Skip to main content
Look up any term used in these docs. Each entry starts with one plain sentence, then gives the precise definition.

A

Account holder

The person who owns the Agent Vault account and its cards. Also called the owner. Only the account holder adds cards, sets agent card rules, approves held purchases and signup requests, enters live security codes, and completes challenges. An agent never acts as the account holder.

Activity feed

The dashboard’s record of what your agents did. The Activity page lists sessions and notification events: fills, refusals, approvals, challenges and key activity. Recording is best effort, so a gap is possible. Session records are kept for 30 days. See Activity.

Agent

The AI program that shops or signs in on your behalf. Any program that drives the proxied browser through the Playwright SDK or the MCP server with one of your API keys. It holds only mock tokens, never real card numbers or passwords.

Agent card

A set of spending rules you attach to one of your real cards. A rules object pointed at a funding card: spending limits, merchant rules, an approval threshold, a webhook and a pause switch. It holds no card number. Several agent cards can govern one funding card. At fill time the fill is permitted if any active agent card permits the merchant. The limits of every agent card that permits the merchant and is active or stopped at its limit then apply. One you paused yourself lends no limits. See Agent cards.

Agent-created login

A website account that an agent opened for you, with your approval. A login created through an approved signup request. The vault generates its password, so the agent never sees it. Only the key that created it may set its username once and rotate its password, for seven days after it was created by default, and only until the account holder replaces the password themselves.

ambiguousOnly

A flag meaning the page might be asking for a code, but the evidence is weak. A challenge detection result where the only signal is a field that might be a code input, judged from the current sign-in’s recent fills. It is not handoff material: requestChallengeHandoff() refuses it with handoff_not_applicable.

anySite

A catalog flag meaning a card is not limited to a list of sites. Set on a card in the token catalog that no agent card governs, or that an active agent card in allow-all or blocklist mode governs, so a catalog narrowed by URL returns it for every host. A card whose agent cards are all paused or at their limit carries neither anySite nor websites, and fills nowhere until one is resumed. It does not promise the fill will succeed: a blocklist card carries it too, and its excluded hosts are refused at fill time with merchant_blocked. A site that is not an authorized destination is refused with destination_unauthorized.

API key

The credential your agent uses to connect to Agent Vault. A key created on the API keys page and sent as a bearer token. It authenticates proxy sessions and API calls, and can be limited by a scope and granted the signup permission. An unknown key and a revoked key are indistinguishable to the caller (invalid_api_key).

Approval threshold

A dollar amount above which a purchase waits for your approval. The confirmAbove setting on an agent card. A fill for a total at or above it is held until the account holder approves or denies it, for up to 5 minutes. A checkout whose total cannot be read is always held when a threshold is set. See Approvals and notifications.

Attested detokenization

An optional mode that reveals a card only inside verified secure hardware. Revealing a card number only inside a hardware-attested confidential computing enclave. Off by default, and non-functional without a provisioned enclave. See Attested detokenization.

Auto-pause

An agent card setting that stops the card when a limit is hit. The autoPauseOnLimit setting, on for every agent card created in the dashboard. While it is on, any breach of the card’s own limits, including a per-transaction one, changes the card’s status to limit, and it stops filling until resumed. With it off, the breaching fill is refused and the card stays active.

B

Binding checks

Final safety checks run just before a card number is typed. After a card is revealed and before it is typed, the vault token’s owner must match the card’s account, and the revealed last four digits must match the card’s. A failure is card_blocked, and nothing is typed.

Blackout

The black image every screenshot becomes after the first fill. The redaction value on a screenshot taken after any token fill in a session. The whole image is a solid black rectangle. It is not a dark page or a failed render. Only a new session clears it. See Redaction.

Blocklist

An agent card rule that allows every merchant except the ones you list. A merchant mode, available on every plan. The listed hosts are refused with merchant_blocked, and are never shown to the agent. See Merchant rules.

C

Challenge

A security step only a person should complete, such as a bank check or CAPTCHA. 3-D Secure, a one-time code, a CAPTCHA, or a silent block (blocked). The proxy detects and classifies these and never solves them. The agent hands off instead. The one exception is the authenticator step of a sign-in for a login with a stored TOTP secret. See Handling challenges.

