AI agent runtime security: fail-closed tool-call interception + offline-verifiable Ed25519 evidence.
Agent → Tool Call → [Guard] → Allow / Block → Signed Receipt
npx -y agent-runtime-guard
npx correctover-scan --demo| Repo | What | Why star |
|---|---|---|
| agent-runtime-guard | Pre-execution MCP interceptor, Ed25519 receipts | Product |
| correctover-scan | MCP / agent config scanner | Fast try |
| [mmcp-atlas-census | 42,163 servers · 229,827 tools (CC0) | Data |
Compare layers: COMPARE.md
| Layer | We | Others |
|---|---|---|
| Discovery / posture | — | Reco / Pluto class |
| Config scan | correctover-scan | many scanners |
| Runtime tool-call block | agent-runtime-guard | HOL Guard class |
| Egress proxy | complement | Pipelock class |
We answer: can this tool call execute now, with a receipt?
Not: full SaaS agent inventory.
- draft-correctover-ccs — personal Internet-Draft (not an RFC)
wangguigui@correctover.com · SkillPay audit
中文: 执行前 fail-closed 拦截 + 可离线验证收据。国内网络不稳时,优先用 GitHub + npx 入口。

