Compute, APIs, SaaS —
without leaking the credit card.

Every coding agent that's ever shipped to production has had at least one near-miss with cost: a runaway training job, an infinite-loop CI run, a model that called itself recursively because the prompt didn't say not to. The blast radius of one of these on a shared corporate card is unbounded until a human notices — usually 24+ hours later when the cost alert fires.

AgentWallet provisions one Visa per service that the agent uses, with a hard daily cap enforced at network authorization time. The OpenAI card simply cannot charge AWS; the AWS card cannot charge Vercel. If the agent goes rogue or its prompt is injected, the maximum loss is the cap on whichever card is in scope — and that cap is enforced by the card network in milliseconds, not by application code that might have a bug.

Many agent workflows want to call an API once, pay for that call, and never see the merchant again — e.g., a one-off geocoding lookup, a one-off LLM judge call, a one-off web scrape. The traditional model (sign up for the API, get a key, pay monthly) breaks for ephemeral or single-call usage.

AgentWallet implements x402 — the proposed HTTP standard for inline payments. The agent hits the API; the API responds with HTTP 402 and a price (in USDC, EUR, USD, or one of 91 currencies). The agent's wallet auto-pays inline (subject to a per-call policy cap) and retries with the proof. Settlement is sub-second on USDC. No subscription, no per-merchant key management, no manual rotation.

The trace ID on each AgentWallet charge is propagated from the agent's call context. If your CI agent is told 'work on PR #4823,' every downstream charge — AWS minutes, OpenAI tokens, GitHub Actions seats — is tagged with PR-4823. At quarter end, your finance team can join AgentWallet's ledger export against your engineering tracker and produce real per-feature unit economics.

This is the difference between knowing 'engineering spent $84k on AI last quarter' and knowing 'feature A cost $2.1k to ship; feature B cost $14.7k to ship and never got merged.' Many teams discover that 5% of their PRs consume 60% of their AI budget — usually from one badly-prompted retry loop — and only see it once attribution is automatic.

If a coding agent is compromised — leaked key, model jailbreak, supply-chain attack on a dependency — the response window matters. AgentWallet's freeze endpoint takes effect within 1 second across all card-network authorization paths, so the maximum additional loss after detection is whatever's already in flight (typically under $50).

Freezes are reversible: POST /api/agents/:id/cards/:cardId/unfreeze restores authorization. The audit ledger records who froze, when, why, and what charges were attempted during the freeze window. This is the difference between an incident postmortem that's 'we lost $4,800 because rotation took 6 minutes' and 'we lost $0 because the freeze landed before the next auth.'