Valta Docs

How Valta Works

Valta puts a governance layer between your AI agents and their financial actions. Three layers enforce every rule, in code, before money moves.

Layer 1: Spending policy

You define the rules. Every agent is assigned a spending policy that sets:

  • Daily limit — maximum cumulative spend in a 24-hour window (resets midnight UTC)
  • Per-transaction maximum — hard cap on any single financial action
  • Monthly limit — maximum spend across a calendar month
  • Approval threshold — actions above this value pause and wait for human sign-off
  • Blocked categories — e.g. gambling, adult, crypto_exchange
  • Allowed domains — if set, the agent can only call these endpoints
  • Blocked domains — always blocked, regardless of other settings

Rules live in your Valta account. They are stored and enforced server-side. The model cannot read them, negotiate around them, or override them by claiming urgency or authority.

Layer 2: Policy enforcement

Every financial action goes through Valta's policy engine before it executes. The check runs in backend code — not in the model's output.

If any single rule fails, the action is blocked. The agent does not retry or escalate. The violation is recorded and the agent receives an error response.

This means:

  • A prompt that says "ignore spending limits" has no effect.
  • An agent that attempts to split transactions to stay under limits is detected (structuring detection is built in).
  • There is no way to spend above a hard cap from within the model.

Layer 3: Audit trail

Every action — attempted, blocked, approved, or executed — is logged to a tamper-evident audit trail. Each entry includes a cryptographic hash of the previous entry, forming a hash chain. Modify any entry and the chain is invalid.

Entries cannot be deleted or modified, including by Valta staff.

Architecture

Developer │ ▼ Valta API │ ▼ Policy Engine ──── Audit Trail (every action logged) │ ▼ Agent Wallet │ ▼ Financial Action (executes only if all rules pass)

What Valta is NOT

What people assumeReality
An AI modelValta uses AI to power agent runs. The model is a component, not the product.
A crypto exchangeValta uses Circle to issue and settle USDC. Valta is not a financial institution.
A bankValta does not hold deposits. Agent wallets are USDC balances on Base.
Just monitoringMonitoring observes after the fact. Valta actively blocks actions before they execute.

The policy enforcement layer and audit trail are always on. They cannot be disabled by an agent, a developer, or through the API.

Next steps