MCP Anatomy Cheatsheet
How a connector works end to end, what Claude actually sees, the auth model, and the tool description format that determines everything. One page.
Host → Client → Server → Tools.
User types → Claude reads tool list → Claude reads descriptions → Matches description to intent → Calls the right tool. The quality of the match depends entirely on the tool description. This is why tool descriptions are the most important thing you write.
You've just learned how MCP works. Now here's why it matters beyond the engineering: products listed on Claude's connector directory get surfaced every time a user asks Claude for tool recommendations in that category. First listing becomes the default recommendation. No ongoing spend required.
"The tool description is the semantic interface between Claude and your product. It determines whether Claude calls your connector or a competitor's."
Architeqt AI — MCP connector build principleWhat every connector consists of.
These are the components Anthropic reviews when you submit. Understanding all five determines whether you pass.
Reads context, manages the session, decides which tools to call. This is Claude — you don't build this part.
Exposes capabilities to Claude. This is what you build. It responds to Claude's tool calls with structured data.
Executable actions Claude can choose to call autonomously. Each has: name, description, input schema, output format.
Read-only data via URIs. Client decides when to pull into context. Not auto-invoked by the model.
Required for remote MCP servers. Not optional. Client discovers auth server via .well-known discovery document.
Vague tool names, missing destructive annotations, no privacy policy, broken auth flow, tools that exceed their stated scope.
Use Claude to design your connector before writing code.
These work in regular Claude.ai today. Run them in order — action mapping before tool descriptions.
My product does [X]. Our API has these capabilities: [describe or paste your API surface]. List the 8 most valuable actions a Claude user would want to take with our product. For each: the action name (max 64 chars), what data it needs from our API, the user outcome, and whether it's read-only, write, or destructive.This is the most important document you'll produce. Anthropic reviewers evaluate your tool list — get it right before writing any code.
Review these MCP tool descriptions: [paste your tool definitions]. For each: Is the description specific enough for Claude to know exactly when to call it? Does it correctly flag read-only vs write vs destructive? Would the name pass Anthropic's review (specific, max 64 chars, no generic verbs)? Rewrite any that fail.Bad descriptions are the most common rejection reason. Run this before submission.
The window to be the default recommendation in your category won't stay open.
400+ connectors on Claude's directory across 30 categories. Most Indian SaaS categories: still open. 6 months old. The teams that get listed while the category is open become the permanent default when users ask Claude for tool recommendations in that space.
Subject: [Product] should be on Claude's connector directory Hi [Product] team, I just read through how MCP connectors work. Your API [does/could do] exactly what Claude's directory is designed for. Your category on claude.com/connectors: [what you found]. Products listed get surfaced when users ask Claude for tool recommendations. First listing in a sparse category = default. No ongoing ad spend. Permanent presence. The team that builds and lists these: architeqt.ai — [Your name]
30 minutes to configure. Results visible in week one.
The 2-Week MCP Build Playbook.
Day-by-day from API audit to directory listing. The decisions, the auth setup, Anthropic's actual rejection criteria, and the tool description format that passes review.
Ready to scope your connector?
15 minutes. I'll check your category saturation, score your readiness against Anthropic's actual criteria, and give you a straight verdict on whether to build now or wait.