Challenge handoff

Passing the live page to you so you can complete a challenge yourself. Streaming the agent’s live page to the account holder, who completes the challenge. While it runs, every other page call from the agent is rejected, so the agent cannot watch what is typed. Afterwards, in a session with no token fill, screenshots are not blacked out, so a page that echoes what was typed stays protected only if the agent does not capture it. See Agent context. The agent waits up to about 10 minutes, then resumes on the same page. Different from a purchase approval, which is a yes or no decision.

Checkout recipe

A saved description of how a store’s checkout works. A JSON document in the recipe registry describing a platform’s or merchant’s checkout: detection signals, where the card fields are, a step per phase, gotchas and any dead end. See Recipe format.

Corroboration

The checks run before a checkout counts as confirmed. What get_session_outcome checks before recording a confirmation: the submit control has left the page, no popup is waiting, no error or challenge is showing, and the success marker is still present. See Confirming outcomes.

Cross-user denial

A refusal because a token belongs to a different account. cross_user_denied: the user id inside a token does not match the API key’s account. Three within five minutes revoke the key automatically (key_auto_revoked).

CVV

The card’s security code, which Agent Vault never stores. The mock cvv token in the catalog has nothing stored behind it. Filling it starts live CVV entry on a deployment that offers it, and otherwise fails with cvv_entry_required. See CVV policy.

D

Dead end

A store where an agent can never complete checkout on its own. The dead-end recipe status, with a deadEnd.type saying why: turnstile, captcha, paypal-only, login-wall, 3ds, stripe-config or automation-blocked. Recording one is as valuable as recording a success. See Recipe statuses.

E

Extra websites

Additional sign-in addresses you allow a login to fill on. Hosts the account holder authorized for a login whose sign-in redirects to a different domain, listed as extraWebsites in the catalog. They match exactly, unlike website, which also matches subdomains. Only the account holder can add one.

F

Fill

Typing a value into a field on a web page. An SDK or MCP call that types into a form field. When the value is a mock token, the proxy substitutes the real value inside its own browser and runs every check at that moment. See fill-time substitution.

Fill-time substitution

Real values are swapped in at the very last moment, inside Agent Vault’s browser. A mock token becomes its real value only inside the proxy’s browser, as it goes into the field. Every policy check (site, account, scope, merchant rules, destination, limits, approval) happens at that moment. A failed check is a refusal with a code, and nothing is typed. A failure after typing starts, such as fill_rejected, can leave part or all of the value on the page.

Fingerprint

A product URL with the private parts removed. The scheme, host and path of a URL, with the query, the fragment and any hosted-checkout session token removed. The only form of a URL to send to the recipe registry. The SDK’s fingerprintUrl() produces it from an absolute URL, and returns a string that does not parse unchanged.

Funding card

A real card you added, which merchants actually charge. A payment card the account holder added in the dashboard. Its number is stored in the vault, never on an Agent Vault server (the proxy uses it for one fill, and keeps it in memory until the session closes only to scrub it from agent-facing output), and it is what the merchant charges. Governed by zero or more agent cards. With none, it can be used for any amount on any authorized destination, even when the total cannot be read. The plan’s payment quota still applies. See Cards.

H

Handoff

See Challenge handoff.

I

Identity

Your name, email and phone details, as tokens. The identity entries in the token catalog: first name, last name, full name, email and phone, plus phone parts. phone is in E.164 form, such as +15550100134, except for an older phone stored as free text, which fills as typed until the account holder saves it again. They resolve on any authorized destination. default is the account profile. The account holder can add more, each under its own label such as work, and each appears in the catalog.

inconclusive

A flag meaning a page scan did not finish. Set on observe() and detectChallenge() results when the scan did not complete. Every false beside it means unknown, not absent. A finished scan omits the flag, so test !result.inconclusive.

L

Live CVV entry

You type the security code yourself while the checkout waits. Live CVV entry is a per-deployment capability. Where it is offered, filling the cvv token pauses the checkout, and the account holder enters the real code from their dashboard for that one fill. The agent never sees it. If it is not entered within about 3 minutes, the fill fails with cvv_entry_required. On a deployment without it, the fill fails with cvv_entry_required at once, and no prompt is raised. See Live CVV entry.

