Prior

Spend gate for other agents. The rules are already published. Over $100 is denied. Nobody approves it.

Remote MCP endpoint: https://prior-mu.vercel.app/mcp

Tools: decide, buy, confirm_payment. A purchase is a Stripe Checkout. The charge is collected only after Stripe reports the session paid. Payments land in the Stripe account attached to this deployment.

checkout-guard-v1 (backend only)

Abuse guard for the public, no-login endpoints. Nothing on the page changes.

• POST /api/v1/checkout and MCP buy: amount capped at CHECKOUT_MAX_USD (default $100, the same line Prior publishes), at most CHECKOUT_PER_IP_PER_HOUR (10) new sessions per address and CHECKOUT_GLOBAL_PER_HOUR (120) overall → 429 with retry-after. The product name shown on Stripe is cleaned (control characters stripped; links, emails, and phone numbers fall back to "Prior"). Sessions carry metadata[app]=prior.
• GET /api/v1/checkout/{id} and MCP confirm_payment: only real cs_test_…/cs_live_… ids, CONFIRM_PER_IP_PER_HOUR (240).
• POST /api/v1/decide and MCP decide: DECIDE_PER_IP_PER_MINUTE (120); agent/vendor/category ≤ DECIDE_FIELD_MAX_CHARS (120) → 422. In-memory history on Vercel is bounded (1000 decisions, 2000 idempotency keys) and today's spend is a running total, so trimming never under-counts the daily cap.
• GET /api/v1/health: pack marker, limits, and block counters.
• Kill switch: PRIOR_GUARD=0. Limits are per warm instance (in memory).

Patterns, not code: Stripe "Protect yourself from card testing" (rate-limit session creation); anomalyco/opencode PR #45007 (MIT) "rate limit checkout session creation"; revanthrajeev/spendveto (Apache-2.0) burst guard idea. All code here is original.