Getting started
evidence-report-mcp is an MCP server with one tool, verify. An agent calls it before acting on something it did not make: a package it is about to install, a dataset it is about to use, a media file it is about to pass on, or a claim it is about to repeat. It gets back an evidence report: what was checked, where each result came from and when, and what was not checked.
It needs Node.js 22.12.0 or later.
Add it to Claude Code
claude mcp add evidence-report -- npx -y evidence-report-mcpThe tool then shows up as mcp__evidence-report__verify. Ask your agent to use it before it acts, for example:
Before you add lodash 4.17.20, verify it, and tell me what was not checked.
Add it to another MCP client
Clients that read a JSON configuration, such as Claude Desktop, start it the same way:
{
"mcpServers": {
"evidence-report": {
"command": "npx",
"args": ["-y", "evidence-report-mcp"]
}
}
}The server talks over stdio. Checking media files needs c2patool; see Configuration.
Read the answer
A call with a package:
{
"target": {
"kind": "package",
"ecosystem": "npm",
"name": "lodash",
"version": "4.17.20"
}
}returns, shortened:
{
"verdict": "fail",
"summary": "Failed: 1/1 checks completed, 4 finding(s).",
"checks": [
{
"id": "osv.vulnerabilities",
"kind": "deterministic",
"status": "fail",
"source": {
"name": "OSV.dev",
"url": "https://osv.dev",
"retrievedAt": "2026-11-19T11:17:18.524Z"
},
"findings": [
{
"severity": "high",
"message": "Command Injection in lodash",
"ref": "GHSA-35jh-r3h4-6jhm"
}
]
}
],
"notChecked": [
"Unpublished or zero-day vulnerabilities",
"Malicious behaviour without an advisory (typosquats, install scripts)",
"Whether this package is the right one for the job"
]
}Read the verdict first, then the findings, then notChecked. A verdict of unknown means not verified, never safe. The evidence report explains every field.
Use it without MCP
The package also exports verify, aggregate and createServer, and the types of the report and its input:
import { verify, type EvidenceReport } from 'evidence-report-mcp';
const report: EvidenceReport = await verify({
kind: 'package',
ecosystem: 'npm',
name: 'lodash',
version: '4.17.20'
});
if (report.verdict !== 'pass') {
console.log(report.summary, report.notChecked);
}See the API reference for each export.
What it checks
| Target | Input | Check |
|---|---|---|
package | { ecosystem, name, version } | Known advisories from OSV.dev |
dataset | { schema, records, assetId? } | A JSON Schema contract, units included |
file | { path } | C2PA content credentials |
claim | { text, sources } | Independent sources |
What is not there yet
- The dataset check's observability adapter and the claim check's source matching are stubs: they report
skipped, so a dataset or a claim is never verified aspassyet. Their pages say what they do today. - Ideas for later:
verifyover a lockfile or a software bill of materials (the OSV batch API), and a signed report (Sigstore), so that other systems can trust the verifier's output and not only the agent's summary of it.