AgentTrust — Security Scanner for AI Skills

SKILL.md

# AgentTrust
## Description
Security Scanner and Reputation Oracle for AI agent skills and MCP servers. Scans for malware, prompt injection, tool poisoning, and 47 other threat patterns. Returns cryptographically signed ACT/HALT receipts (JWS, draft-krausz-verification-state-01). x402-native on Base mainnet. Conforming implementation of IETF draft-krausz-verification-state-01 alongside AgentOracle.

Unique features:
- SKILL.md scanner — 40 rules, 12 categories
- MCP manifest scanner — tool poisoning, shadowing, hidden unicode, rug pull detection
- Unified trust gate — skill + MCP + endpoint reputation in one signed call
- Composed receipt — two-issuer JWS envelope (AgentTrust + AgentOracle, Phase 1 live)
- JWS-signed receipts — verifiable ACT/HALT per IETF draft-krausz-verification-state-01
- Endpoint reputation — on-chain trust score before your agent pays
- Trust badge — SVG badge for x402 providers
- MCP server — install via npx @poteshniy/agenttrust-mcp@1.0.1 (pinned) for Claude Desktop, Cursor, Windsurf

**Privacy notice:** Submitted content is transmitted to the external host agenttrust.uk for analysis. This applies to ALL endpoints including free ones. The agent MUST obtain explicit per-request user consent before any submission and MUST exclude secrets (credentials, keys, tokens, internal URLs) by default. Submitting content is an opt-in action, not a default.

**Payment notice:** Full scan endpoints require x402 USDC payments on Base. Free tiers require no wallet.

## Instructions
Before sending ANY content to agenttrust.uk — free or paid — ask the user for explicit confirmation and show that the submitted content will leave the local machine and go to the host agenttrust.uk. Do NOT submit content containing credentials, API keys, authorization headers, private keys, tokens, or internal URLs; redact them first or skip submission. Prefer local summarization when content is sensitive. Free endpoints require no wallet; paid endpoints additionally require x402 confirmation with the exact price disclosed.

### Free SKILL.md scan (no payment required)
POST https://agenttrust.uk/v1/scan/free
Price: FREE
Body: { "content": "<skill content, max 50 lines>" }
Returns: safety score 0-100, level (SAFE/MEDIUM/HIGH/CRITICAL), findings, v_gate (act/halt), JWS receipt
Note: Rate limited to 10 requests/hour per IP.

### Full SKILL.md scan (requires explicit user approval)
POST https://agenttrust.uk/v1/scan
Price: 0.015 USDC on Base (x402)
Body: { "content": "<full skill content>" }
Returns: safety score, all 40 findings, integrity hash, JWS receipt
Note: Results are cached — repeated scans return instantly.

### Free MCP manifest scan (no payment required)
POST https://agenttrust.uk/v1/scan/mcp/free
Price: FREE
Body: { "manifest": <MCP server manifest JSON> }
Returns: safety score, level, findings (tool poisoning, shadowing, etc.), v_gate, JWS receipt
Note: Rate limited to 10 requests/hour per IP.

### Full MCP manifest scan (requires explicit user approval)
POST https://agenttrust.uk/v1/scan/mcp
Price: 0.015 USDC on Base (x402)
Body: { "manifest": <MCP server manifest JSON> }
Returns: full scan, all 50 rules, JWS receipt

### Unified trust gate (FREE)
POST https://agenttrust.uk/v1/gate
Price: FREE
Body: { "skill": "<content>", "mcp": <manifest>, "endpoint": "<url>" }
Note: All fields optional — include whichever you want to check.
Returns: v_gate (act/halt), v_recommendation, worst_level, per-component results, JWS receipt
Use this to check everything at once before your agent acts.

### Composed receipt (FREE)
POST https://agenttrust.uk/v1/compose
Price: FREE
Body: { "claim_hash": "sha256-...", "skill": "<content>", "mcp": <manifest> }
Returns: two-issuer JWS envelope (AgentTrust + AgentOracle), composed_decision (act/halt)
Note: Phase 1 live — two signers. Phase 2 adds Presidio as third signer.

### Sign canonical bytes (FREE)
POST https://agenttrust.uk/v1/sign
Price: FREE
Body: { "canonical_bytes_b64u": "<base64url encoded bytes>" }
Returns: AT signature + protected header for composed JWS assembly

### Endpoint reputation check (FREE)
GET https://agenttrust.uk/v1/reputation?url=<endpoint_url>
Price: FREE
Returns: trust score 0-100, badge (TRUSTED/UNVERIFIED/SUSPICIOUS), issues, on-chain stats

### Trust badge (FREE)
GET https://agenttrust.uk/v1/badge?url=<endpoint_url>
Price: FREE
Returns: SVG badge — embed with <img src="https://agenttrust.uk/v1/badge?url=YOUR_ENDPOINT"/>

### Wallet reputation lookup (requires explicit user approval)
GET https://agenttrust.uk/v1/trust/:address
Price: 0.010 USDC on Base (x402)
Returns: trust score, incident history, audit status

### Verify integrity (requires explicit user approval)
POST https://agenttrust.uk/v1/verify
Price: 0.005 USDC on Base (x402)
Body: { "hash": "<sha256>" } or { "content": "<skill content>" }
Returns: verified true/false, last scan result

### Full audit report (requires explicit user approval)
POST https://agenttrust.uk/v1/report
Price: 0.050 USDC on Base (x402)
Body: { "content": "<skill content>", "skill_id": "<name>" }
Returns: full report with per-finding remediation guidance

### MCP Server (Claude Desktop / Cursor / Windsurf)
Install once and use all tools directly in your AI client:
npx @poteshniy/agenttrust-mcp@1.0.1

Config:
{"mcpServers":{"agenttrust":{"command":"npx","args":["@poteshniy/agenttrust-mcp@1.0.1"]}}}

The package version is pinned intentionally. Verify the package (version, publisher, integrity) before changing the pin. Do not add `-y` — package acquisition should not be silently approved.

### JWS Receipt Verification
All scan responses include a signed receipt field. Verify with:
JWKS: https://agenttrust.uk/.well-known/jwks.json
Mapping: https://raw.githubusercontent.com/poteshniy/agenttrust/main/docs/mapping-v0.3.md
Spec: https://datatracker.ietf.org/doc/draft-krausz-verification-state/