Anthropic Approved Our MCP Server Without a Human

This week I submitted an MCP server to Anthropic's directory for one of our portfolio companies. I did it by hand on purpose. I didn't want an AI to fill it out and get us rejected.

By the time I started asking Claude how long review usually takes, the listing was already approved. The approval note said it passed on automated checks. No person at Anthropic had to touch it, and that's the reason Anthropic can review connectors at the scale it does.

I think about a lighthouse. One person used to keep the light on for every ship passing by. That's how I expect most companies to work: a single person running what used to take a team. Many will be automated completely, the way most lighthouses are today.

What is an MCP server?

An MCP server is the piece of software that lets an AI like Claude use a company's product: read its data, run its actions and answer questions about it. MCP stands for Model Context Protocol, an open standard Anthropic released in 2024. Anthropic's Connectors Directory lists these servers in one catalog that works across claude.ai, Claude Desktop, mobile, Claude Code and Cowork.

Anyone on a paid Claude plan can submit, with no partner program to apply to first. Anthropic says ranking inside the directory is usage-based, like an app store, and every listing is eligible to be suggested to people inside a chat when it fits their task.

I wrote about where that leaves software companies in software dollars moving from SaaS to custom. The directory is one of the places that shift is already running.

The form checked my answers as I typed

Most application forms wait until you hit submit, and then a person reads what you wrote days later. This one read my answers while I was filling them in.

When I described what the MCP server does, the form compared it against the server itself and flagged where they didn't match. It annotated problems field by field. It felt like a conversation with the form.

The rules behind those flags are public. Anthropic's submission checklist says the portal flags any tool missing a title or a read-only or destructive label. Its review criteria reject a single tool that both reads and writes data. A rule that specific can be checked by software every time.

We rebuilt the server halfway through

Partway through, the portal raised items Anthropic needed to verify. I realized the instructions inside our MCP server didn't line up with what we'd written in the application.

So we stopped, rebuilt the server and refreshed the connector. When I came back, the portal had already pulled in most of the new information on its own. I filled in what was left and hit submit.

The listing existed before it was approved

After submitting, I landed on a draft of our directory listing with an "in review" bar across the top. The portal had built the listing for us. It assumed we'd be approved and showed us the result while we waited.

Anthropic's status page says review time isn't fixed and lists each stage a submission moves through. Ours moved through them before I finished my first question about timing.

How did it approve us so fast?

I don't know exactly what runs behind the portal. My guess is a model built for this one job: it reads the application, visits the URLs we listed, reads the developer docs and contact pages, connects to the MCP server and tests it.

It had been checking our application the whole way through. By the time I hit submit, there wasn't much left to decide.

Why removing people lets Anthropic scale

Think about a normal onboarding flow. Every form, approval queue and support ticket assumes a person reads it. That's why reviews take a week, and why a company has to hire more reviewers as submissions grow.

Anthropic took the person out of the loop. The rules are written down, software checks them while you type, and a passing submission goes live. A thousand submissions cost about the same as ten.

I still did it by hand the first time

Claude could have filled out most of this application. I chose to do it myself because it was our first submission, and I wanted to understand every field before trusting a machine with it.

That's how I approach anything new. Do it manually the first or second time, learn where it breaks, then automate it. But I wasn't needed for this one. The portal could have taken it from start to finish.

What this means for anything purely digital

If a process lives entirely online, people can be removed from almost all of it. Ordering something from a website, signing up for software, logging in, applying for a listing: each one is a set of rules and checks that software can run.

That's why the thesis at Outperform Ventures leans toward companies with a physical connection to the world. I've written about it in physical moats over SaaS. A purely digital workflow is the easiest kind for AI to take over, and I just watched one run without me.

I wrote a companion piece on the bigger picture: dead internet theory is coming for business. If work leaves the internet, what's left for people online?

Questions for your own onboarding

1. Write down every rule your reviewers apply. If a rule can't be written down, ask whether it should exist.
2. Check answers while people type, so problems show up before submit instead of a week later.
3. Show customers their finished account, listing or order before final approval.
4. Keep a person on the cases the rules can't decide, and let software approve the rest.

Our listing was live the same afternoon I submitted it. Most of the steps I took that day could have been done by Claude on our side and Anthropic's checks on theirs.