The 2-Week MCP Connector Build Playbook
From API to Claude's directory — every decision, in order. The decisions happen in week 1. The build happens in week 2. That's not an accident.
Most connectors fail because the decisions were never made.
The code works but the connector feels broken — because the tool descriptions were vague, the actions were poorly separated, or the auth model was bolted on after the build. These aren't engineering problems. They're decision problems that got deferred until they became expensive.
The most important thing you'll write is your tool descriptions. These are the semantic interface between Claude and your product. They determine when Claude calls your connector, what it asks for, and whether the output is useful. Bad descriptions mean Claude calls the wrong action or doesn't call yours at all.
"The tool description is the product. Everything else is implementation."
Architeqt AI — MCP connector build principleAPI audit and action mapping.
These decisions determine everything in week 2. Skipping any of them is how connectors get rejected.
Which endpoints are stable, documented, and owned by you? Mark: stable / unstable / third-party dependency / deprecated.
Not all 40 endpoints — the 8–12 users actually want Claude to do. Ask: what do users do most, and what takes them the longest?
These need different handling, different permissions, and different annotations. Anthropic reviewers check this explicitly.
What should Claude never be allowed to do with your product? Write this list before the build. It shapes every tool description.
Not 'manage_data'. Not 'api_call'. Specific = correct call selection. Vague = Claude guesses wrong.
Remote MCP servers require OAuth 2.1 with PKCE. This is not optional. Design your auth before you write any connector code.
Think through the connector decision with Claude.
Run these in regular Claude.ai before the build starts. They surface the decisions that determine success.
Our product does [X] for [ICP]. List the 10 most valuable actions a user might want Claude to take with our product. For each: describe the specific action, what data it needs from our API, the expected user outcome, and whether it's a read-only or write operation. Flag any that would likely be rejected by Anthropic's safety review.This output becomes your tool definition design. The most important technical document in the build.
Review these tool descriptions for a Claude MCP connector: [paste your tool definitions]. For each: (1) Is the description specific enough for Claude to know exactly when to call it? (2) Does it correctly signal read-only vs write vs destructive? (3) Would this pass Anthropic's safety review? Rewrite any that fail.Run this on every tool definition before submission. Bad descriptions are the most common rejection reason.
The window to be the default recommendation in your category won't stay open.
I'm assembling the builds for August. End-to-end: API audit, action mapping, tool descriptions, auth setup, build, test, submit. Listed on Claude's directory and across the major MCP ecosystems. 1–4 weeks start to published.
Hey [Founder], Two things worth your 5 minutes: 1. claude.com/connectors — search your category and see what's listed. 2. The directory is 6 months old. Less than 5 Indian products in 200+. Products listed there get surfaced when enterprise users ask Claude for tool recommendations. First listing in a sparse category = default recommendation. No ongoing spend. Permanent presence. End-to-end build + listing: architeqt.ai · 1–4 weeks
30 minutes to configure. Results visible in week one.
Before you write a line of code.
15 minutes. I'll check your category, score your readiness against Anthropic's actual criteria, and give you a straight verdict: build now, fix gaps first, or your category is already too crowded.
Want someone to do this for you?
I build, test, and submit MCP connectors for SaaS products — listed on Anthropic's directory and across the major MCP ecosystems. 1–4 weeks, start to published.