SecKit MCP

A small, auditable, read-only MCP server that gives an AI assistant defensive security lookup and analysis tools: check a password against known breaches without revealing it, look up a CVE, explain a MITRE ATT&CK technique, parse a log line, and defang indicators safely.

Defensive and educational only. No scanning, no exploitation, no credential testing, no telemetry, no file writes, no shell execution.

Install

Requires Node 20+ (developed on Node 24).

git clone <this repo> seckit-mcp && cd seckit-mcp
npm install
npm run build
claude mcp add seckit -- node /absolute/path/to/seckit-mcp/dist/index.js

Optional: set NVD_API_KEY in the server environment to raise the NVD rate limit (about 5 requests per 30 s without a key).

Related MCP server: Self-Assembling Forensic MCP Server

Tools

Every result is JSON text with source and retrieved_at. Errors return isError: true with a short message.

Resources: seckit://attack/tactics (tactic list), seckit://safety (the safety policy). Prompt: triage_alert (guided what/who/where/when/scope/next-step checklist).

Example prompts

• "What is T1003.001? Explain it and tell me what to monitor."

• "Search ATT&CK for scheduled task persistence."

• "Give me the CVSS and references for CVE-2024-3094."

• "Parse this log line: Jan 5 10:11:12 web01 sshd[1234]: Failed password for invalid user admin from 203.0.113.7 port 51234 ssh2"

• "What does Windows event 4625 mean, and what ATT&CK technique does it map to?"

• "Defang everything in this text so I can paste it into a ticket: ..."

• "SHA-256 of this string, please."

• "Is the string correct-horse-battery-staple in a known breach?" (see warning below)

Safety and limitations

What it does: k-anonymity password check (SHA-1 is computed locally, only the first 5 hex characters of the hash are sent to api.pwnedpasswords.com, the match is local, the password is never logged, cached, echoed or put in errors); one CVE lookup per call to NVD; everything else is local.

Safeguards enforced in code and tests:

• Outbound host allow-list in src/lib/http.ts: exactly api.pwnedpasswords.com and services.nvd.nist.gov, https only, redirects refused. Anything else is rejected before a request is made.

• Timeouts, a descriptive User-Agent, one retry with backoff on HTTP 429, a 1-hour in-memory cache (never written to disk).

• Strict zod input validation with length limits (4 KB log lines, 64 KB IOC text, 128-char passwords).

• Text fetched from NVD is untrusted: it is returned inside a clearly labelled untrusted_fetched_text field and the tool descriptions say so.

• Nothing is written to stdout except MCP protocol frames (tested); no filesystem writes or shell execution in src/ (tested statically).

Limitations (be honest about them):

• Do not paste a real password into a chat. The assistant and your client still see it even though the network call does not. Use a throwaway or sample string, or a client-side checker. A "not found" result does not mean a password is strong.

• The log parser and IOC extractor are heuristics. Bare-domain defanging uses a common-TLD list, and version strings like 1.2.3.4 look like IPs. Unrecognised input is flagged, not guessed.

• The ATT&CK data is a snapshot (version and generation date are in every result) and goes stale; run npm run update-attack.

• explain_event_id covers common events only and is hand-maintained.

• NVD data can lag or be incomplete; CVSS v3.x only (no v4 or EPSS).

• Detection guidance in recent ATT&CK versions (v20 in the committed snapshot) is expressed as detection strategies and analytics; the snapshot keeps the strategy names and up to four trimmed analytics each.

ATT&CK snapshot

src/data/attack.json is generated from the real MITRE ATT&CK enterprise STIX bundle by scripts/update-attack.ts (fields kept: id, name, tactics, description, platforms, data sources, detection, mitigations, url; techniques and sub-techniques; revoked/deprecated skipped). The committed snapshot is the real generated one (version recorded in meta.attack_version). Regenerate with:

npm run update-attack                # downloads from mitre-attack/attack-stix-data
npm run update-attack -- ./enterprise-attack.json   # or from a local copy

The tools work identically against any file of the same shape (a hand-written fixture would carry meta.fixture: true, which is surfaced in results).

Develop

npm test      # builds, then unit tests (mocked fetch) and end-to-end stdio tests; fully offline
npm audit

Layout: src/index.ts (wiring), src/tools/* (one module per tool), src/lib/{http,ioc,logparse}.ts, scripts/update-attack.ts, tests/.

Data attributions

• NVD: CVE data from the NIST National Vulnerability Database (https://nvd.nist.gov). This product uses the NVD API but is not endorsed or certified by the NVD.

• HIBP: Pwned Passwords by Troy Hunt, Have I Been Pwned (https://haveibeenpwned.com/API/v3#PwnedPasswords), used via the k-anonymity range API.

• MITRE ATT&CK: ATT&CK(R) is a registered trademark of The MITRE Corporation. Content (c) The MITRE Corporation, reproduced under the ATT&CK Terms of Use (https://attack.mitre.org/resources/terms-of-use/).

License

MIT, see LICENSE.