Signed agent identity
BuiltEvery agent proves who it is with its own signing key (OAuth 2.0 private_key_jwt, RFC 7523), never just a name. Unknown agents wait for the business's review by default, and any agent can be blocked at once.
Last updated: 2026-10-20
What Allowfy does to keep AI agents within a business's rules, what it can't do on its own, and what is still planned. Each item says whether it is built today or planned, so you can check it rather than take it on trust.
Every agent proves who it is with its own signing key (OAuth 2.0 private_key_jwt, RFC 7523), never just a name. Unknown agents wait for the business's review by default, and any agent can be blocked at once.
Session tokens are DPoP-bound (RFC 9449) and short-lived, so a stolen token alone is not enough. Account tokens can be revoked, and signing out ends every session signed in with them immediately.
Customers connect an assistant with a short code (RFC 8628) on the business's own page, or with the business's identity provider (OpenID Connect). Passwords and PINs never pass through the agent, and Allowfy doesn't store PINs.
Every action needs a rule. Without one it is refused. Rules can allow, ask the customer, hold for the team or refuse, with limits on amounts and dates, and every published version is kept.
A customer approves one revision of one operation: its amount, currency and details. If the terms change, the approval no longer applies and a new one is needed. Unconfirmed proposals expire.
Rules can cap each customer's total and number of operations over a window of days, across every agent, so a large request can't be split into small ones. A business can replay its real operations of the last 90 days against a draft before publishing it.
Each operation is sent to the business's system with its ID as the idempotency key. Retries reuse the same key. If the outcome can't be learned, the operation is never reported as failed: it waits for a person to settle it.
One switch refuses every agent request, holds confirmed operations that haven't been sent yet, and blocks approvals until the business resumes. A call already on its way can't be recalled; its answer is still recorded.
Every business's data is separated in the database with PostgreSQL row-level security, enforced for each request, with automated tests that try to read across businesses.
Webhook and identity-provider secrets are encrypted with AES-256-GCM before they are stored, and shown once. Account passwords are hashed with scrypt.
Every request Allowfy sends to a business's system is signed with HMAC-SHA256 and a timestamp, so the business can reject forged or replayed requests.
Allowfy only calls public HTTPS addresses, with size and time limits, so an agent or a setting can't point it at internal systems.
Every step is recorded in a hash chain. Exports add a checkpoint signed with an Ed25519 key the database never holds, so a copy the business keeps exposes a rewritten history. This is tamper evidence, not tamper prevention, and not a promise about legal admissibility.
HTTPS everywhere, HSTS, and strict security headers. Accounts need a verified email, and sign-up is protected against bots.
Server and browser errors are reported with personal data scrubbed first, and uptime is checked from outside.
The open-source allowfy package verifies Allowfy's signatures and handles its requests on Node, Deno, Bun, Supabase, Cloudflare and Next.js, with ready templates.
Allowfy runs on Railway in the United States: servers and the PostgreSQL database.
The business's settings, rules, agents, operations and ledger, and a reference to each customer from the business's system (not their password or PIN). Team accounts keep a name, an email and a hashed password.
Operations and the ledger are kept while the business uses Allowfy, because the ledger is a record. Allowfy Check results are deleted after 180 days. See the privacy policy for site data.
The ledger can be exported as a signed file at any time and verified without Allowfy. Operations are available through the API.
Write to hello@allowfy.com to have a business, an account or personal data deleted. We answer within 30 days.
| Provider | What for | Where |
|---|---|---|
| Railway | Hosting: servers and database | United States |
| Resend | Sending account emails (verification, password reset) | United States |
| Sentry | Error reports, with personal data scrubbed | United States |
| Cloudflare | Turnstile bot check on sign-up and forms | Global |
| Optional sign-in with Google | Global | |
| UptimeRobot | Uptime checks of public pages (no personal data) | Global |
The business can pause every agent at once with the emergency stop. We investigate, contain and fix the cause, and tell affected businesses what happened, what it affected and what they should do, as soon as we have confirmed it, and within 72 hours of confirming an incident that affects their data.
Write to hello@allowfy.com with "Security" in the subject. Tell us what you found and how to reproduce it. We reply within 3 business days, keep you informed until it is fixed, and won't take action against good-faith research that avoids harming businesses, customers or their data. Our security.txt is at /.well-known/security.txt.
By an outside firm, before rolling out to enterprise customers. We will publish a summary.
A standard agreement for businesses that need one.
We don't hold SOC 2 or ISO 27001 yet. We will say here when an audit starts.