netdiag MCP

netdiag-mcp

English | 日本語

MCP server for on-demand network diagnostics — DNS lookups (with a DNSSEC AD-bit check), ping, an mtr-based path report, TCP port checks, HTTP status/redirect checks, TLS certificate inspection, and WHOIS, all from one server.

Built for triaging "can't reach X" / "is DNS propagated yet" reports without shelling into a jump host for each one-off dig/ping/curl.

Tools

All tools are read-only and single-target (no batch/sweep mode) — this is a convenience wrapper around checks an operator would run by hand, not a scanning tool. nmap-style multi-host/multi-port scanning is intentionally out of scope; deliberately probing many hosts or ports is a different, higher-blast-radius action that deserves its own tooling and approval flow.

tcp_port_check, http_check, http_get and tls_cert_check use Python's own socket/ssl/httpx stack rather than shelling out to nc/curl/openssl, so those four tools work even on a host with only the dig/ping/mtr/whois binaries installed (or none of them — health_check reports which are missing without failing the whole server).

dns_lookup/dnssec_check support DNS-over-TLS and DNS-over-HTTPS via transport="dot"/"doh" (dig's +tls/+https). This needs dig from BIND 9.18+ — an older dig rejects the flag outright rather than silently falling back to plain DNS, so a stale binary fails loudly instead of giving a false sense of having checked over an encrypted transport.

tls_cert_check/http_check against a bare IP address can fail TLS handshake with a "handshake failure" or similar error on SNI-hosted / CDN-fronted origins (e.g. behind Cloudflare) — TLS's SNI extension only carries hostnames, so an IP literal can't route to the right certificate on a shared edge. This is normal TLS behavior, not a tool bug; check by hostname when the target is CDN-fronted.

Setup

1. System dependencies

dns_lookup, dnssec_check, ping_host, traceroute_path and whois_lookup shell out to dig, ping, mtr and whois respectively. Install whichever of these you want available:

# Debian/Ubuntusudo apt install dnsutils iputils-ping mtr-tiny whois

mtr needs raw-socket access. Debian/Ubuntu's mtr-tiny package grants cap_net_raw to the mtr-packet helper at install time, so it normally works for an unprivileged service user without further setup — verify with getcap "$(command -v mtr-packet)" if traceroute_path reports a socket permission error. Without that capability, traceroute_path fails cleanly with a ToolError rather than crashing the server.

2. Install

pip install netdiag-mcp# oruv tool install netdiag-mcp

3. Claude Code (plugin)

This repository doubles as a single-plugin marketplace, so Claude Code can install the server for you:

/plugin marketplace add shigechika/netdiag-mcp/plugin install netdiag-mcp@netdiag-mcp

The plugin launches uvx netdiag-mcp. No environment variables are required — the only prerequisite is the system dependencies above, and the TCP, HTTP and TLS checks work even without them.

uvx must be on the PATH of the process that runs Claude Code — a login shell usually has it, but a GUI-launched app may not; install uv system-wide if the plugin fails to start.

4. Claude Code (manual)

claude mcp add netdiag -- netdiag-mcp

No environment variables are required.

5. Codex

codex mcp add netdiag -- netdiag-mcp

Any other MCP client that launches stdio servers works the same way: give it netdiag-mcp as the command.

Optional: credentials for http_get

http_get can send an Authorization header to URLs under a configured https prefix, for reading a protected JSON endpoint without handing the credential to the model. Set NETDIAG_HTTP_AUTH to a JSON object mapping prefix to header value:

NETDIAG_HTTP_AUTH='{"https://reports.example.ac.jp/daily/": "Bearer <token>"}'

• The header is chosen per request and per redirect hop by comparing the parsed URL (https only, exact host and port, path prefix on a / boundary). A redirect to anywhere else drops it, along with any cookie set while it was attached. user@host URLs never match, and neither does a path with anything beyond letters, digits, ., _, ~, - and / (no %, ; or ./.. segments), so a server cannot normalise a matched URL out of the prefix. The longest matching prefix wins; two prefixes that differ only in case, :443 or a trailing dot are rejected as duplicates. Values must be printable ASCII.
• netdiag-mcp never puts the value into its own output, errors or health_check; the result's status line says auth=applied when the final hop carried it, and health_check lists the configured prefixes in canonical form (not the values). Errors name a bad prefix by position, not by text. A response body under the prefix has the value (and its token part) masked as [redacted], but an endpoint that transforms it (for example base64) would still reveal it, so point a prefix only at endpoints you control.
• A malformed value fails loudly: http_get returns an error and health_check reports degraded.
• It is read from the environment only. If several instances run as the same user (for example one open to many users and one restricted), set it only in the environment of the instance that should send it — a file would be shared by all of them.

CLI

netdiag-mcp --version   # print versionnetdiag-mcp --check     # report missing binaries and a malformed NETDIAG_HTTP_AUTH (exit 0 when all binaries are present and the setting, if any, parses)

Security notes

• Every external-binary call passes an argv list (never a shell string), so no tool argument can break out into shell syntax.
• Hostname/IP and port arguments are validated and size/range-clamped before use — tool input is model-driven and treated as untrusted, the same as any other tool-calling surface.
• tcp_port_check connects to exactly one host:port per call; there is no loop or range argument, by design.
• The optional http_get credential never leaves the configured https prefix and netdiag-mcp never prints it; a response body that echoes it is the server's doing (see above).

Development

Live smoke test

Unit tests check logic against fixtures; they cannot tell you that a tool has stopped returning real data (a dead dig/ping/mtr/whois binary, a broken TLS trust store, a network that blocks outbound ICMP). scripts/ smoke_test.py runs every registered tool against real public endpoints and fails on empty, malformed or error answers:

uv run python scripts/smoke_test.pyuv run python scripts/smoke_test.py --only ping --traceback

• No inventory, so every target is a fixed public endpoint — Cloudflare's 1.1.1.1 and IANA's example.com (reserved for documentation/testing use, RFC 2606). This server takes no config and has nothing to discover a target from, unlike a device-fleet MCP server in this family.
• tests/test_smoke_probes.py is the offline half: it only checks that every registered tool has a probe spec (and vice versa), so CI catches a tool added without deciding how anyone would know it works, without needing network access.

License

MIT