Autospend
Programmable cards for AI agents

Give your agents spending power. Keep every control.

Autospend issues scoped, single-use virtual cards to every agent you run — each bounded by its own policy, daily and exposure ceilings, and a kill switch that stops every agent at once. No approving each purchase. No shared card. No blast radius.

Works anywhere Visa does · one MCP tool · settles on stablecoin rails

Built by the team behind Figo

Agents · control plane demo
Active exposure
$0 / $2,000 ceiling
Audit chain
4 agents · simulated demo
Runs in Claude· Cursor· any MCP runtime· Settles on Visa rails· USDC
Built for many agents, not one

Approving every payment doesn't scale past one agent.

A personal assistant that asks before each purchase is fine when there's one of them and one of you. Run fifty autonomous agents and human-in-the-loop becomes the bottleneck — so the controls have to move into policy, not your inbox.

Approve-each-payment

One agent, one human, every purchase confirmed.
  • You are in the loop for every single charge — your agents moves at your typing speed.
  • One shared wallet or card behind everything; a bad prompt reaches all of it.
  • No per-agent policy — every agent has the same reach as every other.

Policy-once, spend-continuously

All your agents spending inside limits you set once.
  • Set each agent's policy at key-mint time; it spends autonomously within it, forever.
  • Every card is single-use and scoped — blast radius is one capped card, not the account.
  • Freeze every agent or revoke just one instantly; the rest never notice.
How it works

Four steps from prompt to audited purchase.

Fund the tenant balance once. From there every card an agent asks for runs the same path — checked against policy, issued scoped, reconciled on settlement, and written to an append-only audit chain.

1

Mint a scoped key

Give each agent a key with its own per-txn cap, daily cap, merchant categories, and card lifetime. That policy is the agent's whole reach.

2

Agent requests a card

At checkout the agent calls one tool with the order total. Policy and account-wide ceilings are re-checked server-side before anything is issued.

3

Single-use card issued

A virtual card capped to the amount, scoped to the categories, live for its TTL. Used once at checkout, then it auto-closes.

4

Settle & audit

Settlement reconciles real spend and releases exposure; every decision — accepted, declined, frozen — lands hash-chained and attributable.

The card

A real Visa card, born for one purchase.

Every approved request mints a fresh virtual card — capped to the order total, locked to its merchant categories, and gone after a single charge. A leaked number is worth almost nothing.

  • Issued the moment your agent reaches checkout
  • Capped to the amount, scoped to the category
  • Auto-closes after the charge — or when its TTL expires
Use-case presets

One agent, one key, one policy.

Each use case is just a scoped key. A grocery agent physically cannot book a flight; a travel agent cannot buy gift cards. Give each its own policy and revoking one never touches the others.

Food delivery
MCC 581258145411
per-txn
$75
daily
$300
ttl
1h
SaaS & infra
MCC 573458174816
per-txn
$200
daily
$1,000
ttl
none
Rideshare
MCC 4121
per-txn
$100
daily
$400
ttl
2h
Travel & flights
MCC 451130xx7922
per-txn
$1,500
daily
$5,000
ttl
24h
One tool · any agent

A single MCP tool your agents already speak.

Drop the Autospend MCP server into Claude, Cursor, or any MCP runtime with one config block. Your agent gets a request_card tool — the card comes back scoped and single-use.

agent · request_card
// the agent, at checkout, calls one tool
request_card({
  amountInUSDCents: 1840,          // $18.40 — the order total
  allowedMccs: ["5812"],           // restaurants only
  memo: "DoorDash — dinner order"
})

// → single-use card, capped, scoped, live once
{ last4: "4242", status: "active",
  pan: "•••• •••• •••• 4242" }

// over the cap? no card — a reason instead
policy_rejected: "amount exceeds per-txn cap $75.00"
The controls

Every guarantee an operator asks for.

Kill switch

Freeze in one call

Freeze a tenant or revoke a single agent key and its live cards are locked at the network — instantly, across every agent or surgically.

Exposure ceiling

Bounded stock, not just flow

Daily caps bound issuance; the exposure ceiling bounds the total value of all live cards at once, settlement-aware and always conservative.

Audit chain

Hash-chained, per tenant

Every decision is appended to a tamper-evident chain — accepted, declined, frozen — each attributable to the exact agent key.

Single-use

Blast radius of one card

Cards auto-close after a charge and expire on a TTL. A leaked PAN is worth one capped purchase, in one category, for a short window.

Settlement

Spending power tracks reality

As purchases settle, exposure is released back. Your agents' headroom reflects what actually cleared, not worst case.

Isolation

No shared card

Each agent hangs off its own key and policy. Compromise one and it cannot reach another's budget, categories, or cards.

Private beta

Let your agents spend — on your terms.

Autospend is in private beta. Leave your email and we'll reach out with early access and the MCP setup guide.

Built by the team behind Figo