Bgp Registry Intel

BGP & Registry Intelligence

MCP Server

• Server: bgp-intel-mcp (NetClaw-authored, spec 081 / roadmap R9)
• Command: $BGP_INTEL_MCP_CMD
• Transport: stdio
• Credentials: none. Every source is a public unauthenticated API
• Mode: read-only. There is no write path

The one rule that matters most

RPKI not-found is NOT invalid.

Most of the internet has no ROA. Unsigned address space is the overwhelmingly common case, not an anomaly. If you report not-found as a hijack, a misconfiguration, or a security finding, you will generate false incidents at scale and the operator will stop trusting you.

The two invalid reasons are different findings. as means someone else is announcing your space; length usually means you announced a /24 under a /22 ROA. Never collapse them.

validation_unavailable is not not_found. If the validator is unreachable you get the former, and the RPKI state is genuinely unknown. Do not infer "unsigned" from "could not ask", and do not fall back to guessing from routing or registry data.

Tools (10, all read-only)

Three more "this is not what you think it is"

• Registry data is allocation, not routing. RDAP tells you who address space is registered to. It says nothing about who is announcing it. Use routing_announced_prefixes for that. Treating an RDAP holder as a routing fact is the same category error as treating FortiManager intent as device state.
• PeeringDB is self-reported. No record means nobody filled in the form — not that the network does not peer. Many networks peer extensively and publish nothing.
• Visibility is RIPE's collectors, not the internet. Low visibility has legitimate causes: scoped announcements, no-export, anycast, a recent change. The tool will never call it a leak, and neither should you without more evidence.

Every response carries its source

{ "source": "rpki-validator.ripe.net", "retrieved_at": "...", "outcome": "ok",
  "cached": false, "cache_age_seconds": null, "data": {...}, "caveats": [...] }

caveats carries the statements above — read them, they are not decoration. resource_report gives each section its own source; never attribute one section's data to another's origin.

Failures name the source that failed. source_unavailable is never "no record found" — a dead API and an empty registry are different facts.

Workflow: an unfamiliar prefix appears

1. rpki_validate with the prefix and the origin AS. Validation is always of the pair.
2. Read the state carefully against the table above. If not_found, stop treating it as a problem.
3. registry_lookup — who holds it, and who to contact.
4. routing_announced_prefixes on the origin AS — is this consistent with what it normally announces?
5. peering_network — is this a transit provider, content network, or enterprise? It changes what "normal" looks like.
6. Escalate only on invalid, and only after checking that the origin AS is what you think it is and the ROA is current. This skill does not declare incidents — that is your judgement.

Workflow: who do I complain to?

1. registry_abuse_contact on the address.
2. If none is published, try the covering allocation with registry_lookup.
3. peering_network often has a technical contact when the registry does not.

Boundaries — which tool owns what

Those first two are load-bearing, not politeness. gtrace already does quick ASN and geolocation lookups and this skill deliberately does not duplicate them (Principle VII). For general Atlas/Globalping probe availability by location, or to run any measurement, go to Globalping.

Being a good citizen

RIPE NCC and PeeringDB are volunteer- and membership-funded. The server holds itself to ≤ 4 requests per second per source, strictly serial — even resource_report runs its sections one after another.

Do not use these tools to enumerate, sweep, or bulk-harvest registry data. Look things up in service of a specific operational question. The rate limiter enforces politeness mechanically; the judgement about what to ask is yours.

Repeated lookups come from a cache with per-source lifetimes — RPKI 5 minutes, registry 24 hours — and a cached answer says so and reports its age. Pass fresh=true when a ROA was just published and you need to see through the cache.

Important rules

• not_found is normal. Say so when you report it.
• Keep the two invalid reasons apart. Different cause, different fix.
• Never say hijack or attack. Report state and evidence; escalation is the operator's.
• Private and reserved addresses are refused locally, before any request leaves — sending internal addressing to a public registry is a disclosure even if the lookup fails.
• Single validator. Results are never corroborated; they say so.