mcpField
V
vuln-triage-mcp

vuln-triage-mcp

Local
github.com/kimmawlaSecurity0Updated 1w ago
Repository

About vuln-triage-mcp

vuln-triage-mcp is an MCP server for AI agents deciding which vulnerabilities to fix first. It combines CVSS, FIRST EPSS, and CISA KEV data, and reports the basis for its priority score and any missing data.

Key features

  1. 01Look up CVE and GHSA advisories
  2. 02Find vulnerabilities affecting a package
  3. 03Retrieve EPSS exploit-probability scores
  4. 04Check CISA KEV known-exploitation listings
  5. 05Score priority and explain the reasoning

Use cases

  1. 01Rank multiple vulnerabilities for remediation
  2. 02Check vulnerabilities affecting a package version
  3. 03See whether a vulnerability is exploited in the wild

MCP support

It exposes vulnerability lookup, package search, EPSS and KEV checks, and triage tools to MCP clients. Data comes from OSV.dev, FIRST EPSS, and CISA KEV.

Configuration

Local server config is the same for every client. Copy it into your client’s MCP config file.

mcp.json · Universal config
{
  "mcpServers": {
    "kimmawla-vuln-triage": {
      "command": "uvx",
      "args": [
        "vuln-triage-mcp"
      ]
    }
  }
}
Run commanduvx vuln-triage-mcp

Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.

Packages
pypivuln-triage-mcp0.1.0

Tools

No tool list yet

Local servers expose tools only once running on your machine, and the maintainer has not provided a tool list.

Details

From the project README

vuln-triage-mcp

Tells your AI agent which vulnerabilities to fix first, using OSV.dev, FIRST EPSS, and CISA KEV.

PyPI version Python 3.12+ License: MIT Tests

Why this exists

Two or three sentences. CVSS alone ranks by severity, not by likelihood of exploitation. This server adds exploit probability (EPSS) and known exploitation (KEV), and shows its reasoning. Example: a CVSS 9.8 nobody exploits ranks below a CVSS 8.8 that is being exploited in the wild.

Quick start

Claude Code

claude mcp add vuln-triage -- uvx vuln-triage-mcp

Claude Desktop / other clients

JSON config snippet using uvx. Requires: Python 3.12+, no API keys.

Try it

Three example prompts and what to expect:

  • "Which is most urgent: CVE-2021-44228, CVE-2023-4863, CVE-2019-10744?"
  • "Is lodash 4.17.15 vulnerable? What should I upgrade to?"
  • "Is CVE-XXXX actively exploited?"

Tools

ToolArgsUse it whenPrefer another when
lookup_vulnerabilityvuln_idraw advisory details for one CVE/GHSAuser wants prioritization -> triage_vulnerability
find_vulns_for_packagename, ecosystem, version="""what vulns affect package X (version Y)"user already has an ID
get_epss_scorecve_idexploit-probability onlyuser wants overall priority -> triage_vulnerability
check_known_exploitedcve_id"is this exploited in the wild" (CISA KEV)same
triage_vulnerabilityvuln_idany "how urgent / what should I fix first" questionraw details only -> lookup_vulnerability

How scoring works

  • Formula: CVSS (base) x 40 + EPSS x 35 + KEV +25; KEV-listed floors at 80
  • Buckets: see the table below
  • Missing data: component omitted, named in reasoning, lowers confidence
  • Link to docs/scoring-spec.md
BucketWhen
criticalscore >= 80, or listed in CISA KEV (always floored to 80, whatever the confidence)
highscore 60-79
mediumscore 40-59
lowscore < 40
unknownnot KEV-listed, and either no CVSS severity or low confidence (0-1 of 3 signals). Too little data to classify; the score is still returned, and reasoning names what is missing. Not a low-risk verdict.

unknown takes precedence over high, medium, and low; it never applies to a KEV-listed vulnerability.

Reliability and limits

  • Partial results: if EPSS or KEV fails, the result names the missing source
  • Truncation: truncated, found_count vs returned_count, ranking_note
  • Stale KEV cache: stale and catalog_fetched_at
  • Timeouts and retry policy (one line)
  • Package queries return at most 50 advisories, sorted by severity

Security

  • Advisory text is untrusted third-party content (prompt injection vector): sanitized, length-capped, returned under untrusted_text, and flagged to the model in the server instructions
  • Input validation on all IDs and ecosystems
  • No credentials required or stored; read-only calls to three public APIs
  • Reporting vulnerabilities in this project: SECURITY.md or email

Evaluation

Golden set: 17 entries (16 CVEs, 1 GHSA) with expected buckets, confidence, and ordering, recorded 10/21/2026. Result: 17/17 match on the recorded fixtures (uv run pytest tests/test_golden.py), and 17/17 against the live APIs on 10/21/2026 (uv run pytest -m live tests/test_golden.py). EPSS drifts daily, so the live run can diverge over time; re-record the fixtures with uv run python scripts/record_golden_fixtures.py when the golden set or scoring spec changes.

Data sources and attribution

OSV.dev, FIRST EPSS (cite per their usage terms), CISA KEV. Not affiliated with any of them.

Development

uv sync, uv run pytest, running the Inspector, live-test marker.

License

MIT

Comments

Markdown supported

New to MCP?

From what the protocol is to installing your first server, these answers cover most first questions.

An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.