Pre-launch · early access open

Every agent payment, checked before it moves.

HalvorAI checks each payment an AI agent makes against what it was authorised to do and how it normally behaves. Then it hands attributed, audit-ready events to the monitoring you already run.

Request early access → Try a mandate
Across CardsStripex402StablecoinsBank transfers
Agent payments 00:00:00 UTC
Agent Payee · reason Amount Decision
    Checked0
    Approved0
    Held0
    Blocked0
    Example data. Shows how HalvorAI decides, not live traffic.
    The gap

    Your monitoring sees the card. It doesn't see the agent.

    Transaction monitoring was built to profile people. When an agent pays with a person's card, the bank sees a normal small payment. It can't see which agent paid, who it acted for, or what it was told to do.

    What your monitoring sees
    Account
    Card ending 4421
    Payee
    gift-cards-direct.shop
    Amount
    $200.00
    Velocity
    Within customer norms
    Agent
    Unknown
    Instruction
    Not in the data
    ApprovedLooks like the cardholder.
    What HalvorAI sees
    Agent
    agt_21 · shopping assistant
    Acting for
    Card ending 4421 (owner)
    Mandate
    Groceries and household, $150 per order
    Instruction
    "Reorder paper towels"
    Behaviour
    6th identical payment in 40 seconds
    Payee
    New, outside usual merchants
    BlockedOutside mandate. Logged for review.

    Attribution

    A person and three agents on one card blend into one profile.

    Mandate

    The instruction an agent was given never reaches the bank's data.

    New threats

    Prompt injection, hijacked agents and runaway loops look like ordinary payments.

    Timing

    AML checks mostly run after the fact. Agents pay many times a minute.

    False positives

    Fast, repetitive agent payments trip velocity rules built for humans.

    How it works

    Four steps between an agent and the money.

    One API call before each payment. We sit alongside your existing transaction monitoring and case tools, not in place of them.

    1

    Register the agent

    Each agent gets an ID, an owner and a mandate: spend limits, allowed merchants and categories, and a time window.

    agt_07 · owner: Northwind Travel limit $500 / trip · travel only
    2

    Check before it pays

    Every payment request is scored in real time against the mandate and the agent's normal behaviour.

    POST /v1/checks → approved · held · blocked
    3

    Monitor continuously

    Velocity, drift from scope, sudden changes in behaviour, and sanctions screening on every payee.

    drift: 0.82 vs baseline payee screened · no match
    4

    Hand off with context

    Attributed events go to your monitoring and case tools, with an audit trail a regulator can follow.

    which agent · for whom under which instruction
    Try a mandate

    Set the rules. Watch a procurement agent's day.

    Change the limits and see each payment re-checked. The last two rows stay within the rules, and get caught anyway on behaviour.

    $500
    $2,500
    Allowed categories
    agt_31 · procurement · Tuesday 0 approved0 held0 blocked

      What we catch

      Small payments that look normal, and aren't.

      Limits catch the obvious. Most agent failures stay inside them, so we watch how each agent behaves as well as what it spends.

      Prompt injection

      A web page tells the agent to pay someone. The owner never asked for it.

      instruction.source = page:supplier-reviews.example
owner_instruction = none
      Blocked

      Runaway loop

      A retry bug repeats the same payment until a limit stops it, or nothing does.

      repeats = 6 in 40s · same payee, same amount
baseline = 1 per day
      Blocked

      Scope drift

      A travel agent starts buying software. Each payment is small. The pattern is new.

      category = software · mandate = travel
drift = 0.74 over 3 days
      Held

      Hijacked agent

      Same agent ID, different behaviour: new hours, new payees, new amounts.

      active_hours = 02:00–04:00 (usual 09:00–18:00)
new_payees = 4 today
      Held

      Sanctioned payee

      Every payee is screened before money moves, with a case file ready if it matches.

      screening = possible match, 94% name similarity
case = opened, evidence attached
      Blocked
      Cross-rail

      One audit trail, whichever rail the agent uses.

      Agents already pay through at least six protocols that don't talk to each other. Network tools see their own rail. We give you one view and one record across all of them.

      CardsVisa, Mastercard agent tokens
      StripeAgentic checkout, machine payments
      x402Pay-per-call over HTTP
      StablecoinsUSDC and on-chain wallets
      Bank transfersACH, SEPA, Faster Payments
      Early access

      Tell us what your agents pay for. We'll build for that first.

      We're working with a small group of teams before launch. Your use case decides which rails, checks and integrations we ship first.

      • First access to the API and sandbox
      • A direct line to the people building it
      • A say in the roadmap before it's fixed
      Enter a work email, like name@company.com.
      Add your company name.
      Add your role.
      Takes about two minutes.
      Questions

      Before you sign up.

      Is HalvorAI live?

      Not yet. We're onboarding a small early-access group first and shaping the product around their use cases.

      Does it replace our transaction monitoring?

      No. We add the agent layer your monitoring can't see, then send attributed events into the tools you already use, such as your TM system and case manager.

      Which rails come first?

      Cards and Stripe first. x402, stablecoins and bank transfers follow, in the order early-access teams need them.

      Will it block payments by mistake?

      You start in monitor-and-alert mode. HalvorAI flags what it would hold or block, and you decide when to let it act.

      What does it cost?

      Pricing isn't set yet. There will be a free developer tier. Tell us in the form what would make it worth paying for.

      Who is it for?

      Teams building agent payments, AI products whose agents spend money, and companies running agents for procurement or finance.