Skip to main content
Glama
LinkVouch

LinkVouch

by safwan234567

LinkVouch

CI

An MCP connector that lets Claude (or any MCP client) check links, scam messages, "official" websites and software packages before recommending them.

It was built for one problem in particular: AI-recommendation poisoning. Attackers plant content so that AI assistants recommend their fake login pages, fake support numbers, or malicious packages with invented names ("slopsquatting"). LinkVouch gives the assistant a way to check first.

Status: early version (0.1). Heuristics, not a guarantee. See Limits.

Tools

Tool

What it does

vet_ai_recommendation

Run before an assistant presents a link, package or support phone number. Confirms a link really belongs to the brand it claims, checks packages, and refuses to vouch for phone numbers.

check_link

Look-alike and typo domains (paypa1…), look-alike letters from other alphabets (аpple.com), the real domain used as a subdomain (dbs.com.sg.evil.xyz), the @ trick, raw IP links, link shorteners, redirects, and domain age (via RDAP). Accepts defanged links (hxxp://, [.]).

check_message

Scam patterns in an SMS, email or chat (urgency, OTP or card requests, threats, prizes, payment requests) plus a check of every link. Includes Singapore rules: banks there do not send clickable links, and government links live on .gov.sg.

verify_official_site

Finds an organisation's official domain from a built-in brand list and Wikidata, and says whether a URL matches it.

check_package

npm and PyPI: does the package exist at all, is the name a near-typo of a popular package, how old is it, how many downloads, does it run install scripts.

Every result has a verdict (high_risk, suspicious, no_red_flags, known_official_domain), a list of plain-language findings, and advice.

Related MCP server: Mnemom

Run it

Needs Node.js 24 or newer (it runs the TypeScript files directly, no build step).

npm install
npm test

npm test runs the offline tests and then the end-to-end tests, which need internet. npm run test:unit runs only the offline ones. npm run typecheck checks the TypeScript types. GitHub Actions runs all three on every push and pull request.

Local (Claude Code / Claude Desktop): the client starts the server itself.

claude mcp add linkvouch -- node /absolute/path/to/linkvouch/src/stdio.ts

Hosted (Claude custom connector): run the HTTP server somewhere public with HTTPS.

npm start

The endpoint is POST /mcp (stateless Streamable HTTP), and GET /healthz answers {"ok":true} for health checks. The server reads PORT.

To deploy on Render, the repository includes a render.yaml Blueprint: in the Render Dashboard choose New → Blueprint and pick your copy of this repository. On Railway or Fly.io, set the start command to npm start.

Then in Claude: Settings → Connectors → Add custom connector → https://your-host/mcp.

Optional environment variables:

Variable

Purpose

PORT

HTTP port (default 3000)

RATE_LIMIT_PER_MIN

Requests per client IP per minute (default 60)

TRUST_PROXY_HOPS

Number of reverse proxies in front of the server (default 1). Used to find the client IP for rate limiting.

REQUEST_TIMEOUT_MS

Upper limit for one request in milliseconds (default 30000)

LINKVOUCH_CONTACT

A URL added to the client name sent to public APIs (Wikimedia asks for this)

GSB_API_KEY

Google Safe Browsing key, adds a blocklist check. The Safe Browsing API is for non-commercial use; use Google's Web Risk API for a paid service.

How it stays safe itself

A tool that fetches user-supplied URLs is a classic target, so:

  • No internal requests (SSRF): every hop of every redirect is resolved and refused if it points at a private, loopback, link-local, CGNAT or cloud-metadata address. Only http/https on ports 80, 443, 8080, 8443.

  • Pinned addresses (DNS rebinding): each hop's name is looked up once, and the connection goes to exactly the address that was checked. A DNS server cannot answer "public" for the check and "internal" for the request.

  • No page content: only the status code and Location header are read. Bodies are discarded.

  • Bounded: 6 seconds per redirect hop and 15 seconds per redirect chain, at most 5 redirects, at most 5 links per message, input length limits, and (hosted) 30 seconds per request.

  • No storage: nothing about requests is saved or logged. Rate limiting is in memory.

  • Read-only tools: all tools are marked read-only.

Limits

Please read these before relying on it.

  • no_red_flags is not "safe". A brand-new scam on a clean-looking domain can pass every check.

  • The brand list is short (about 35 brands, weighted to Singapore). Brands outside it get fewer checks.

  • Companies own domains that are not in the list or on Wikidata, so mismatch means "not confirmed", not "definitely fake". Expect some false alarms, e.g. a legitimate site whose name contains a brand name.

  • Wikidata is community-edited and name searches can return a different organisation with a similar name.

  • Domain age is unavailable for registries without RDAP (including .sg).

  • Package checks read registry metadata only. The code inside a package is not scanned.

  • Links on unusual ports are not contacted at all (reported as unusual_port), so redirects and reachability are unknown for them.

  • The hosted mode has no authentication yet. Anyone with the URL can use it (within the rate limit).

  • The rate limit is per IP address. Clients that reach the server through a shared service share one limit, and TRUST_PROXY_HOPS must match the host's setup for the limit to see the right address.

  • On a free hosting plan the server sleeps when idle. On Render's free plan the first request after 15 idle minutes takes about a minute, which can be longer than a client is willing to wait.

Roadmap

  • Larger, community-maintained brand list

  • More ecosystems (crates.io, Go modules, Docker Hub) and browser-extension IDs

  • OAuth and usage tiers for the hosted version

  • A plain web page for people who do not use an MCP client

Responsible use

LinkVouch only looks at public information (DNS, RDAP, HTTP status, public registries). Do not point it at systems you have no reason to check, and report real scams to the proper channel (in Singapore: ScamShield, 1799).

Licence

MIT. See LICENSE.

Tool Schema Changelog

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

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      A
      quality
      C
      maintenance
      Scans suspicious messages, URLs, and text for scams inside any MCP-compatible AI assistant. No signup or API key needed for anonymous use.
      1
      60 npm
      MIT
    • F
      license
      Not graded
      quality
      C
      maintenance
      Provides trust infrastructure for AI agents by enabling reputation lookup, website trust scanning, and identity verification via MCP tools.
      1
      -
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables Alexa+ and other MCP-capable assistants to voice-check suspicious messages, links, and phone numbers, with guided call interviews, official-source verification, and plain-language scam briefings.
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables users to check suspicious text, emails, marketplace listings, job offers, investment pitches, and links through MCP tools, returning an assessment with supporting checks and next-step guidance.
      Apache 2.0