Developer docs
Ask before your agents act.
Authentify sits between your AI agents and the actions they take. Before an agent does something that matters, your code asks Authentify. It answers allow, deny or escalate to a person, and seals the decision into a tamper-evident record you can verify without trusting us.
How it works
- Describe each agent once. Which actions it may take, on which resources, how fast, and which actions need a person.
- Ask on every action. Your code (or an SDK guard around the agent's tools) calls
POST /api/v2/authorizebefore running the action. - Act on the answer. Run it on
allow, stop ondeny, hold it onescalateuntil an approver (or the person the agent works for) decides. - Keep the evidence. Every decision, vote and consent change is sealed into a hash chain, independently timestamped, and exportable for auditors.
Authentify never runs the action itself. Treat the decision as the gate in your code, and optionally require a decision token at your API so an agent can't skip the check.
Quickstart (sandbox)
- Create a sandbox account.Sign up with no card. The sandbox is free for 30 days.
- Create an agent and copy its key.The setup wizard (or Dashboard → Govern → Agents) creates the agent with its allowed actions and resource scope and shows its key once. Sandbox keys start with
aut_sandbox_. - Ask before acting.Or with the TypeScript SDK (bash
curl -X POST https://www.authentify.bz/api/v2/authorize \ -H "x-api-key: $AUTHENTIFY_AGENT_KEY" \ -H "Content-Type: application/json" \ -d '{ "action": "SCHEDULE_APPOINTMENT", "resource": { "id": "patient-1042", "type": "patient" }, "context": { "clinic": "north", "slot": "2026-12-03T09:30" } }'npm install @authentify2026/sdk):tsimport { AuthentifyClient } from '@authentify2026/sdk'; const authentify = new AuthentifyClient({ apiKey: process.env.AUTHENTIFY_AGENT_KEY!, environment: 'sandbox' }); const result = await authentify.authorize({ action: 'SCHEDULE_APPOINTMENT', resource: { id: 'patient-1042', type: 'patient' }, context: { clinic: 'north', slot: '2026-12-03T09:30' }, }); - Handle all three decisions.An HTTP 200 alone never means "go": always read
decision What your code does allowRun the action. Optionally report what happened afterwards. denyDon't run it. reasonsays why (out of scope, rule, consent, rate limit, revoked agent…).escalateHold it. A person decides; the response carries an approvalRequestId. Not a denial.decision. - Look at the record.Dashboard → Monitor → Requests shows each request with the checks that ran, the votes, and its independent timestamp. Audit log shows the sealed chain.
Two kinds of key
| Key | Who uses it | For |
|---|---|---|
| Agent key | One agent | Asking /v2/authorize, reporting outcomes, minting short-lived OAuth tokens. Can't read your audit log or approve anything. |
| Account key | Your backend | Account-level calls: consent sessions, end-user links, audit logs, request status, rules. Never give it to an agent. |
Both are sent as x-api-key, except consent sessions, which take the account key only as Authorization: Bearer. The prefix picks the environment: aut_sandbox_… or aut_production_…. Base URL: https://www.authentify.bz/api. Approving or rejecting is never possible with any key; only a signed-in person can vote.