PiggybankTry the demo

Financial IAM for AI agents

Give every AI agent an allowance.

Piggybank gives autonomous systems programmable budgets, deterministic approvals, and a proof trail for every economic action.

Try an authorizationSee how it works

The model proposes. Deterministic code decides.

Allowance / monthly$100.00authority, not customer funds
Who

is acting?

Why

for which task?

How much

authority remains?

Prove it

what happened afterward?

The missing layer

API keys quietly became spending power.

Agents can call models, search the web, rent compute, and buy data. Each tool has its own limit and invoice—but none carries the full intent, mandate, or shared allowance.

More autonomy without bounded authority creates invisible risk.

A decision you can inspect

One request. One explainable answer.

Compare three outcomes from the same $5 research task. Every result includes the reason, remaining allowance, and an immutable trace.

Authorization requestExample fixture

Summarize competitor product pages

Agent
Research sub-agent
Vendor
Model provider
Category
model
Requested
$0.43
HumanMandateAgentTask
Deterministic decision

APPROVED

APPROVED_BY_POLICY

Known category and amount are within the task and transaction limits.

Task remaining
$4.57
  1. Identity active
  2. Mandate in scope
  3. Budget reserved

The control loop

Permission stays attached to intent.

Authorization is not a dashboard setting. It is a deterministic step in the agent's execution path.

  1. 01 / Propose

    Name the action.

    The agent submits its task, amount, vendor, category, and purpose.

  2. 02 / Authorize

    Evaluate authority.

    Identity, mandate, policy, and available allowance resolve atomically.

  3. 03 / Execute

    Run the action.

    The customer or provider executes only after an approved reservation.

  4. 04 / Reconcile

    Record the truth.

    Capture actual cost, release the remainder, and preserve the audit chain.

Multi-agent authority

Delegation divides an allowance. It never duplicates it.

Give specialist agents bounded authority beneath a parent. Every allocation, reservation, and return conserves the original limit—even when requests arrive concurrently.

Main agent$1,000
Research$150
Coding$500
Testing$200
$150 remains unallocated—and cannot be copied.

The architecture boundary

Control permission. Never custody the money.

Piggybank records economic authority and decisions. Regulated providers and vendors continue to hold and move real value.

Piggybank controls

  • Agent identity and mandates
  • Allowances and delegation
  • Policies and approvals
  • Authorization journal and audit

Providers execute

  • Model and API usage
  • Card and payment movement
  • Cloud and compute
  • Settlement and reconciliation data

Built for the execution path

Ask before the agent acts.

A stable API returns approved, requires approval, or denied—with machine-readable reason codes and safe retries.

Inspect the architecture

authorization-request.json

{
  "agent_id": "agt_research",
  "task_id": "tsk_competitors",
  "amount": "1.250000",
  "currency": "USD",
  "category": "dataset",
  "purpose": "Compare competitor pricing"
}
202requires_approvalNEW_VENDOR_APPROVAL

The build starts now

Help define the control plane for agent economies.

We're looking for design partners, founding contributors, and teams giving autonomous agents real work.

Run the demo