Skip to main content
Glama

mcp-vet

A trust-scan framework for MCP servers: evidence-backed verdicts before you install — usable from any harness.

uvx --from mcp-vetting mcp-vet @modelcontextprotocol/server-fetch   # one command
uvx --from mcp-vetting mcp-vet ./my-server --harness hermes          # verdict + config for YOUR harness
uvx --from mcp-vetting mcp-vet pypi:fastmcp --json --gate            # machine-readable, CI-friendly

Verdicts: SAFE_TO_INSTALL · REVIEW_BEFORE_INSTALL · DO_NOT_INSTALL — every finding carries file:line and plain-English evidence. A verdict is a gate for humans, never a black box.


The flow

  1. Vet any target: npm / PyPI / GitHub / local directory.

  2. Read the verdict + findings (evidence, not vibes).

  3. Configure: pick your harness, paste the emitted config block.

  4. Gate it in CI with --gate (exit 1 on DO_NOT_INSTALL).

Related MCP server: MCPskills

Three surfaces, one engine

Surface

Use

Library

from mcp_vet import vet, Policy → VetResult (verdict, findings, provenance)

CLI

mcp-vet <target> [--json] [--gate] [--strict] [--harness ...]

MCP server

mcp-vet-server — tools: vet_server, vet_directory, get_verdict, list_scanners, list_harnesses, get_config

Targets

npm:pkg            npm registry + tarball        (bare names default to npm)
pypi:pkg           PyPI JSON + sdist/wheel
gh:owner/repo      GitHub metadata + source
./path             local source directory (offline)

Bare names that fail to resolve on npm (e.g. PyPI-only servers like
`mcp-server-time`) automatically fall back to PyPI before reporting failure.

Harness adapters (any harness, paste-ready)

--harness emits the exact config block for your tool:

harness

output

generic

universal mcpServers block

claude-code

claude mcp add <name> -- <cmd> <args>

cursor

.cursor/mcp.json

vscode

.vscode/mcp.json (VS Code servers schema)

hermes

~/.hermes/config.yaml mcp_servers block

--harness all prints every adapter. Programmatic: config_for(harness, ServerSpec(...)).

Policy (calibrated defaults, --strict to disable)

Static analysis must not cry wolf. Defaults, learned by vetting the real ecosystem:

  • Examples/tests/docs are informational — examples/, tests/, docs/ code can't block a package on its own (strict restores raw findings).

  • Fake secrets don't count — sk-test-…, example, xxxx literals are placeholders, not exfiltration.

  • Trusted-host auth is normal — credentials sent to a host named in the file (constant or literal) is client auth; HIGH only when the destination host appears nowhere in the code.

  • Host interpolation → REVIEW — https://${host}/… is a strong signal, but static analysis can't prove the host is user-controlled; a human decides. Strict mode blocks on it.

JSON schema (stable)

{
  "target": "npm:some-server", "kind": "npm", "version": "1.2.3",
  "verdict": {
    "level": "REVIEW_BEFORE_INSTALL",
    "summary": "3 finding(s); 0 high, 1 medium.",
    "findings": [
      {"scanner": "ssrf", "severity": "medium", "message": "...",
       "file": "dist/index.js", "line": 41, "evidence": "fetch(url)"}
    ]
  },
  "provenance": {"version": "1.2.3", "license": "MIT",
                 "source_url": "git+https://...", "source": "npm"},
  "files_scanned": 214, "duration_s": 1.7
}

Scanners

ssrf · exec · secrets · auth · provenance · deps — static, local-first (no telemetry, no cloud round-trips; the cache is a local SQLite file under ~/.cache/mcp-vet).

Verification

python3 -m unittest discover -s tests    # 48 checks: golden corpus is the quality bar

The golden corpus is the contract: known-good fixtures must never carry HIGH findings; malicious fixtures must always block. The ecosystem scan (scripts/ecosystem_scan.py) keeps the tool honest against real packages.

Honest limits

Static analysis has false positives and negatives. A verdict is a gate for humans, not a replacement for them — read the evidence. Sandboxed execution, egress audit-trail, and live CVE lookups are planned for v2.

Available Tools

6 tools
get_configA

Emit the ready-to-paste MCP config block for a harness (generic|claude-code|cursor|vscode|hermes).

Inputs: name (required), harness (required), env, args, +1 more. Returns: object

get_verdictA

Return the cached verdict for a target if fresh, else vet it now.

Inputs: target (required). Returns: object

list_harnessesC

Harnesses with ready-to-paste config adapters.

Inputs: no named parameters. Returns: result

list_scannersC

The static scanners this framework runs.

Inputs: no named parameters. Returns: result

vet_directoryC

Vet a local MCP server source directory (offline). Returns the VetResult dict.

Inputs: path (required), strict. Returns: object

vet_serverA

Vet a remote MCP server (npm:pkg, pypi:pkg, gh:owner/repo, or bare npm name). Returns the VetResult dict: verdict level/summary/findings, provenance, files scanned.

Inputs: target (required), strict. Returns: object

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.5/5.0

    Scored across 6 tools

    Disambiguation5/5

    Each tool has a clearly distinct purpose: vetting remote vs local, retrieving cached verdicts, listing scanners, listing harnesses, and generating configs. No two tools overlap in what they do, and the descriptions reinforce their unique roles.

    Naming Consistency5/5

    All tools follow a consistent verb_noun snake_case pattern: vet_server, vet_directory, get_verdict, list_scanners, list_harnesses, get_config. The verbs (vet, get, list) are descriptive and uniform.

    Tool Count5/5

    With 6 tools, the set is well-scoped for a vetting framework, covering both operational actions and informational queries. There is no bloat or redundancy.

    Completeness5/5

    The tool surface covers the core lifecycle: vetting remote and local servers, retrieving cached results, enumerating available scanners and harnesses, and generating configuration. No obvious gaps hinder a user from completing a vetting workflow.

    Maintenance

    ActivityInactive
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      B
      maintenance
      Scans MCP servers for prompt injection, supply chain attacks, excessive permissions, and code execution risks. Includes an offline blacklist that catches known-compromised packages like LiteLLM 1.82.7/1.82.8 and Trivy with zero latency.
      19
      MIT
    • A
      license
      A
      quality
      C
      maintenance
      Pre-install trust scoring & safety scanning for MCP servers, AI skills & npm packages — 15 signals incl. OSV/KEV/EPSS vuln intel and an auto-gate go/no-go.
      9
      191 npm
      1
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      Scans MCP servers for security vulnerabilities, prompt injection, and tool poisoning, providing risk scores and protection.
      4
      MIT
    • F
      license
      Not graded
      quality
      D
      maintenance
      Enables users to search and audit MCP servers by quality metrics such as installs, security, and activity, helping them find trustworthy servers and avoid ones with no installs.
      -