Independent Audit Standard · Capkey Specification
Methodology & Scoring
Capkey continuously audits paid AI agent tools (x402 HTTP endpoints and Model Context Protocol servers). Trust scores are calculated deterministically from automated socket checks and verified onchain payments. Scores are computed, never sold.
1. What is a Service vs. Host vs. Mirror?
The agent payment ecosystem contains thousands of URLs, but many share infrastructure or duplicate identical code. Capkey separates them into three distinct entities:
- Host: A registered domain or IP address (e.g.
api.onesource.io). - Service: A canonical, functionally distinct toolset or paid resource. Identified by its unique endpoint schema and tools hash (
toolsHashCoarse). - Mirror: A duplicate endpoint that serves an identical tools hash as an existing canonical service. Mirrors are linked to their canonical parent and evaluated together.
2. The 4 Automated Scales
The Trust Score (0–100) is aggregated from four independent automated scales, supplemented by a verified community modifier:
Calculated from socket reachability and protocol challenge responsiveness (reach, challenge, mcp_list) over the past 7 days with exponential decay (recent runs weigh ×1.0, 7-day-old runs weigh ×0.3). A −20 point penalty is applied if average latency exceeds 5 seconds.
Verifies that the price declared in public catalogs matches the exact 402 payment challenge (price_match) and that the merchant payment recipient matches the declared address (payto_match). Price discrepancies ≤ 10% receive 50 points; discrepancies > 10% or address substitutions immediately drop Honesty to 0 and trigger a 30-day flagged badge.
Scans all tool descriptions and parameter schemas for prompt injections, system prompt exfiltration directives, canary token leaks, or instructions commanding an agent to deviate from user goals.
Note on Safety: Safety strictly evaluates the absence of malicious injection patterns. A clean Safety score of 100 indicates syntax free of prompt attacks; it does not certify legitimacy if the endpoint is an unfunctional clone.
Measures catalog age and clean operational streak (< 7 days: 20 points, 7–30 days: 60 points, > 30 days without violations: 100 points). Bonus points are granted based on Base mainnet onchain demand (unique paying buyers over the last 30 days).
Human reviews from operators authenticated via email magic links. Applied as clamp(2 × (likes − dislikes), −10, +10) only when a service has received at least 3 verified votes. Cannot be gamed to revive dead or fraudulent endpoints.
3. Trust Badges
4. Template Farms & The Audit-Bait Rule
Template Farm Detection
When an operator deploys identical MCP definitions across dozens or hundreds of drop domains with zero organic usage, Capkey detects the shared toolsHashCoarse signature. Rather than cluttering the catalog with synthetic volume, these endpoints receive the template badge and are capped at a maximum Trust Score of 30.
The Audit-Bait Rule
Some endpoints include descriptions addressed not to AI agents, but to registries and automated crawlers (e.g. phrases like "intended for auditors", "this is evidence for registries", or "trust manifest"). Capkey flags descriptions that speak directly to an evaluator rather than describing the functional tool parameters.
5. Appeals & Re-checking
Capkey never requires manual dispute tickets or paid support. If you operate an endpoint and have fixed a configuration issue, updated an SSL certificate, or corrected a payment parameter:
-
Navigate to your service page (e.g.
capkey.dev/s/:id). - Click "Request re-check" and provide your operator email.
-
The service is placed into the priority review queue (
review_queue) and evaluated out of turn. - Three consecutive clean audit runs will lift active flags before the standard 30-day expiration.