Login

A saved website username and password. A credential record in the vault with a website, optional extraWebsites, a password, and optionally a username, email and TOTP secret. It fills only on its registered site. See Logins.

M

MCP server

The Agent Vault tool server for AI assistants. A Model Context Protocol server that exposes Agent Vault as tools (open_page, fill_checkout, get_token_catalog and others) to any MCP client. See MCP setup.

Merchant

The online store an agent buys from. In recipes and rules, a merchant is identified by its host, for example shop.example.com.

Merchant recipe

A recipe describing one specific store. A recipe with kind: "merchant", naming the store’s hosts, the platform recipe it uses, any selector overrides and any dead end. See Recipe format.

Merchant rules

Which stores an agent card may be used at. An agent card’s merchant mode: Allow all, Blocklist (any merchant except the listed hosts, every plan) or Allowlist (only the listed hosts; choosing or editing it needs Enterprise, and one set before a downgrade stays enforced). A refusal is merchant_blocked in both list modes, so the agent cannot map your settings.

Mock token

A stand-in value the agent uses instead of real card or login details. A string such as av_card_1_number_u7x2k that the agent fills in place of a real value. Stable and never expiring while the record or field behind it exists, and useless outside the proxy: anywhere else it is just text. See Token grammar.

N

Network tokenization

Using card-network tokens instead of card numbers. Not active. Support for scheme-issued network tokens exists in the data model but is inactive: no card carries one, and cards are filled with their card number. See Card storage.

O

One-shot authorization

Permission for exactly one security code fill, right after the card number. Filling a card number arms a single authorization for the security code that follows. The next action on the page uses it up. Reads such as screenshots, observe() and count() do not, except a screenshot that returns securityCodeAuthorizationReleased: true. Then fill the number again before the code. Something in between means a second approval.

P

Page hint

Telling Agent Vault where something is on the page. A declaration of where something is, never what it is. Used to point at an order total the vault cannot find, or at an otpauth:// code during 2FA setup. See Page hints.

Partial

A recipe status: the card fields were reached and filled on a real run. The partial recipe status. A real checkout reached a filled card form. No purchase was completed.

Phantom spend

Spending that counts against a limit even though no order was placed. A card fill whose order never completed. It still counts against the agent card’s spending windows until they roll past it, because a fill is recorded the moment it lands.

Plan quota

The usage allowance that comes with your plan. Account-level limits set by the plan: payments per rolling 30 days and the number of agent cards linked to a funding card that still exists. Reaching one gives plan_quota_reached. See Limits and timeouts.

Platform recipe

A recipe describing how a whole store platform’s checkout works. A recipe with kind: "platform", for example Shopify or WooCommerce, shared by every store on that platform. Changes are reviewed by a maintainer.

Playwright SDK

The code library for driving Agent Vault from your own program. The @agentvault/playwright package. It mirrors Playwright’s API but runs the browser inside the proxy. See Install the SDK.

Precondition

A step a store requires before checkout, which an agent can complete. A clearable gate recorded on a recipe: delivery_location, region_select or account_required for checkout. Unlike a dead end, it does not make the store unusable.

Proxy

Agent Vault’s secure browser service, where real values are typed. The service that runs the browser the agent drives, substitutes real values for mock tokens at fill time, and enforces every rule. The agent talks to it over a WebSocket.

Proxy session

One browser run, from launch to close. The browser, its tabs and its cookies inside the proxy. It ends after 15 minutes idle or 60 minutes in total, and starts with an empty profile every time. See Sessions and timeouts.

Purchase approval

A yes or no decision you make on a held purchase. A held fill at or above an agent card’s approval threshold, waiting up to 5 minutes for the account holder to approve or deny it. Different from a challenge handoff, where the account holder completes a step on the live page.

R

Receipt

A screenshot of an order confirmation, kept with the session. A redacted image of a confirmation page stored on the session record. It is stored only when no sensitive value was filled anywhere in that session.

Recipe registry

The shared, public library of checkout recipes. The open registry of checkout recipes, readable without an API key at GET /v1/recipes and grown through screened submissions. See Checkout recipe registry.

