← Go back

Guide // MandateGuard

How does a one-time spend token work after approval?

By the Ingenious AI team · Published 2026-11-14 · Reviewed 2026-11-14

▸ Run the free live demo See pricing

Free covers 1 agent and 1 policy with the full authorize flow, signed receipts and the approval inbox — no card required to start.

Short answer

A one-time spend token is what a human approval physically becomes in MandateGuard. When an approver approves a pending request, the system mints a token bound to that vendor and that amount, stores it only as a hash, and lets the agent redeem it exactly once to execute that purchase. It cannot be inflated, transferred to another vendor or saved for later — so an approval authorises one specific purchase, not a category of spending. If the facts change, the honest path is a new request and a new decision. The request, approval and redemption all appear in the receipt chain as one story.

What is a one-time spend token?

A spend token is a single-use credential minted when a human approves a pending MandateGuard request. It is stored server-side only as a hash, is bound to the approved vendor and amount, and is redeemed by the agent exactly once through the redeem endpoint to complete that purchase. It is not a balance, a session or a general permission: it authorises one execution of one already-decided purchase, and then it ceases to exist.

The gap the token closes

Approvals have a classic implementation flaw: the system records that a human said yes, then sets a flag the agent can spend against — and flags leak. The agent retries with a larger amount, reuses the yes for a sibling purchase, or executes hours later in a changed context, and the audit trail shows an approval that does not quite match the spend. The one-time token exists to close that gap with cryptography-grade literalism. Approval does not change the agent's permissions at all. It creates a token whose entire meaning is: this vendor, this amount, once. Execution consumes it. Anything else requires a new decision. The human's yes and the machine's spend become the same object, not two loosely related events.

Binding: vendor and amount

Binding is the security property doing the work. Vendor binding means the token cannot drift to a different counterparty — the approved supplier is the payable supplier. Amount binding means the token cannot inflate — the approved sum is the payable sum, so a price change between approval and execution surfaces as a new decision instead of a quiet overcharge. Together they make approval review fast and safe: the approver judges a specific purchase, and the system guarantees that the specific purchase is the only thing their judgement can release. When you evaluate any approval mechanism, ask what its yes is bound to. Unbound approvals are delegations; bound tokens are decisions.

Single use and hashed storage

Single-use redemption removes the second classic flaw: the reusable approval. A token that survives its purchase becomes standing credit, and standing credit in an agent's hands is indistinguishable from a raised limit — except it never went through a policy decision. MandateGuard stores tokens only as hashes, redeemable exactly once via the redeem endpoint; the plaintext exists transiently for the agent to present. This mirrors the system's treatment of license keys and workspace credentials: hashes at rest, values shown once, revocation and expiry handled by state rather than by hoping copies were deleted. If a token is ever exposed, its maximum value is one already-approved purchase — a bounded, receipted event.

The agent-side contract

Agents must treat tokens as single-purpose instruments. Redeem promptly for the approved purchase, never cache tokens for later convenience, never attempt redemption against a modified cart, and treat redemption failure as a terminal state requiring human attention rather than a retry loop. The failure cases are informative: already redeemed suggests a duplicated execution path in the agent; binding mismatch suggests the purchase facts changed after approval; unknown token suggests state confusion between workspaces or agents. Each is a bug or a process break worth seeing clearly, which is why the redeem endpoint's refusal should surface loudly in agent logs alongside the receipt reference of the original request.

The receipt story

The token completes a three-part receipt narrative. The original authorize call records the request and its REQUIRE_APPROVAL decision. The approval records the human decision with the facts that were visible. Redemption records the execution against the token. Because all three live in the hash-chained ledger, an auditor can follow one purchase end to end and confirm that what executed is what was approved, in amount and counterparty. This is the property that lets a business honestly say a human approved this spend — not a similar spend, not a spend in that week, but this one, for this much, to this vendor.

Tokens among MandateGuard's other instruments

The token is one of three scoped instruments in the system, and confusing them weakens all three. Policies are standing authority — they say what the agent may do unsupervised. Spend envelopes are task-scoped authority — they say what one job may consume, with its own approval conditions. Tokens are execution-scoped authority — they say this single approved purchase may now happen. Intents extend scoping across organisations, with budgets and authority windows that merchants can verify. Each instrument narrows a different dimension: time, task, transaction or counterparty set. A mature agent deployment uses the narrowest instrument that fits the decision in front of it.

Five actions to take now

  1. Treat tokens as secrets in transit: never log them in plaintext
  2. Redeem exactly once, for the approved vendor and amount only
  3. On redemption failure, stop and escalate — never request-and-redeem in a loop
  4. Build agents to request a fresh approval when facts change
  5. Audit approvals by their receipts: request, decision, redemption as one story

Worked example: one approval, one seat, nothing else

An owner approves a 64-dollar purchase from the workspace's helpdesk vendor for an extra agent seat. The token minted is bound to that vendor and 64 dollars. The agent redeems it and the seat is bought; the receipt chain closes the story: request, REQUIRE_APPROVAL, approval, redemption.

Later the same agent, mid-task, attempts to apply the flow to a 64-dollar analytics add-on from a different vendor. There is no shortcut to take: the old token is spent, and a different vendor would not have matched its binding anyway. The add-on makes its own request, waits for its own decision, and is denied on its merits. The approval the owner gave bought exactly one seat — nothing adjacent, nothing later.

Frequently asked questions

What if the price changes after approval?

The token is bound to the approved amount. A genuinely different amount is a new request with a new decision — by design.

Can a token be transferred to another agent?

Tokens belong to the approval that minted them, bound to vendor and amount within the workspace's flow. Treat any other use as invalid.

What does the agent see on a second redemption?

Failure. The token has already been redeemed; the correct behaviour is to stop and surface the state to a human.

How are tokens stored?

Only as hashes server-side, so a database read does not hand out live spending credentials.

Do tokens expire conceptually if never redeemed?

An unredeemed approval leaves the purchase unexecuted; the safe operational habit is to clear stale pending items by decision rather than letting them linger.

Keep reading

Put the policy in force: spend envelopes · intent registry · approval inbox · how the proof works · all MandateGuard guides