SponsorFinder MCP Server

An authless, read-only remote MCP server that lets AI assistants (Claude, ChatGPT, Cursor, …) check whether a company holds a UK or Netherlands work-visa sponsorship licence. It proxies the public SponsorFinder API, shaping responses into clean verdicts, and keeps the upstream API key server-side so clients connect with no credentials.

What it does

SponsorFinder tracks two official government registers of licensed work-visa sponsors:

• UK — the Home Office register of licensed sponsors, rebuilt from the CSVs the Home Office publishes (checked daily). Carries routes (e.g. Skilled Worker), ratings (A/B) and locations.

• Netherlands — the IND public register of recognised sponsors, checked daily. Lists recognised sponsors and their sponsor type (WORK / EXCHANGE / STUDY / RESEARCH); much thinner than the UK data — no routes, ratings, or locations.

This MCP server exposes that data as four tools. It is read-only and authless for clients: the upstream x-api-key is a server-side secret that MCP clients never see. There is no OAuth, no per-user state, and no write tools.

Related MCP server: Companies House MCP

Tools

Every tool is annotated readOnlyHint: true and returns both a human-readable text block and structuredContent, each stamped with the data-freshness date and a source note.

Example prompts

Natural-language things you can ask an assistant once the server is connected:

• "Does Google hold a UK sponsorship licence?"

• "Is ASML a recognised sponsor in the Netherlands?"

• "Search UK Skilled Worker sponsors in Manchester."

• "List Dutch WORK-type sponsors matching 'shell'."

• "What does a B rating mean?"

A note on name matching. The fuzzy match tolerates typos (e.g. googel uk), but it needs a reasonably complete name to resolve confidently — "Google UK" or "Google UK Limited" resolves cleanly, whereas a single bare word can match many companies. The server deliberately refuses to guess on weak or tied matches: instead of silently picking one, it returns an ambiguous verdict with the candidate list and asks you to disambiguate (or to call get_sponsor_details with the right id). Absence from a register is itself a meaningful answer: it means the company cannot currently sponsor that visa type.

Use it (hosted)

The server is hosted at:

https://mcp.sponsorfinder.io/mcp

Claude Code

claude mcp add --transport http sponsorfinder https://mcp.sponsorfinder.io/mcp

claude.ai — Settings → Connectors → Add custom connector → paste the URL above. No authentication is required.

ChatGPT — Settings → Connectors (or a custom GPT's Actions) → Add a custom/remote MCP connector and paste the URL above.

Any MCP client that speaks Streamable HTTP can connect the same way — point it at https://mcp.sponsorfinder.io/mcp.

Run locally

Prerequisites: Node.js 24 and pnpm 11.

pnpm install
cp .env.example .env
# then edit .env and fill in:
#   SPONSORFINDER_API_BASE   e.g. https://api.sponsorfinder.io
#   SPONSORFINDER_API_KEY    your upstream x-api-key (server-side secret)

Run one of the two transports in watch mode:

pnpm dev:stdio   # stdio transport (Claude Desktop / Claude Code / MCP Inspector)
pnpm dev:http    # Streamable HTTP on http://localhost:3001/mcp

Add the local stdio build to Claude Code:

pnpm build
claude mcp add sponsorfinder -- node dist/entry/stdio.js

Inspect the tools interactively with the MCP Inspector against either transport:

npx @modelcontextprotocol/inspector

Configuration

All configuration is via environment variables (see .env.example):

SPONSORFINDER_API_KEY is the one true secret. MCP clients connect authless; the key lives only on the server (a Docker env var or a Cloudflare Workers secret) and is never surfaced to clients.

Self-host / deploy

The Web-standard core runs on both Node and Cloudflare Workers; only the entry file differs.

Docker (primary) — image ghcr.io/alirezahamid/sponsor-finder-mcp, config in Dockerfile and docker-compose.yml:

docker compose up -d

Put a reverse proxy (Caddy/nginx) in front of mcp.sponsorfinder.io and pass POST, GET and DELETE through to /mcp. Streamable HTTP needs response buffering off (proxy_buffering off; in nginx; Caddy's defaults are fine).

Cloudflare Workers — config in wrangler.jsonc:

wrangler secret put SPONSORFINDER_API_KEY
pnpm deploy:worker

Development

Analytics (optional)

The server can report usage to Google Analytics 4 via the server-side Measurement Protocol. Because an MCP server has no browser, a GA/GTM JavaScript tag cannot run in it — the server sends events over HTTP instead. It works on both Node and Cloudflare Workers.

Analytics are off unless both GA_MEASUREMENT_ID and GA_API_SECRET are set. Each tool call emits one mcp_tool_call event with categorical parameters — tool, status, verdict, country, mcp_client, latency_bucket, error_kind. By default no company names or free-text queries are sent to GA; the searched name is only included (as the query param) if you opt in with CAPTURE_QUERY_NAMES — see below.

Setup:

1. Create a GA4 property (free) and a Web data stream.

2. In Admin → Data streams → your stream → Measurement Protocol API secrets, create a secret. Copy the stream's Measurement ID (G-XXXXXXXXXX) and the secret value.

3. Set GA_MEASUREMENT_ID and GA_API_SECRET (env / Docker / wrangler secret).

4. In GA4, register the event params above as custom dimensions (Admin → Custom definitions) so they appear in reports. Use GA_DEBUG=true to send to GA's validation endpoint while testing.

Unmet-demand analysis: setting CAPTURE_QUERY_NAMES=true records the searched company name to the server's structured logs and sends it to GA4 as the query event param (register a query custom dimension to see it). Off by default. Caveats: raw names are high-cardinality in GA (bucketed as (other)) and person-named queries may count as PII under GA's terms — a private log store is usually the better home for this, and you should add a privacy-policy line before enabling it.

How it works

The server is a thin, stateless proxy with response shaping. An MCP client connects over stdio or stateless Streamable HTTP; a small Hono app (with @hono/mcp) constructs an MCP server per request, calls the SponsorFinder API with the server-side key, validates every response with zod (a contract-drift guard), and shapes it into a compact verdict. Register stats and filter values are cached in-process with a short TTL. Because the core uses only Web-standard APIs (fetch, URL), the same code runs on Node and Cloudflare Workers — only src/entry/* differs.

MCP client (Claude / ChatGPT / Cursor)
        │  stdio  or  stateless Streamable HTTP
        ▼
SponsorFinder MCP server  (Hono + @hono/mcp)
   • 4 read-only tools, zod-validated
   • x-api-key added server-side
   • cached /status + /filters
        │  HTTPS  (x-api-key)
        ▼
SponsorFinder API  →  UK Home Office register + Dutch IND register

Data & disclaimer

The data comes from the official UK gov.uk register of licensed sponsors and the Dutch IND public register of recognised sponsors, refreshed daily.

This tool is informational only and is not legal advice. Register data can lag official publications, and a licence does not guarantee a company will sponsor any given role. Always verify against the official sources before making decisions.

Contributing

Issues and pull requests are welcome at github.com/alirezahamid/sponsor-finder-mcp. Please run pnpm check before opening a PR.

Publishing note: server.json is the manifest for the official MCP registry. Before running mcp-publisher publish, confirm the exact $schema URL against the current registry.modelcontextprotocol.io docs — the schema version pinned here may have moved on.

License

MIT © 2026 Alireza Hamid