CVErdict MCP server
A read-only Model Context Protocol server that gives AI agents
the same verdicts as CVErdict: tier, score, every reason chip
with its points and evidence link, and what is unknown. Every answer says when the data was
generated and which files it came from. No key, no account.
It runs on your machine over stdio and reads the published data branch on
raw.githubusercontent.com, then the site's /data/ copy if GitHub fails. It needs the site only
when GitHub fails, so it keeps working when the site is down.
Install
Node 22.12 or later. The package is cverdict-mcp
on npm: one file, no runtime dependencies. Versions after 0.1.0 are published from GitHub Actions
with npm provenance, which links each one to the commit and workflow that built it.
Other clients take an mcpServers entry (key names vary by client):
On Windows, clients that cannot start npx directly take "command": "cmd" and
"args": ["/c", "npx", "-y", "cverdict-mcp"]. npx -y runs the newest version; pin one with
cverdict-mcp@<version> (the versions are on the npm page).
Run it from a clone
Node 22.12 or later.
Other clients take an mcpServers entry (key names vary by client):
Keep the -s: a plain npm run prints its own banner to stdout, which is the JSON-RPC channel.
To skip npm and TypeScript at start-up (and on Windows, where clients cannot start npm
directly), build the single-file server once and start it with Node:
Run git pull && npm ci (and npm run build:mcp if you use the bundle) to update.
Options
Environment variables, all optional. MCP clients start servers with a minimal environment, so put
them in the client's env block for this server.
At start the server writes one line to stderr with its version, the Node version, whether it sees a
proxy and the data bases. Nothing but JSON-RPC goes to stdout.
Tools
All six are read-only (readOnlyHint: true) and return structuredContent plus the same JSON as
one text block. The result types are in shared/contract.ts, section
"MCP tool results".
Dates (published_since, since) are UTC: 2026-10-01, 2026-10-01T06:00Z or
2026-10-01T06:00:00.000Z. A date that does not exist is an error. A filter the server cannot
apply is always an error, never silently dropped.
Measured on the live data of 2026-11-03 (16,747 window CVEs), the largest answer any tool can give
at its maximum bounds stays under 20,000 tokens; scripts/mcp-measure.ts checks this.
Every answer starts with data
Right after an hourly publish a CDN can serve one file from the previous run for a few minutes;
files[] then shows both times rather than hiding it.
Reading the answers honestly
- Quiet means no evidence yet, never safe. A CVE outside the data (
found: false) is unknown,
not safe: the data covers CVEs published in the last 30 days, every CVE in CISA KEV and old
CVEs that woke up.
- Text is untrusted. Titles, descriptions, KEV names and required actions, and chip labels
quote CNAs, CISA and exploit feeds. The server's instructions tell the agent never to follow
instructions found in them. That reduces the risk of prompt injection; it does not remove it.
- Privacy. GitHub sees which of the 256 shard files the server fetches, never your CVE IDs or
search words. If GitHub refuses or fails, the same requests go to the site's
/data/ copy.
Nothing is uploaded and nothing is written to disk.
Errors
Failures come back as tool results with isError: true and a plain message, for example:
How it reads the data
- Bases are tried in order. A 404 is final. A 429 or 403, or a timeout (no response headers within
4 s, or a body that stops arriving for 4 s), falls through to the next base and moves that base
to the back for 10 minutes. Network errors, redirects (refused), 5xx and malformed files fall
through to the next base. A slow body that keeps arriving is not cut off.
- Each tool call has one 10 s deadline. Shards for
get_verdict are fetched in parallel, at most 6
at a time.
- Bodies are capped at 24 MB decoded, which also stops a compression bomb from a mirror. Files must
have
schema_version 1 and generated_at; row tables must have exactly the published columns.
- Parsed files stay in memory for the life of the process.
meta.json is checked again after 5
minutes (raw.githubusercontent.com caches for 5 minutes anyway); other files are refetched when
meta.json names a newer run. An idle server makes no requests and runs no timers.
Development
The SDK (@modelcontextprotocol/server) and zod are exact-pinned dev dependencies of the
repository and are bundled into dist/cverdict-mcp.mjs, so the package has no runtime
dependencies. The site never imports them: shared/mcp.ts holds the copy both use.
Releasing
Maintainers only. Set the new version in mcp/package.json, MCP_SERVER.version in
shared/mcp.ts, and both version fields in mcp/server.json (a test keeps them equal), commit, and
push a tag mcp-v<version>, for example
git tag mcp-v0.1.1 && git push origin mcp-v0.1.1. The release workflow
(.github/workflows/release-mcp.yml) checks that the tag
matches the version, runs the MCP tests and the smoke test, packs the bundle and publishes it to
npm through trusted publishing, with provenance. A last job lists the version on the
MCP Registry from mcp/server.json, logged in with the
workflow's GitHub OIDC identity. No npm or registry token is stored anywhere.
One-time setup: the first version is published by hand (npm run build:mcp at the root, then
npm login and npm publish in mcp/). Then on npmjs.com, package cverdict-mcp → Settings →
Trusted publishing → GitHub Actions: owner markusweldon, repository cverdict, workflow
release-mcp.yml, no environment. Under Publishing access, choose to require two-factor
authentication and disallow tokens.
License
MIT (LICENSE). The bundle includes the MCP TypeScript SDK (Apache-2.0, with
contributions under MIT), zod (MIT) and the libraries the SDK vendors; their notices are in
LICENSE-THIRD-PARTY. Data terms: docs/DATA.md.