I built a free MCP server for testing (so you don’t have to)

A zero-auth, public MCP endpoint for testing agents, MCP clients, and CI/CD evals

While building agents, MCP clients, and writing evals for MCP clients, I kept hitting the same annoying problem again and again.

All I wanted was a simple MCP server — something like a “Hello MCP” — with no auth, no signup, nothing. Just a URL I can point my client to and start calling tools. But I couldn’t find one. And honestly, I didn’t want to spin up a server myself just for basic testing.

Running one on my local machine also didn’t solve the real problem — it stays stuck on my laptop, and I can’t use it in my CI/CD pipeline for eval runs.

At one point I was actually using OpenAI’s MCP server for testing :-) It’s simple and straightforward, but the responses are lengthy, which makes life difficult when you’re using LLM-as-judge for evals. I was somehow managing, but it wasn’t enough.

And what if I want to pass parameters and test? Then again I have to dig through the OpenAI MCP documentation and construct complex parameters. On top of that, the token cost for evals also goes up because I’m hitting an actual, heavyweight MCP server just for testing.

So I built one

To solve this, I hosted a simple MCP server myself — and made it a public endpoint, so any developer or QA facing the same problem can use it too.

**https://mcp.heptafox.com/**

I’d suggest opening the site after reading this article — everything I’m describing below (the tools, the parameters, the copy-paste config for Claude Code and Claude Desktop) is right there on a single page, and it’ll make much more sense once you see it.

Paste the URL into any MCP client (Claude, Cursor, or anything that supports Streamable HTTP) and you’re done. No signup, no auth, nothing to install.

There are two basic tools:

• identity — the simplest possible call, no parameters
• echo — takes parameters and sends your text back inside a fixed wrapper

One thing I deliberately added: every response contains a fixed canary string. So when you’re writing evals, you can assert on the exact string and know the response genuinely came from the server — not something the model hallucinated. This makes LLM-as-judge (or even a plain string match) dead simple.

Testing auth too

I didn’t stop with just the plain server. I added a bearer token check on the same endpoint.

The default tools work without any token, but there’s a dedicated auth_check tool that tells you whether your client actually forwarded the Authorization header or not. The token is public and shared — it's only for testing whether your client's auth wiring works.

Don’t trust a random website? Run it yourself

If you’re in a secured environment behind a proxy, or you simply don’t want to trust some random public endpoint (fair enough!), the full source code is open source under the Apache-2.0 license:

**https://github.com/heptafox/mcp-sandbox**

Clone it, spin it up, and use it locally — or host it inside your own infra and wire it into your CI/CD.

If you feel it needs improvements or more features, please fork the repo and help me improve it. PRs are welcome.