15 MCP Tools Traced: The Good Ones Are All in Coding — For One Reason
A joke circulating in the community: MCP is a technology with more developers than users.
The quip isn't baseless. By early 2026, a security analysis of roughly 15,000 public MCP servers concluded that nearly half were misconfigured. Many servers are practice projects abandoned right after being written. By the numbers, it does look like an oversupplied technology.
But apply that line to the coding domain and it breaks down. I recently traced 15 tools from an MCP roundup video back to their official repos and documentation, and saw a different picture. Background first, then conclusions.
What Is MCP, in Thirty Seconds
MCP stands for Model Context Protocol, an open protocol released by Anthropic in November 2024. It solves a small but critical problem: AI models calling external tools used to require custom interfaces per vendor, with glue code written for every integration; MCP unified the interface. A developer writes a server once, and every MCP-capable client — Claude Code, Codex, Cursor — can use it directly. When Claude Code or Codex helps you operate a database, run a browser, or deploy a website lately, that's this protocol underneath.
In December 2025, Anthropic donated MCP to the Linux Foundation's Agentic AI Foundation, alongside Block's goose and OpenAI's AGENTS.md. The protocol is now neutral, and the spec is still revised quarterly — most recently July 2026.
The 15 Tools Form a Pipeline
After checking each tool's official background, I found these aren't isolated tools. Arranged by the software development workflow, they form a complete pipeline:
- Design: Figma MCP hands design layout data directly to AI coding tools. There are official (built into the Figma desktop app) and community versions — Framelink, an independent open-source project with ~16k stars.
- Coding / docs: Context7 (by Upstash) feeds up-to-date library docs to the model on demand, fixing "the model remembers the old API." ~63k stars, the highest of the 15. Ref does similar doc retrieval; the company does only this one thing — the most niche entry on the list.
- UI: ShadCN MCP lets the AI pull components from a registry; the official CLI already includes this, plus a community version with ~3k stars.
- Data: Neon and Supabase database MCPs. Neon was acquired by Databricks for ~$1B in May 2025 — the acquisition announcement called Neon a "database designed for the agentic AI era." Database companies now treat agents as their #1 customer.
- Deployment: Vercel, Cloudflare, and Tencent Cloud EdgeOne Pages — all official servers. Vercel's is purely remote-hosted: one URL at mcp.vercel.com, no local config needed.
- Payments: Stripe's official agent-toolkit. Security: Semgrep, for vulnerability scanning. Orchestration: GitHub's official MCP, ~33k stars.
The Conclusion in One Sentence
Without exception, behind every one of these MCPs is a platform wrapping its existing APIs in a form agents can call directly.
Chrome DevTools MCP was written by Google's Chrome team itself, released September 2025, hitting ~53k stars within a year. GitHub MCP is official GitHub. Stripe, Cloudflare, Supabase, Vercel, Tencent EdgeOne — all platforms showing up in person.
So MCP is to platform APIs what apps once were to websites — not a new capability, but a new distribution form. It looks like you're "using MCP," but really an agent is calling the platform's API on your behalf. Platforms writing MCP servers means opening their services onto the new street of the agent ecosystem.
Why the Good Ones Cluster in Coding
The common explanation is "developers love new toys." I think the stronger explanation is verification cost.
When code is written, run it and you know if it works. Database connected? Deployment succeeded? Tests passed? Every step yields a deterministic success or failure signal. Agents working in this domain don't need a human referee. Compare: an agent booking flights or replying to email — mistakes surface only when a human notices; an agent creating a database and writing queries — a schema error retries on the spot. The cheaper verification is, the faster the loop turns and the more productivity is realized.
Coding happens to be one of the lowest-verification-cost domains. So MCP blooming here first isn't chance — it's structure: wherever verification is cheap, agents deliver value first.
Pitfalls Found Along the Way
A few things the video didn't mention, worth knowing before choosing:
1. Semgrep's official MCP repo's last commit was late October 2025 — nearly a year dormant. Set expectations accordingly. EdgeOne Pages MCP has also been inactive ~3 months. 2. Figma and ShadCN each have an official and a community version with meaningfully different capabilities — don't install both blindly. 3. Item #15, the MCP SDK, isn't strictly a tool — it's the SDK for building tools (official TypeScript and Python implementations, ~38k stars combined). Placed last, it conveniently closes the loop: the first 14 use MCP, the last one makes MCP.
The Joke: Half Right, Half Wrong
I couldn't find the original source of "more developers than users," so treat it as folklore. It's half right and half wrong.
Wrong half: in coding, MCP's user profile was never consumers. Developers configure it for themselves; agents run it for developers; the user is the developer. The loop needs no one else. Right half: the configuration barrier (editing JSON, getting tokens) does keep ordinary people out, and general-purpose MCP use is indeed quiet.
15,000 servers, half misconfigured — that's the ecosystem's floor. Chrome DevTools at 53k stars in a year — that's the ceiling. Floor and ceiling coexisting doesn't look like a quiet technology; it looks like an industry just past its founding stage: low barriers, clear leaders.
If your agent has no MCP configured, two公认的 starting points: Chrome DevTools (lets the AI see the web pages it writes) and Context7 (lets the AI read the latest docs). Starting from these beats any "toolbox list."
---
Appendix: Verification Table for the 15 Tools (stars as of GitHub snapshot, 2026-10-04)
| # | Tool | Backed by | Stars | One-line role | |---|------|-----------|-------|----------------| | 1 | Chrome DevTools MCP | Google Chrome DevTools team | 52,946 | Browser control / web debugging for agents | | 2 | Neon MCP | Neon (acquired by Databricks for ~$1B, 2025-05) | 648 | Create DBs, alter tables, run SQL | | 3 | Supabase MCP | Supabase official | 2,929 | Full Postgres suite management | | 4 | Figma MCP | Official (desktop built-in) + community Framelink | 15,951 (community) | Design-to-code | | 5 | Context7 | Upstash | 62,666 | Live library docs fed to the model | | 6 | Ref | ref.tools (company product) | not open source | Doc retrieval for agents | | 7 | Replicate MCP | Replicate official (hosted + npm) | — | Run Veo/Flux generative models | | 8 | Vercel MCP | Vercel official (mcp.vercel.com) | remote-hosted | Deployment management, purely remote | | 9 | EdgeOne Pages MCP | Tencent Cloud EdgeOne | 435 | One-click public HTML deployment | | 10 | Cloudflare MCP | Cloudflare official | 4,353 | Workers/KV/R2 management | | 11 | GitHub MCP | GitHub official | 33,357 | Full repo/PR/issue operations | | 12 | Stripe MCP | Stripe official (inside agent-toolkit) | 1,853 | Payments and billing | | 13 | ShadCN MCP | Official CLI built-in + community version | 3,021 (community) | Direct registry component access | | 14 | Semgrep MCP | Semgrep official | 687 | Code security scanning (⚠️ ~1 year inactive) | | 15 | MCP SDK | MCP official | TS 13,512 / Py 24,481 | The SDK for building MCP servers |
Sources: each tool's official repo and documentation checked individually; ecosystem size and the "~half misconfigured" figure come from January 2026 public reporting; "more developers than users" is a circulating saying with no traceable origin; Neon acquisition amount per Databricks' official announcement and media coverage.