Short positioning for reviewers and list maintainers.
Pre-execution, local, fail-closed tool-call interception with offline-verifiable Ed25519 receipts.
| Layer | Typical tools | Question answered |
|---|---|---|
| Discovery / posture | Reco, Pluto, Zenity-class | What agents exist and what can they reach? |
| Config scanners | correctover-scan, mcp-guard scanners | Is this MCP/agent config risky before it runs? |
| Egress / proxy firewalls | Pipelock, gateway products | What leaves the network boundary? |
| Runtime tool-call guard | agent-runtime-guard, HOL Guard-class | Can this specific tool call execute now? |
| Kernel / sandbox | eBPF MCP sandboxes | Can the process make this syscall? |
| Agent Runtime Guard | Pipelock-class | |
|---|---|---|
| Primary surface | MCP / tool-call arguments | Mediated HTTP, WS, MCP, tunnels |
| Install | npx MCP server |
Binary / proxy topology |
| Receipts | Ed25519 per decision | Mediator-signed action receipts |
| Best when | Agent already has MCP tools attached | You control egress path |
Complementary: proxy stops bulk exfil on the wire; we stop the bad tool call before it runs.
Scanners find bad configs. We enforce at call time. Use both.
Discovery does not block a single curl|sh. We do not map your SaaS agent estate. Use both.
- npm:
npx -y agent-runtime-guard - Census data: [mmcp-atlas-census (229,827 tools, CC0)
- Standard draft: draft-correctover-ccs
- Not “blocks 100% of attacks”
- Not a full SaaS posture platform
- Not an IETF standard (personal Internet-Draft only)