforcer repository the application repository cannot write to: it holds the production credential and deploys only the artifact a person approved.
Let agents prepare the work. Use INTYGA to require a human signature for the exact payment, permission change or production release before your system executes it.
INTYGA does not make agents trustworthy. It makes trusting them unnecessary — when your own service verifies a human's signed approval before it acts, and holds the only credential that can act.
Shared golden test vectors prove that implementations agree with the fixtures, not that they conform to the specification, its wire format or its rules about time. An audit of INTYGA's five receipt verifiers against the published specs found three defects the vectors missed: WebAuthn fields decoded in a form the gateway never sends, forward-dated offline approvals accepted, and a Merkle padding forgery revived by hex case-folding in the TypeScript reference. The fixes added production-format encoding, explicit evaluation times and new refusal cases.
ChainDrop began in 11 compromised npm packages and propagated into 433 more — 444 in roughly four hours, with valid build provenance on every release. To stop autonomous republishing, a separately administered release workflow must require a fresh human approval bound to the exact artifact before obtaining publishing authority. A signature check in repository-controlled code is not enough if the attacker can change or bypass it.
A keyless witness log records approvals signed by each approver's own authenticator, so INTYGA never holds a key that could make one. A log signed by the party being audited is tamper-evident, not non-repudiable; client-side signatures plus externally anchored Merkle roots close the gap. Teams can verify the approval receipts offline with the open-source @intyga/verify library.
Prompt-level safeguards are not a security boundary: high-risk MCP tool calls need human approval enforced by the service performing the action. With the INTYGA gateway, calls such as execute_sql_mutation trigger a WebAuthn challenge bound to the exact tool arguments; a person approves with Touch ID or a YubiKey, and your service verifies the receipt offline before executing it once. The agent must have no other route to the action.
Static API keys cannot answer which human authorized a specific irreversible run. Passkey-signed, parameter-bound approvals can — and your code can verify them offline before it executes.