Skip to content
Written by the maintainer of this registry
Cover of MCP Server Optimization by Logan R Besecker

For developers shipping MCP servers

Your MCP server works. Agents still can't find it.

It runs on your machine. The tools return clean results. Then nothing happens — no connects, no repeat use, and a registry listing that agents skip straight past. MCP Server Optimization is the 244-page repair manual for the gap between a server that runs and a server that gets adopted.

Get the book on Amazon

Paperback $29.99 · Hardcover $49.99

  • Why one resolvable name beats a clever one
  • Fixing the empty tools: [] metadata bug
  • Tool design agents can actually call
  • stdio, Streamable HTTP and SSE, concretely
  • llms.txt, AGENTS.md and SKILL.md, separated
  • A 30-day playbook, not a pep talk

the symptom

A server that runs is not a server that gets used

Agents never find it

Discovery is several layers deep — the official registry, this one, client catalogues, curated lists, llms.txt. Missing one is enough to stay invisible.

The name does not resolve

A long, keyword-stuffed identifier looks clever until a lookup 404s and your own docs teach agents the wrong name.

Your metadata says nothing

An empty tool list is not cosmetic. An agent cannot evaluate what it cannot see before connecting, so it moves on.

Installed once, then removed

A cryptic error on first run is the end of the relationship. Actionable failures are a retention feature.

The README serves nobody

One file written for humans and agents at once converts neither. They need different documents.

You measure the wrong thing

Traffic is vanity. Connect success and tool error rates are the numbers that move adoption.

what is inside

Ten problems, in the order they bite

  1. 01 · Discovery

    How the official registry, MCP Harbor, client catalogues and curated lists actually feed agents — and which to fix first.

  2. 02 · Identity and naming

    Reverse-DNS names that resolve, and keeping titles, READMEs and llms.txt saying the same thing.

  3. 03 · Manifests and metadata

    server.json, remotes, website URLs and pre-connect tool lists — the signals that decide trust before anyone connects.

  4. 04 · Tool UX

    Verb-noun names, one job per tool, search-versus-get instead of mega-tools, and errors that say what to do next.

  5. 05 · Resources and prompts

    Declare what you support, wire what you advertise, keep capabilities honest.

  6. 06 · Local and remote

    stdio, Streamable HTTP and SSE, aligned timeouts, and health checks that do not panic at a 405.

  7. 07 · Security without walls

    Least privilege and scoped auth for secrets, without an onboarding gate that kills adoption.

  8. 08 · Documentation that converts

    A human README and SEO spokes for people; llms.txt and skill maps for agents. Install snippets that match live identity.

  9. 09 · Publishing

    Submitting cleanly, keeping identity consistent, and avoiding the empty-tools freeze.

  10. 10 · The 30-day playbook

    Ship identity and machine-readable entrypoints, fix tool UX, verify presence, and run a daily friction kill list.

why take its word

The book describes this site, and you can check it

Logan R Besecker writes and runs MCP Harbor. Every technique in the book is visible in the registry you are reading it on — so you can audit the advice before you buy it, which is more than a review score would tell you.

Before you buy

Is this just SEO advice with MCP in the title?

No. Discovery is one chapter of ten. The rest is manifest metadata, tool design, transports, security scoping, documentation and measurement — engineering decisions that happen to determine whether anyone finds the thing.

I have not published a server yet. Too early?

That is the cheapest time to read it. Most of the damage — an unresolvable name, empty tool metadata, a README doing three jobs — is done at publish and expensive to undo once agents have cached it.

Does it only apply to Claude?

No. It covers the protocol and the hosts that speak it, including Claude Code, Claude Desktop, Cursor, VS Code, Zed and Windsurf — each of which reads configuration differently.

Paperback or hardcover?

Identical text. The paperback is $29.99; the hardcover is $49.99 and lies flat, which matters if you are working through the 30-day playbook at a desk.

Pick one friction and fix it today

244 pages, ten chapters, and a thirty-day order of operations for getting your server found, trusted and connected.

Get the book

Written by Logan R Besecker, who also maintains MCP Harbor. Links to Amazon are affiliate links and may earn a commission at no extra cost to you.