MCP server exposing free OSINT sources to AI agents — no API keys required
Strong foundation with 15 well-named tools covering distinct OSINT operations. All tools have clear, detailed descriptions (avg 180 chars) and complete JSON schemas with typed parameters. Naming follows verb_noun convention (wayback_*, crt_*, dns_*, host_*, rdap_*). However, output schemas are not documented, LLMs cannot predict response structure. Error handling is generic (status codes only, no recovery guidance). No tool annotations (readOnlyHint, etc.). Parameter descriptions are good but lack explicit constraints (ranges, patterns). The envelope pattern (ok/err with pivots) is sophisticated but not visible in tool definitions themselves.
Raw Certificate Transparency entries for a domain: issuer, validity window, serial and all SANs per certificate. Use when the certificates themselves matter — spotting a rogue issuer, dating when a host first appeared, or pivoting on a shared certificate. For plain subdomain discovery use crt_subdomains instead, it returns far less.
List unique subdomains of a domain from Certificate Transparency logs. Best first move when mapping an organisation's attack surface — finds hosts that DNS bruteforce misses. Argument is a bare registrable domain ('example.com', not a URL and not a subdomain). Does not verify the hosts still resolve; pair with dns_records for that.
Look up one CVE by id and return severity, CVSS score and vector, description and dates. Argument format is the full identifier, 'CVE-2021-44228'. This is the tool to reach for after host_vulns hands you ids — it tells you which of them actually matter.
Search CVEs by keyword, newest first. Good for 'what is known about product X' when you have a name but no ids. Keep keywords specific — 'apache' returns thousands, 'apache struts rce' returns something you can read. NVD rate-limits unauthenticated callers, so avoid running this in a tight loop.
Resolve one DNS record type for a name. Argument `rrtype` is a standard record type ('A', 'AAAA', 'MX', 'TXT', 'NS', 'CNAME', 'SOA', 'CAA'). Use dns_records instead if you want the usual set at once — it is one call rather than six.
Output schemas not documented. LLMs cannot predict response structure (ok/status/data/pivots/truncated/hint envelope). Tool descriptions mention 'results' but do not specify field names, types, or nesting.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). All 15 tools are read-only but this is not declared in the MCP schema, forcing LLMs to infer safety from descriptions.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | <=2025-11-25 | v2 |
Resolve the common record types for a name in one call: A, AAAA, CNAME, MX, NS, TXT, SOA. This is the normal starting point for a host — it answers 'does this resolve, where to, and who runs it' in a single round trip. Queries run in parallel.
Reverse-resolve an IPv4 address to its PTR record. Tells you what a host calls itself, which often reveals the hosting provider or an internal naming scheme. Many addresses have no PTR — that is normal, not an error.
Open ports Shodan last observed on an IPv4 address, without the rest of the profile. Use when you only need the port list — for example when checking a range of addresses and the CVE data would be noise.
Everything Shodan knows about an IPv4 address in one call: open ports, hostnames, tags and CVEs. The normal follow-up once DNS gives you an address. Data comes from Shodan's last crawl, so treat it as a lead to verify, not a live scan result.
CVEs Shodan associates with an IPv4 address, based on the service banners it saw. These are inferred from version strings, so false positives are common — a backported patch keeps the banner but fixes the bug. Always confirm before reporting anything as vulnerable.
Registration data for a domain over RDAP (the structured successor to WHOIS): creation and expiry dates, registrar, nameservers and status flags. Registrant identity is normally redacted under GDPR — do not expect names or emails. A very recent creation date is itself a signal.
Registration data for an IP address: which organisation holds the netblock, its range, country and allocation type. Answers 'whose infrastructure is this' — cloud provider, hosting reseller, or the organisation's own allocation.
One-shot first pass over a domain: pulls subdomains from Certificate Transparency, resolves the most promising ones, then profiles the addresses found (open ports, CVEs, PTR). Replaces roughly fifteen individual calls. Start here when handed a new domain, then use the specific tools to drill into whatever looks interesting.
List archived snapshots of one specific URL over time, with timestamps and status codes. Use to see when a page changed or disappeared — for example to date when an endpoint was pulled offline.
List distinct URLs the Wayback Machine has archived under a domain. Surfaces deleted pages, forgotten admin paths and old parameter names without sending a single request to the target. Argument is a bare domain; results cover its subpaths.
Parameter constraints not explicit in descriptions. 'limit' parameters lack min/max bounds (e.g., wayback_urls accepts 1-2000 but description does not state range). 'full' boolean lacks guidance on when to use it.
Error handling is generic (status codes only). server.py catches exceptions and returns err(500, ...) with no recovery guidance. LLMs cannot determine if errors are retryable, user-fixable, or fatal.
STDIO transport only. Server is not remotely accessible and cannot be used by hosted MCP clients. Limits deployment to local/CLI-only scenarios.