Recipe status

How much has been proven about a recipe. One of verified, partial, unverified or dead-end. See Recipe statuses. A login you have on another part of the same website. A login registered on a sibling host of the same site as the URL the catalog was narrowed to. It carries hosts only, no tokens, and will not fill on the page you are on: navigate to its website.

Run report

A summary of what a session did, without any values. The proxy’s own record of a session: which token kinds were filled, how many fills were refused, whether a confirmation was recorded, and the order total. Kinds and counts, never values.

S

Scope

The list of cards and logins one API key may use. An API key’s allowlist of card and login records. Unset means all, an empty list means none, and a list means only those. A record outside it is refused with scope_denied.

Signing secret

The secret that proves a webhook came from Agent Vault. A per-agent-card secret starting whsec_, created when the first webhook URL is saved. It keys the HMAC-SHA256 signature in X-AgentPay-Signature. See Webhooks.

Signup permission

Permission for an API key to request new website accounts. A per-key setting, off by default, that lets an agent file signup requests and maintain the logins it created. Without it, those calls fail with signup_not_enabled. See Agent signup.

Signup recipe

A recipe describing a website’s account signup form. A recipe with kind: "signup" or "signup-platform". It describes a form and is not permission to create an account. See Signup recipes.

Signup request

An agent’s request to open a website account for you. A request filed with POST /v1/logins/requests or createLogin(). It usually starts pending and creates nothing until the account holder approves it, and expires after 24 hours. For a site and email on the holder’s Pre-approved signups list, it starts approved and the login already exists. At most five can wait at once per account.

Spending limit

The most an agent card may spend in a period. Per transaction, per day (Plus and Enterprise), per week and per rolling 30 days, in US dollars, set on an agent card. Every window is rolling. A breach is refused with limit_reached. Unmetered deployments (preview, development, test and staging) enforce only the per-transaction cap.

T

Token catalog

The list of mock tokens your agent may use. The tokens one API key may use, plus the non-sensitive details needed to choose between them. Safe to give an agent. See Token catalog.

TOTP

The six-digit codes from an authenticator app. Time-based one-time password (RFC 6238). The only second factor Agent Vault can complete without a person: the backend derives the current code from the stored secret and sends only that code to the proxy, which types it into the site. The proxy and the site receive the code; the agent and the account holder never see it. A secret already stored never leaves the backend. The one exception is enrolment by an agent: the site shows the new otpauth:// secret on a page the proxy is rendering, and the proxy reads it there and sends it to the backend to encrypt. The agent never receives it. See TOTP.

Tunnel

A route that makes the browser’s traffic leave from your own machine. A connection back through the agent’s machine, so merchants see that machine’s IP address instead of the proxy’s. Used by default when available. Set SELF_AGENT_PAY_TUNNEL=1 to require it. See Egress and tunnelling.

U

unanswered

The proxy stopped waiting for one call. The session is still fine. Not terminal: the session, its pages and its cookies are intact. Contrast request_timeout, the SDK’s own deadline, which closes the session.

Unverified

A recipe status: worked out from the platform, but never run. The unverified recipe status. Derived from the platform’s known structure and never tested on this store.

URL locking

Each saved login fills only on its own website. A credential resolves only where it is registered. Logins fail closed: they fill only on their website, its subdomains and exact extraWebsites. Cards resolve wherever their agent card rules allow, and only on an authorized destination. See URL locking.

V

Vault

The secure store where your card numbers are kept. Cards are tokenized in the account holder’s browser by the vault provider’s iframe, so the number never reaches the Agent Vault backend. The proxy types it for one authorized fill, and keeps it in memory until the session closes only to scrub it from agent-facing output. See Card storage.

Verified

A recipe status: a real purchase was proven with a receipt. The verified recipe status. A real purchase completed, proven out of band by a merchant receipt or card statement. Only a maintainer can set it. A confirmation page is not proof.

W

Webhook

A message Agent Vault sends to your server when something happens on an agent card. A signed HTTPS POST to the endpoint set on an agent card, for fills, refusals, limit events, auto-pause and held purchases. Plus and Enterprise only. See Webhooks.