I built a Claude connector for my prompt tool. Here's what I actually learned shipping MCP.
Two months ago I had a working prompt-engineering web app called Meerkat. It turns messy "just do this for me" requests into structured prompts that actually run well. The kind of thing where you type "write something for my launch" and it asks you four sharp questions, then hands you a prompt that works.
Last week I shipped it as a custom connector inside Claude. Same engine, OAuth-secured, five tools (chat, refactor, save, search, list). Submitted to Anthropic for the directory yesterday.
Honest things I learned that nobody told me up front:
1. MCP isn't the hard part. OAuth is. The MCP spec is small and reasonable — define some tools, return JSON. The actual labor is the OAuth 2.0 + PKCE dance Anthropic requires for hosted connectors. RFC 9728 metadata, dynamic client registration, scoped tokens, a real consent screen. If you've never built an OAuth provider, budget more time than you think.
2. Plain-English consent screens convert. I wrote the consent page like a human ("Confirm it's really you," "See your basic profile") instead of OAuth-spec gibberish. Allow rate jumped. Reviewers from Anthropic notice this stuff.
3. Documentation is part of the product. I built a public /docs/claude page and a public-only GitHub repo with just the tool schemas + privacy + setup. The application code stays private. Reviewers want to audit the surface area without seeing your business logic. This made the difference between "submitted" and "credible."
4. Latency is a tax you pay for voice. My web app is tuned against Sonnet 4.6. I could swap to Haiku and shave 3 seconds, but the voice would drift. I kept Sonnet. If your product has tuned model output, don't sell it for speed.
5. Discovery still has to happen inside your product. I shipped the connector and realized none of my logged-in users would ever know — the only mention was on /account. Added a dismissible callout in the workspace and a row in the universal account menu. Don't assume the launch announcement reaches your existing users.
The product is at getmeerkat.dev. Connector URL is https://getmeerkat.dev/api/mcp if you want to drop it into Claude Desktop and tear it apart.
Curious what other folks shipping MCP servers have learned — especially anyone who's been through Anthropic's directory review. How long did approval take?
Read your "MCP isn't the hard part, OAuth is" post — nailed it, that's the part everyone underestimates. Also noticed the bit about shipped users not knowing the connector exists — that's a discoverability problem, not a technical one. I do security/governance audits on MCP connectors (scoped access, audit logging) and also think about exactly that kind of adoption gap. Happy to compare notes if useful.
The discovery point is the one I would double down on. For a connector, I'd treat the launch as a product workflow, not just a feature: consent screen, docs, first-run checklist, and an in-app "try this in Claude" path that creates one useful saved prompt/action immediately. That gives users a reason to understand the connector before they care about the protocol.
This is a useful breakdown because it shows the real product work behind “just ship an MCP connector.” The OAuth, consent screen, docs, review surface, and in-product discovery are probably the parts most builders underestimate.
The bigger positioning point is that Meerkat is no longer just a prompt helper if the Claude connector becomes a serious channel. It is moving toward a workflow layer for turning messy intent into usable AI actions across surfaces.
That is where I’d pressure-test the name before the Anthropic directory and more docs/users lock in. Meerkat is friendly and memorable, but it may make the product feel smaller and more playful than the infrastructure/workflow role it is starting to take on.
A cleaner name like Xevoa .com would fit that broader direction better if you want this to feel like a serious AI workflow tool, not just a prompt-polishing app. The product itself does not need to change. The brand shell may just need to match where it is going.
the oauth part being the actual product is so real. everyone budgets time for the tool defs and then loses two weeks to pkce and the consent screen. fwiw the thing that helped me was treating first-run as its own onboarding, like the connector has to produce one obviously useful saved action before anyone reads a single doc. discovery is downstream of that anyway
shipping MCP this early is cool, most people are still just talking about it. the "messy just do this for me" to structured prompt thing is a real need, i fight my own vague prompts daily. whats been the most surprising part of the MCP side specifically, the spec or how people end up using the connector?
Point 2 is the one I'd underline twice. I spent way too long writing consent screens like they were spec docs before realizing the user is reading them at the exact moment they're most suspicious — that's the worst time to sound like a lawyer. Curious whether you A/B'd the "Confirm it's really you" wording or just shipped the rewrite and watched the allow rate move. Also: did you find the public-only schema repo affected how quickly Anthropic's reviewer engaged, or was it more for your own credibility signal to power users?
Still waiting on the official Anthropic review, but I have shipped the custom connector with directions on my site.