Fintech
A reconciliation agent queues a $52,400 vendor wire. Two of three finance approvers sign it on their phones before a cent moves — and the receipt is what you show the auditor.
HumanAuth is human-in-the-loop for AI agents. It holds an agent's risky actions — the wire, the payout, the merge — until the right people approve them: a biometric yes on their own phone, a cryptographically signed receipt for your records. When an auditor asks who approved this, you have the answer.
Wire $52,400 to vendor-payouts
Your agents hold credentials, run code, and reach your customer data.
Nothing is asking permission.
Authenticate the CLI, drop the MCP server into your agent config, and install an approver on your phone or Mac. Your agent can now ask, collect, and inform.
npx @humanauth/cli login // Add to your agent's MCP config — that's it
{
"mcpServers": {
"humanauth": {
"command": "npx",
"args": ["@humanauth/mcp"],
"env": { "HUMANAUTH_API_KEY": "ha_live_..." }
}
}
}
// Three new tools: human_authorize, human_collect, human_informYour agent calls auth.authorize() with the action it wants to perform. The platform fans the request out to your devices. The ask is async — your agent gets a receipt to verify once a human approves, seconds or hours later.
A push notification with a plain-language summary of what the agent wants, and an anomaly flag if it is acting outside its pattern.
One tap — Face ID or Touch ID. The device signs the response with its own key, generated on the device and never sent anywhere. We countersign.
Drop the receipt into the service that executes the action. It verifies offline against public keys. Nothing runs that a human didn't sign.
2FA is authentication — it proves who you are, once, at login. This is authorization: a human decides whether this exact action may happen, and your systems get the signed proof. An agent isn't an impostor; it holds valid credentials. That is why login security cannot answer this question. More in the FAQ →
Pause-and-ping tools stop your agent until someone clicks a button, then trust the agent to behave. The approval is a boolean in a database and nothing downstream can check it. Here the approval is a dual-signed receipt your resource server demands before executing — verified offline, bound to the exact parameters the human saw, burned after one use.
An agent acts on someone's behalf, and that person's intent is a different consent than the approver's sign-off. HumanAuth captures both: the principal confirms they asked for this on their own device before approvers ever see the request, and the receipt carries the full chain. A request nobody asked for dies at "This wasn't me", one tap, before it can farm a single approval. How on-behalf-of works →
No receipt, no action. One receipt, one action.
Every receipt carries two signatures over the same bytes: the platform's, and one from a key that exists only on the approver's device. Compromise our servers and you still can't mint "a human approved this." Verification runs offline, against public keys.
The receipt is cryptographically tied to the parameters the human saw. A parameter swap is rejected.
Composite jti plus idempotency key. The same key retries safely; a different key is an attack.
EdDSA and JWKS. Zero round-trips in your hot path. Fast at the edge.
Drop-in MCP server. Zero code changes.
Human-in-the-loop authorization tool.
Gate critical crew actions with biometrics.
Authorization checkpoints in multi-agent flows.
Drop-in authorization for native agents.
Community node: pause a workflow for a human yes.
TypeScript SDK — or the documented HTTP API from any language.
No login walls, no inscrutable JSON. Your approvers see what the agent wants in human language. Anomalies are flagged when a request deviates from an agent's pattern. Tap to approve, biometrics confirm, receipt logged.
A reconciliation agent queues a $52,400 vendor wire. Two of three finance approvers sign it on their phones before a cent moves — and the receipt is what you show the auditor.
An intake agent wants to release records to a specialist. A named clinician approves with a biometric — so every movement of patient data traces to a person, not a prompt.
A drafting agent is ready to send the engagement letter or file with the court. The supervising partner's approval is on the record before anything leaves the firm.
A claims agent recommends an $18,000 payout. The adjuster approves routine claims solo; exceptions route to a supervisor quorum — every decision reasoned and receipted.
A coding agent wants to merge to main, run a destructive migration, or promote to prod. On-call approves from the menu bar; the gate itself is part of your test evidence.
An ops agent onboards a vendor and raises a PO above threshold. Budget owners approve in seconds from wherever they are — no ticket queue, no rubber stamp.
A support agent resolves the ticket by issuing a refund, crediting the account, or pulling up a customer's record. The agent drafts; a human on the desk approves the ones that move money or touch personal data, in the same breath as the reply.
The tempting version takes an afternoon: post to Slack, wait for the thumbs-up, flip a flag in the database. It works — right up until an auditor asks who approved the $52,400 wire, and the honest answer is a message anyone in the channel could have clicked, recorded in a table your own admins can edit. If the approval has to be provable, here is the parts list:
We built the parts list so you don't have to. The approver apps, the cryptography, the verifier, the runbook — they're done, they're open source, you drop them in.
The part you couldn't outsource? Your agent's intelligence, your product's voice, the integration with your data. That's still yours.
The receipt spec, the verifier, the SDK, the CLI and the MCP server are MIT licensed and published on npm. Receipts verify with the open verifier no matter who runs the servers — your exit is guaranteed by construction. Self-host the platform under an Enterprise agreement, or use our managed instance.
We onboard teams one at a time, by hand. Leave an email and we will get you set up and walk you through the first approval.
Tell us what your agent does and we will get you an API key and a working approval in the same call.