Security

How agent keys, sign-in, decisions and webhooks are protected, and what an agent can never do.

An approval service is only worth having if an agent cannot get around it. These are the guarantees the service keeps, and how. Each is checked by automated tests that try to break it.

Only an explicit yes means go

  • allowed becomes true in exactly two ways: a person approves the request, or an “always allow” rule a workspace owner wrote in advance approves it, recorded in that owner’s name.
  • Pending, denied, expired and withdrawn all mean the agent must not act. Silence is never a yes.
  • While a workspace or an agent is paused, nothing is allowed, whatever was approved. Any member may pause; only an owner may resume.
  • A decision is final. Nobody can change it afterwards.

Agent keys

Strength
Each key holds 244 random bits from the operating system, with the prefix gl_.
Storage
Only a SHA-256 hash is stored. The key is shown once, when it is created, and cannot be shown again.
Identification
The first characters are kept in clear, so the app can show which key is which.
Revocation
An owner can revoke a key at once. A revoked key is refused on its next call.
Scope
An owner can limit a key to some actions. Anything else it asks for is refused before it is stored.
Expiry
A key can stop working after a set time. An expired key is refused like a revoked one.

How to create, limit, expire and rotate keys is in Agent keys.

An agent key can ask for approval, and read, withdraw and report on its own requests. A request another key made answers 404, as if it did not exist. A key cannot list or decide requests, read the inbox, write rules, create keys or change any setting. A key on a person’s endpoint is refused with 403.

People and sessions

  • People sign in with a link sent to their email; there is no password to steal or reuse. Companies on the Enterprise plan can sign in through their own identity provider instead (SAML single sign-on).
  • Every call from a person carries a short-lived token, checked against the sign-in service’s public signing keys. Only asymmetric signatures are accepted.
  • Who decided always comes from that token, never from anything in the request, so a decision cannot be put in someone else’s name.

Rules that cannot be talked around

  • A rule matches the key an agent asked with, never the name the agent gives itself. An agent cannot rename itself into another’s “always allow”.
  • When rules disagree the most cautious wins: deny, then ask, then allow.
  • An amount limit only works in the cautious direction. Leaving the amount out, or stating another currency, never gets an agent past a limit.
  • Only owners can write rules.

Workspaces are kept apart

Every query on workspace data filters by workspace, so an agent or a person in one workspace can never see a request, key or rule from another. A request id from another workspace answers 404, as if it did not exist.

The database is reached only through the API. Every table has row level security switched on with no policies, so the database provider’s own data API cannot read product data.

Where your data is

VariantGate runs on a server in a Hetzner data centre in Germany, and its database is stored there, inside the EU. Cloudflare delivers the website and passes traffic on to that server, emails go out through Resend’s EU region in Ireland, and Stripe in Ireland takes payments. To keep everything on your own machines, run it yourself.

Invitations

A join link is a one-time secret, stored only as a hash like an agent key. It works once, for seven days, and can be withdrawn. A workspace always keeps at least one owner.

The Approve and Deny links in an email, and the link that turns off a weekly summary, are secrets stored only as hashes. An Approve or Deny link works once, decides nothing until you confirm on the page it opens, and records the decision in the name of the person it was sent to. An unsubscribe link can only turn off that person’s weekly summary of that one workspace, from then on, until they turn it back on under Settings. See Notifications.

Notification webhooks

  • A webhook address is a secret: it is stored, never shown again and never written to a log.
  • Only Slack, Microsoft Teams workflow and Discord webhook addresses are accepted, over https on those services’ own domains, and redirects are not followed, so the service cannot be made to call addresses on its own network.
  • What an agent wrote is posted as plain text, so a request cannot put a link or a mention in your channel.

Signed webhooks

  • Every event is signed with the webhook’s own secret, the Standard Webhooks way, so your side can tell it came from here and is recent. The secret is shown once, when it is made or replaced.
  • The address and the secret are never shown again or logged. People see the host and the label.
  • Only public https host names are taken. Each time an event is sent, every address the name leads to must be on the public internet, and only those addresses are connected to. Redirects are not followed.
  • An event carries no allowed: only the agent’s own read of its request says whether it may act. See Webhooks.

Logs and exports

  • Logs never contain tokens, agent keys, request bodies or email addresses. People and things are identified by id.
  • Every response carries an x-request-id. A server error shows only that id; its cause goes to the log.
  • The CSV export makes values that could be read as spreadsheet formulas safe, so a request’s text cannot run anything when the file is opened.

Deleting data

  • Only an owner can delete a workspace, and only by sending its exact name. Everything in it goes in one step, and its agent keys stop working at once.
  • A paid plan is cancelled with Stripe before anything is deleted. If Stripe cannot cancel it, nothing is deleted.
  • Anyone can delete their own account by sending their email address. Their sign-in is removed last, after their data, and the sign-in service is asked first whether it can be reached, so a failure there stops the deletion before anything is gone.
  • A sign-in token stays valid until it expires, so the service keeps a deleted account’s id for a week and refuses its tokens.
  • The key that removes sign-ins is kept like every other secret: never sent back and never logged.

What stays after an account is deleted is described under Teams and in the record.

Rate limits

Each agent key may make 600 calls a minute. A runaway agent gets 429 with a Retry-After header instead of flooding the inbox. Details are on Limits and errors.

The service answers the question an agent asks; it does not sit between the agent and its tools. Put the check in the code that performs the action, or use a hook, so the model cannot skip it.