Overview
The Model Context Protocol is how an AI coding agent talks to tools it did not ship with. The Vulnetix MCP server is one of those tools: it puts a vulnerability database of 150+ aggregated sources (CVE.org, NVD, VulnCheck, CISA KEV, GitHub Security Advisories, OSV, EUVD and more) directly inside your agent’s reach.
The problem it solves
Ask any coding agent about a CVE and you get an answer from training data. That answer has no idea whether the vulnerability is being exploited right now, whether CISA has attached a remediation deadline to it, whether a fixed version exists yet, or whether the specific version in your lockfile is affected.
Worse, the failure is invisible. The model answers confidently either way.
With this server connected, the same question routes through live data:
- Is it being exploited? Per-source exploit counts across ExploitDB, Metasploit, Nuclei, VulnCheck XDB, CrowdSec, GitHub PoCs, HackerOne and 13 more, plus dated in-the-wild sighting counts.
- Is there a deadline? CISA KEV membership with the binding due date and required action, alongside the EU and VulnCheck catalogues.
- How urgent, really? EPSS probability, SSVC decision, exploitation maturity, threat exposure, and how long this class of issue historically took to be weaponised.
- What do I change? Ranked remediation actions with concrete steps, distribution patches, registry fixes and workarounds.
What you get
| 31 tools | Vulnerability lookup, exploit intelligence, remediation plans, KEV catalogues, package risk gating, IOCs, sightings, ATT&CK mapping, advisories, lifecycle timelines, threat feeds, full-text search, and an escape hatch to any read endpoint. |
| 33 skill prompts | The complete Pix security-workflow library, delivered verbatim: CVE triage, dependency gating, incident response, compliance bundling, secure-code coaching. |
| Nothing to install | Stateless HTTP. No daemon, no container, no npx shim, no version to keep current. |
What it does not do
It does not scan your repository. A Cloudflare Worker has no filesystem.
Sixteen of the 33 skills need one: SAST, secret scanning, IaC and container scanning, SBOM generation, applying a fix, resolving a dependency conflict, publishing VEX. Those ship as prompts that instruct your own agent to run the Vulnetix CLI locally. Each one says so in its first paragraph.
Why responses look different from the API
The database is built for machines with disks, not context windows. A raw record
for CVE-2021-44228 is 2,606,571 bytes, around 650,000 tokens.
Every tool reduces its response on the server before it reaches you:
| View | Bytes | Of raw |
|---|---|---|
| Raw upstream record | 2,606,571 | 100% |
detail: "full" | 207,604 | 7.97% |
detail: "summary" (default) | 10,613 | 0.41% |
The summary keeps every field a decision rests on and replaces the bulk enumerations (676 affected packages, 499 package URLs, 33 cloud resource locators) with counts, groupings and samples, each pointing at where to page the rest. Nothing you would act on is dropped. See Response shaping for exactly what is kept and why.
Where this sits alongside the other products
| Product | Runs | Best for |
|---|---|---|
| MCP server | Nowhere, you connect to it | Any AI agent, zero install, intelligence questions |
| Pix plugin | Claude Code | Deep local workflows with hooks and slash commands |
| CLI | Your machine and CI | Scanning code, SBOMs, applying fixes, gates |
They share one database and one credential. Using the MCP server does not stop you using the CLI; several prompts assume you have both.