Mcpreflight

mcpreflight

Point it at any MCP or x402 endpoint and get a verdict in one command. Read-only: no API key, no wallet, no tool calls.

npx -y @eidonze/mcpreflight schema https://your-mcp-server.example/mcp

Status: PASS | FAIL | UNKNOWN   exit 0 | 1 | 2, so CI can gate on it

Reach for it when an agent client, a buyer, or a CI job is about to depend on an endpoint whose source you cannot read, including endpoints you do not own. It reads only what the endpoint already sends to any anonymous caller: the initialize handshake, tools/list, and the HTTP 402 / x402 payment document.

No install to try it: run the check in the browser, free and without signup: check-your-mcp-endpoint

Listed in the official MCP Registry as io.github.xka0085-byte/mcpreflight (browse) and on the GitHub Marketplace.

What it is

Read-only MCP and x402 preflight: it reads initialize + tools/list and x402 payment documents. No API key, no wallet, no payment, no tool calls.

This is an endpoint preflight, not a source audit. Tools that read your repository or fuzz a server you launch need a checkout and a start command; this one needs only a URL, so it can check the deployment an agent is about to connect to, including one you do not own.

Same name, different job

Search "MCP preflight" and you also meet kenimo49/mcp-preflight, now published as mcp-scorecard. Both check an MCP server before something depends on it; the input and the findings differ.

One grades the source you have; the other checks the endpoint you are about to depend on.

Check an endpoint you do not own

The same read-only check runs weekly against every remote endpoint listed in the official MCP Registry, and the results are published as pages and as JSON:

• Trust Directory — 35,325 endpoints, one report each: status, protocol version, tool count, findings, and whether credentials are required.
• JSON API — /trust/api/{slug}.json per endpoint, plus batch indexes and a JSON Schema. No key, no signup, no rate limit.
• x402 census dataset — the payment slice of the same sweep: 1,583 of 35,327 endpoints advertise x402 or HTTP-402, with a weekly series, a status distribution against the registry-wide baseline, and downloadable CSV/JSON.

Read the vocabulary before quoting any number: in the 2026-11-02 scan, 20,701 endpoints answered anonymously (PASS), 10,236 refused with 401 because they require credentials by design (AUTH_GATED), 167 were reachable but invalid (FAIL), and 4,223 gave no anonymous handshake (UNKNOWN, usually an auth wall we cannot see past). A "dead MCP servers" count taken from this data would be wrong — 41% of the registry cannot be checked without credentials at all.

Use it as an MCP server

The stdio entry point is bin/mcpreflight-mcp.mjs, published as the mcpreflight-mcp bin in the same npm package. It implements the MCP lifecycle a client expects: initialize returns serverInfo mcpreflight with the package version, notifications/initialized is accepted, and tools/list returns two tools.

Add it to an MCP client:

{

"mcpServers": { "mcpreflight": { "command": "npx", "args": ["-y", "-p", "@eidonze/mcpreflight", "mcpreflight-mcp"] } } }

Both tools do the same work as the CLI and return the same findings, so an agent can check an endpoint before it connects to it. Read-only by construction: the transport issues only GET and POST to the URL you pass, holds no key and has no wallet.

Seen in use

• joxo serves an MCP endpoint that is auth-gated by design. Their maintainer read the 0.2.0 source to confirm the check was two plain POSTs with no dependencies, then ran npx @eidonze/mcpreflight@0.2.0 schema https://joxo.ai/api/mcp and got UNKNOWN: MCP initialize returned HTTP 401 against a deliberate OAuth 2.1 gate (#50). That report is the reason 0.3.0 classifies an OAuth challenge as AUTH_GATED, with the RFC 9728 resource metadata verified, instead of reading as a dead endpoint. They re-checked on 2026-11-01 and closed it: "The endpoint is sign-in only, so no anonymous handshake is expected. Thanks for the AUTH_GATED handling; closing." (#50).
• A maintainer of cueframe-mcp reran the scan after a fix and reported the result back: @eidonze/mcpdoctor@0.1.8 (the earlier npm name) now exits 0 on their endpoint, with the anonymous tools/list payload down from 1,399,189 to 550,377 bytes — under the scanner's 1 MiB limit (#11).
• Listed in awesome-mcp-lists under MCP Toolkits / Testing Tools (#226).

These are examples of the check being run, not endorsements or a claim that a PASS makes a server safe.

Start here

Exit codes are stable for CI: 0 PASS, 1 FAIL, 2 UNKNOWN/network, 3 usage. Requires Node.js 18+.

mode is schema (MCP handshake plus tool schemas), inspect (one-off x402 preflight), baseline (record the current 402 challenge shape) or drift (fail the job when that shape moves away from the committed baseline). The report is also appended to the job summary, and steps. .outputs.status carries the exit code so later steps can branch on it without parsing logs.

Find the right guide

• Public documentation site
• MCP schema checker
• x402 endpoint inspector
• x402 payments in the MCP Registry: census dataset
• MCP security CI
• Troubleshooting
• OAuth scope and current limits

Run it now

Check an MCP server without calling any tools:

npx @eidonze/mcpreflight schema https://your-mcp-server.example/mcp --json

Check an x402 endpoint without paying:

npx @eidonze/mcpreflight inspect https://your-api.example/paid --method=POST --json

npx @eidonze/mcpreflight inspect https://your-api.example/paid --x402 # payment layer only, one request

Requires Node.js 18+. No API key or wallet is needed.

Example schema run against a live endpoint:

$ npx @eidonze/mcpreflight schema https://your-mcp-server.example/mcp

MCP Tool Schema Report

• Status: FAIL
• Endpoint: https://your-mcp-server.example/mcp
• Protocol: 2025-06-18
• Tools: 12

• FAIL REQUIRED_PROPERTY_UNDEFINED: tools[3] requires undeclared property query
• WARN PROPERTY_DESCRIPTION_MISSING: tools[7].limit has no description

The exact same run against a healthy server prints Status: PASS and exits 0, so CI can gate on it.

Which command?

Exit codes are stable for automation: 0 PASS, 1 FAIL, 2 UNKNOWN/network error, 3 usage error.

What it does NOT check

mcpreflight is a protocol- and schema-level preflight, not a security audit. It does not:

• Detect malicious code, prompt injection, or supply-chain risk in the server's implementation
• Call tools, so it cannot prove runtime behavior, data handling, or permission scope
• Bypass, test, or evaluate authentication and authorization
• Verify payment settlement or delivery for x402 endpoints
• Certify compliance with any spec beyond the response shapes it inspects

A PASS means the observed handshake and schemas were well-formed at scan time. Nothing more.

GitHub Actions

Add a read-only check to pull requests:

name: MCP Trust Check

on: [pull_request]

jobs: mcp-trust: runs-on: ubuntu-latest steps:

• uses: xka0085-byte/mcpreflight@v1 with: endpoint: https://your-mcp-server.example/mcp mode: schema format: markdown

Use mode: x402 when you only care about the payment layer: one request, the 402 challenge and its accepts[], no handshake, no discovery probes. Use mode: inspect for the full report. The action fails on FAIL or UNKNOWN by default; set fail-on: false for an informational check. Looking for it in the Marketplace? Search "MCP Trust Preflight" or "mcp server test", or browse the mcp-trust-preflight listing.

Starter template

Start a new MCP project with this check already wired in:

• Secure MCP Server Starter

The check is read-only and should target a public test endpoint. Do not place private keys or credentials in workflow inputs.

What the report means

PASS means the observed response matched the checks in this version. It is not a security audit, protocol certification, payment-success guarantee, or proof that an endpoint is s