Skip to content
_AI AGENT PAYMENTS/

Let agents move money without losing control of it

Give AI agents a ledger to act on: budgets they cannot exceed, transactions they cannot duplicate, and an immutable record of everything they did and why.

agent intent
budget check
atomic posting
audit trail
_THE PROBLEM/

Agents are getting wallets before they get guardrails

Autonomous money movement without a system of record is a liability engine.

A stack of new protocols is racing to let agents pay (mandates, agent cards, agent wallets, agentic checkout), but they all solve authorization, not accounting. A retried tool call can still double-pay, a hallucinated parameter can still overspend, and when the auditor asks why money moved, a chat log is not an answer. With fleets running tens of thousands of tasks, setting limits by hand is not humanly possible.

01

Agent spend limits live in prompts, and prompts are not controls.

02

Retried and replayed tool calls double-pay because nothing enforces idempotency.

03

Nothing ties a settled transaction back to the agent, the task, and the mandate that authorized it.

04

Fleet spend fragments across rails (an agent card here, a stablecoin wallet there) with no single record of what the fleet spent.

_PROOF/ Formance ships an MCP server, so agents post to the ledger through the same guarantees as any other client, never around them.

_FLOW OF FUNDS/

Intent, guardrail, posting, proof

The agent proposes; the ledger constrains, executes, and remembers.

The agent's payment intent, carrying its mandate, hits a budgeted account with ledger-enforced limits, posts atomically with a deterministic idempotency key, and lands in an immutable log attributed to the agent, the task, the delegating human, and the credential that authorized it.

agent guardrails
agent intenttool call
budget checkledger-enforced
atomic postingidempotent
audit trailwho · what · why
_CAPABILITIES/

The primitives agentic money needs

Built on the open-source Formance ledger.

Ledger-enforced budgets, replay-safe execution, and full attribution, composed for autonomous financial workflows.

_AGENT BUDGETS/

Each agent operates from a funded account inside a budget hierarchy (org → team → user → workflow → agent) with balance constraints the ledger enforces: an agent structurally cannot spend money it does not have.

enforced, not prompted

_REPLAY-SAFE EXECUTION/

Deterministic idempotency keys collapse retried and duplicated tool calls into one transaction: double-pays are impossible by construction.

exactly-once

_FULL ATTRIBUTION/

Every posting carries agent, task, session, delegating principal, and the mandate or credential that authorized it (an AP2-style intent, an agent-card token, an x402 receipt): the "why" behind every cent, queryable forever.

who, what, why

_MCP & APPROVALS/

Agents connect through the Formance MCP server, query their remaining budget and request limit changes programmatically mid-task, and high-value flows route through human approval: human on the loop by policy, in the loop by exception.

human on the loop

Read the AI & MCP docs
_BY THE NUMBERS/

Agent-grade controls

double-pays0idempotent by construction
action attribution100%agent · task · principal
spend limitsenforcedledger constraints · not prompts
human approvaloptionalper flow · per threshold
_NUMSCRIPT/

Authorize the maximum, settle the actual

The x402 upto pattern, atomic end to end.

The agent authorizes a maximum spend, actual consumption settles to the vendor, and the unused balance returns to the agent in the same atomic operation. A partial execution is not a state the ledger can be in, and an agent that cannot cover the authorization fails cleanly, with no overdraft to clean up.

upto-settle.num
// x402 upto: authorize max, settle actual, refund rest
vars {
  monetary $authorized   // agent's signed maximum
  monetary $consumed     // actual usage settled
  account  $agent
  account  $vendor
  string   $auth_id
}

// 1. Escrow the authorized maximum
send $authorized (
  source = @agents:$agent:budget
  destination = @escrow:requests:$auth_id
)

// 2. Settle actual consumption to the vendor
send $consumed (
  source = @escrow:requests:$auth_id
  destination = $vendor
)

// 3. Refund the unused balance to the agent
send [USDC/6 *] (
  source = @escrow:requests:$auth_id
  destination = @agents:$agent:budget
)
_VOICES/

Trusted by builders

LiberisLiberis
Formance enabled us to ship new lending products faster by providing a customizable foundation that let us reliably express the complexities of our flow of funds.

Embedded finance · reconciliation across 14 countries

See customer story
DoctolibDoctolib
Formance is the foundation of our Financial OS, the open-source approach lets us retain control over this key component.

Healthcare · Financial OS

PayflipPayflip
Formance helped us kick off our move into fintech, with robust infrastructure and intuitive developer tooling.

Benefits · fintech infrastructure

_TRUST/

Built for regulated money.

Enterprise controls, the certifications auditors ask for, and an immutable record, out of the box.

See our trust center
THE COMPLIANCE STACK
IAM, RBAC & SSO
Non-repudiation
Audit logs
Observability & monitoring
Real-time events
Admin console
_AUDIT-LOG/ NON-REPUDIATION

09:41:07Z AUDIT gateway POST /api/ledger/v2/main/transactions 200 sub:ops@acme.io

09:41:09Z LOG id:4093 NEW_TRANSACTION ledger:main

09:41:12Z LOG id:4094 SET_METADATA ledger:main

09:41:15Z LOG id:4095 REVERTED_TRANSACTION ledger:main

09:41:18Z AUDIT gateway GET /api/ledger/v2/main/logs 200 sub:audit@acme.io

CERTIFICATIONS
AICPA SOC IICERTIFIED
ISO 27001CERTIFIED
DORACOMPLIANT
REGULATORY FRAMEWORKS
EMIFCAMTLNYDFS Trust CharterOCC CharterMiCAGENIUS
_ATTESTED/
FAQ

Agentic payments, answered

01 / HOW IS THIS DIFFERENT FROM PROMPT-LEVEL LIMITS?

Prompts are suggestions; ledger constraints are physics. An agent's account balance and sending rules are enforced at posting time, no matter what the model decides.

02 / WHAT HAPPENS WHEN AN AGENT RETRIES A PAYMENT?

Every intent carries a deterministic idempotency key, so retries, replays, and duplicated tool calls collapse into one transaction, exactly-once by construction.

03 / CAN A HUMAN APPROVE BEFORE MONEY MOVES?

Yes: flows can route through approval steps above a threshold, where the agent proposes, a human confirms, and the ledger executes and records both.

04 / HOW DO AUDITORS SEE AGENT ACTIVITY?

Every posting is immutable and attributed (agent, task, session, delegating principal, authorizing mandate), so agent spend audits are ledger queries, not log archaeology. The record is rail-agnostic: agent cards today, stablecoin wallets and x402 tomorrow, bank rails throughout.

_GET STARTED/

Give your agents a ledger, not a liability

See how Formance fits your agentic stack: book a demo or start building.