Thinking out loud about AI, architecture, and building things that work
Research, case studies, and notes from the work.
What is a Manifest and well-known
A manifest is the configuration file that tells clients like Claude what an MCP server can do. It declares the tools, resources, and prompts the server offers, plus some basic metadata like its name, version, and description. In short, it’s how an MCP server introduces itself and explains how to talk to it. Until now, you wouldn’t have seen one because clients had to actually connect to a server before discovering its capabilities. That’s changing with the November 2025 spec update, which introduces .well-known URL discovery. This is a big deal. It means MCP servers will be able to publish their manifest in a predictable public location—like how websites use sitemap.xml—so tools, registries, and even search engines for MCP can index what’s out there without needing a live connection. That shift makes the manifest more than just internal config. It becomes the public face of an MCP server—the thing that lets the ecosystem browse, catalog, and connect everything together. If you’ve never heard of manifests before, that’s normal. They’ve been working quietly in the background. But the new .well-known requirement is about to make them front and center in how AI systems discover and connect across the MCP network.
How We Test
LLMs don’t just need to answer questions—they need to choose the right tools when connected to an MCP server. Our testing system is designed to measure exactly that, giving us a clear picture of how well models navigate real-world tool use.
Case Study: How Manifest Alignment Boosted Tool Use Performance from 70% to 83% for Cooking MCP
Large language models don’t just rely on their training—they also depend on how tools are described and exposed through manifests. Even small wording choices in a manifest can make the difference between a model using the right tool confidently or skipping it altogether.
Case Study: How a Single Manifest Change Boosted LLM Accuracy from 33% to 83% in a Key Use Category
## The Challenge We were testing a simple MCP todo server with three tools: createTodo, listTodos, and a completion tool. The original completion tool was named completeTodo with the bare description: “Mark a todo item as completed.”
Deterministic vs. Interpretive MCP Servers: Speaking the Same Language
MCP isn’t just about exposing tools. It’s about creating an efficient and effective understanding of a server in the eye of the ClientAI. This communication is guided by the manifest and, by extension, the initialization and tool calls. For simplicity, we’ll use “manifest” to represent that entire package. In effect, the manifest becomes the dictionary of the language between server and AI.
Potential and Pitfalls: Security Lessons from Real-World MCP Tools
The Model Context Protocol (MCP) ecosystem has exploded. In just months, developers have published hundreds of servers and thousands of tools. It’s a remarkable show of creativity: everything from lightweight utilities that return the time of day to heavyweight servers that can administer databases, manage cloud infrastructure, or even control a computer.
MCP Tool Distribution
We analyzed 1,945 tools across 238 servers in the MCP Registry to understand how tools are distributed.
MCP Tool Research
This week, we’re launching a new phase of research at Lyr3: digging into the tools that power the MCP ecosystem. So far: - We’ve collected data on > 420 MCP servers. - Across ~300 of them, we’ve identified 1,400+ distinct tools, tool descriptions and inputs.
Link: You are Doing MCP Wrong
Good read of what MCP is and is not from Jim Clark at Docker. link.
MCP Registry
Lyr3.com is happy to present its MCP Registry as a beta.
Let’s figure out what AI
actually looks like for you
No pitch deck. Just a conversation about where you are, what you're trying to do, and whether I can help.
Work with me