Smorster

The Best MCP Servers of 2025

AI LABS · 8m 11s · Watch on YouTube · 21 sources

Decision Card

Effort: A few hours of setup spread across whichever servers fit your stack — each is a copy-paste install command plus an OAuth/API-key step (Context7 ~5 min, Supabase/Notion ~10 min each, Docker MCP Toolkit ~30 min to wire up a gateway).

Honest take: This is a useful shortlist, but it’s a sponsored, single-creator “favorites” video, not a benchmark — and one of its headline claims is shaky: it credits Docker as “the first one to fix” the context-bloat problem with code mode, when Cloudflare published the Code Mode concept and a working two-tool MCP server itself. Docker’s own docs also warn that code mode is “in early development and not yet reliable for general use,” which the video glosses over.

Concrete next steps (per item):

  • Context7 — adopt: free for open-source libs, fixes hallucinated/outdated APIs; install via context7.com. Skip if you only work in a stable, well-known stack the model already knows cold.
  • Ref — skip: marketed as a context-efficient Context7, but the video itself says the free tier is too limited and Context7 wins unless you need its extra web-scrape/search features.
  • Docker MCP Toolkit / code mode — try: real fix for tool bloat (docs), but code mode is explicitly early/unreliable — pilot it, don’t depend on it. Skip if you run only 1–2 MCP servers (no bloat to solve).
  • shadcn registry MCP — adopt if you build React/Tailwind UIs: official ui.shadcn.com/docs/mcp. Skip if you’re not on a shadcn-compatible front end.
  • Google managed MCP servers — try: remote, zero-config BigQuery/Maps/GKE (announcement), but closed-source and GCP-only. Skip if you’re not on Google Cloud.
  • Notion / Obsidian MCP — adopt for knowledge/PM workflows; skip if your team doesn’t live in either tool.
  • Supabase MCP — adopt if Supabase is your backend (supabase.com/docs/guides/ai-tools/mcp); use read-only/dev-project mode for safety. Skip if you don’t use Supabase.

TL;DR

A sponsored year-end roundup of six MCP servers the AI LABS creator adopted in 2025: Context7 (live docs), Docker MCP Toolkit (tool-bloat fix via code mode), the shadcn registry server (UI components), Google’s new managed servers, Notion/Obsidian (knowledge), and Supabase (database). Most are quick installs and genuinely useful, but the picks are personal favorites, not benchmarked, and one origin claim about code mode is overstated.

Key Points

  • MCP was released by Anthropic in late 2024 and rapidly “blew up,” with many products built around it 00:10
  • Context7 injects up-to-date, version-specific docs and code examples into the coding agent, avoiding dependency mismatches; free tier covers open-source libraries 01:22
  • Context7 uses a frequently-updated vector database plus semantic search rather than vague web-search results 02:11
  • “Ref” is pitched as a context-efficient Context7 alternative, but its free credits are limited, so Context7 is the better default 02:22
  • Docker MCP exposes just two tools (MCP find / MCP add) so connecting hundreds of MCP tools doesn’t bloat the context window, and everything runs sandboxed in Docker 02:58
  • The creator says Cloudflare and Anthropic highlighted context bloat and Cloudflare named the “code mode” concept, but claims Docker was first to actually fix it 03:54
  • The shadcn registry MCP fetches and installs UI components correctly (with dependencies) and supports custom registries like Aceternity and Magic UI via components.json 04:24
  • Google announced fully managed MCP servers alongside Gemini 3 — Maps, BigQuery, Compute, and Kubernetes — that are remote and not open source 05:23
  • The Notion MCP installs with one command and one-time auth, and the creator uses it with Claude to manage content pipeline and ideas; Obsidian has an equivalent 06:25
  • The Supabase MCP lets the AI editor handle schema management, SQL, and project setup without manually writing queries 07:14

Notable Quotes

“Context 7 pulls all the up-to-date version specific documentation and code examples directly into the AI coding agent.” 01:22

“So now your context window consists of only two tools, even if the MCPS you have connected in Docker contain hundreds.” 03:36

“It eliminates the need to manually write SQL queries or manage database schemas and configurations.” 07:26

Verified Claims

Claim: Anthropic released the Model Context Protocol in late 2024 and it saw rapid adoption. 00:10 Sources: Anthropic — Introducing MCP, Model Context Protocol (Wikipedia) Verdict: Confirmed (announced November 2024; thousands of community servers since).

Claim: Context7 pulls version-specific docs/code examples into the agent via a vector DB + semantic search to prevent outdated/hallucinated APIs. 01:22 Sources: upstash/context7 (GitHub), Context7 clients docs Verdict: Confirmed.

Claim: Docker MCP reduces exposed tools to ~two while keeping access to a catalog of verified servers, running them sandboxed. 02:58 Sources: Docker MCP Catalog & Toolkit docs, Docker — Dynamic MCPs blog Verdict: Confirmed (100+ verified tools; find + code-mode keeps the context window small).

Claim: Cloudflare introduced the “code mode” concept and Docker was “the first one to fix” the tool-bloat problem. 04:00 Sources: Cloudflare — Code Mode: the better way to use MCP, Cloudflare — give agents an entire API in 1,000 tokens Verdict: Disputed — Cloudflare both named and shipped a working Code Mode MCP server (two tools, ~1,000 tokens for the full API); the “Docker was first to fix it” framing is questionable.

Claim: The shadcn registry MCP installs components with correct dependencies and supports custom/third-party registries via components.json. 04:24 Sources: shadcn/ui — MCP Server docs, shadcn CLI 3.0 + MCP changelog Verdict: Confirmed.

Claim: Google announced fully managed MCP servers (Maps, BigQuery, Compute, Kubernetes) alongside Gemini 3, remote and not open source. 05:23 Sources: Google Cloud — Announcing official MCP support, TechCrunch coverage Verdict: Confirmed (rollout began Dec 10, 2025, following Gemini 3).

Claim: The Supabase MCP lets the AI handle schema management and SQL without manual queries. 07:14 Sources: Supabase MCP docs, Supabase MCP announcement blog Verdict: Confirmed (tools include list_tables, apply_migration, execute_sql).

Tools, Papers & Standards Mentioned

Follow-up Questions

  1. Now that both Cloudflare and Docker ship code-mode implementations, which is more reliable and lower-token in real agent workloads — and does Anthropic’s own code-execution-with-MCP approach beat both?
  2. How safe is granting an MCP write access to production data (Supabase, Notion, GKE)? What’s the least-privilege setup (read-only mode, scoped tokens, sandboxing) for each of these servers?
  3. For teams running 10+ MCP servers, does the Docker gateway’s two-tool approach measurably improve task success rates and cost versus loading all tools directly?

Sources