Smorster

The Best MCP Servers of 2025

AI LABS · 8m 11s · Watch on YouTube · 22 sources

Decision Card

Effort: Afternoon sampler — each MCP is a single install command plus one auth step in your AI client (~15–30 min per server); trying all six is roughly half a day.

Honest take: The video is sponsor-funded and repeats vendor marketing uncritically; its claim that “Docker was actually the first one to fix” MCP context bloat is shaky, since Cloudflare published Code Mode in September 2025 and Anthropic published its own code-execution-with-MCP approach, with Docker’s Dynamic MCP arriving in Docker Desktop 4.50 around the same period. It also overstates Context7 setup — an API key is optional (it only raises rate limits), and the free tier was later cut to ~1,000 requests/month.

Concrete next steps:

  • Context7 — adopt if you code with AI agents against fast-moving libraries; install per context7.com/plans and the repo docs (~15 min). Skip if your agent already has a docs-lookup tool.
  • Ref — try only if you need private-repo/PDF doc indexing or hit Context7’s context-bloat limits; see github.com/ref-tools/ref-tools-mcp (~20 min). Skip otherwise; free credits are thin.
  • Docker MCP Toolkit — try if you already run Docker Desktop 4.50+; enable Dynamic MCP per docs.docker.com/ai/mcp-catalog-and-toolkit/dynamic-mcp/ (~30 min). Skip if you connect fewer than ~3 MCP servers — the two-tool gateway solves a problem you don’t have.
  • shadcn registry MCP — adopt if building React/Next.js UIs with shadcn; setup at ui.shadcn.com/docs/mcp (~15 min incl. adding registries to components.json). Skip for non-React stacks.
  • Google managed MCP servers (Maps, BigQuery, Compute, GKE) — try only if you’re already on Google Cloud; quickstarts via the announcement (~30 min incl. IAM). Skip otherwise — remote, closed-source, GCP-account-gated.
  • Notion MCP — try if Notion is your task/content hub; OAuth to https://mcp.notion.com/mcp per developers.notion.com (~10 min). Skip if you use Obsidian — an equivalent MCP exists.
  • Supabase MCP — adopt if Supabase is your backend; follow supabase.com/docs/guides/ai-tools/mcp (~20 min), and scope it to a dev project, not prod. Skip if you don’t use Supabase.

TL;DR

A year-end roundup of six MCP servers the creator considers most impactful for AI-assisted development: Context7 (live library docs), Docker MCP Toolkit (a two-tool gateway that fixes context bloat), the shadcn registry MCP (UI components), Google’s new managed MCP servers (Maps, BigQuery, Compute, GKE), Notion MCP (workspace management), and Supabase MCP (database/schema automation). The recurring theme is context-window economics — the best MCPs either inject precisely the right context (docs, component specs, schemas) or reduce how many tool definitions the agent must carry.

Key Points

  • MCP, released by Anthropic in late 2024, became the foundation for a wave of 2025 developer tooling. 00:12
  • Context7 injects up-to-date, version-specific documentation and code examples into coding agents, avoiding dependency mismatches; its free plan is limited to open-source libraries. 01:22
  • Context7 works by maintaining a frequently updated vector database of docs queried via semantic search, unlike vague web-search results. 02:14
  • Ref is a context-efficient alternative that returns only relevant snippets and bundles web search/scraping, but its free credits are limited, so the creator still prefers Context7. 02:23
  • Docker MCP acts as a gateway: your client sees only two tools (mcp-find, mcp-add) even if hundreds of MCPs are connected behind it, and everything runs sandboxed in Docker. 02:58
  • Cloudflare articulated “code mode” as the fix for tool-bloated context windows; the video credits Docker as first to ship it, including dynamic MCP where agents write JavaScript tools that call other MCP tools. 03:56
  • The shadcn registry MCP gives agents component-specific context and installs components directly, including third-party registries (Aceternity UI, Magic UI) added via components.json. 04:26
  • Google launched fully managed, remote (closed-source) MCP servers alongside Gemini 3 — Maps, BigQuery, Compute, Kubernetes — plus open-source MCPs for Workspace, Firebase, Analytics, and Flutter. 05:26
  • Notion MCP handles search/create/update across a workspace after one-time auth; the creator uses it with Claude for content pipeline management, and an Obsidian MCP offers similar capabilities. 06:25
  • Supabase MCP removes manual SQL and schema management — the agent can create projects, manage schemas, and run operations end-to-end from prompts. 07:19

Notable Quotes

“I want to share the six best MCPs that truly changed the way that I fundamentally look at development now.” 00:20

“The fundamental problem faced while using MCP is a bloated context window due to many tools exposed in the context window while only a few are actually needed.” 03:46

“So now your context window consists of only two tools, even if the MCPS you have connected in Docker contain hundreds.” 03:34

Verified Claims

  1. Claim: Anthropic released the Model Context Protocol in late 2024. 00:12 Sources: Anthropic announcement (Nov 25, 2024), Wikipedia: Model Context Protocol Verdict: Confirmed.

  2. Claim: Context7’s free plan is limited to open-source libraries and requires signing up for an API key. 01:40 Sources: Context7 Plans & Pricing, Context7 docs: Plans & Pricing Verdict: Partially confirmed — free tier is indeed public-repos-only, but an API key is optional (it raises rate limits), and the free allowance was later cut to ~1,000 requests/month.

  3. Claim: Docker MCP exposes only two discovery tools (mcp-find, mcp-add) and runs connected MCPs sandboxed in Docker containers. 02:58 Sources: Docker docs: Dynamic MCP, Docker blog: Dynamic MCPs Verdict: Confirmed — Docker Desktop 4.50+ ships mcp-find, mcp-add, and code-mode tools; the gateway runs servers in isolated containers with restricted privileges.

  4. Claim: Cloudflare gave the general concept of “code mode,” and Docker was the first to actually fix the context-bloat problem. 03:56 Sources: Cloudflare: Code Mode (Sept 2025), Anthropic: Code execution with MCP Verdict: Disputed — Cloudflare coined the term and shipped a working implementation in its own blog post; Anthropic published its code-execution approach as well. Docker shipped a productized version, but “first to fix” is an overclaim.

  5. Claim: Google launched fully managed, remote, non-open-source MCP servers alongside Gemini 3, covering Maps, BigQuery, Compute, and Kubernetes. 05:26 Sources: Google Cloud: Official MCP support for Google services, TechCrunch coverage (Dec 2025) Verdict: Confirmed — launch set was Maps, BigQuery, Compute Engine, and GKE, delivered as managed remote servers following the Gemini 3 launch (and later expanded to 50+ servers).

  6. Claim: Ref is a more context-efficient docs MCP that exposes only relevant snippets instead of injecting large documents. 02:23 Sources: ref-tools/ref-tools-mcp on GitHub Verdict: Confirmed with caveat — Ref claims 60–95% token reduction and ~5k-token-max page extracts, but those figures are vendor-reported.

  7. Claim: The shadcn MCP server lets agents pull and install components from multiple registries (e.g., Aceternity UI, Magic UI) configured in components.json. 04:52 Sources: shadcn/ui MCP docs, shadcn/ui Registry docs Verdict: Confirmed.

  8. Claim: Supabase MCP lets the AI editor handle schema management, SQL operations, and even project creation on its own. 07:19 Sources: Supabase MCP docs, Supabase MCP blog Verdict: Confirmed — official server supports SQL execution, schema/migration management, project creation, Edge Function deploys, and log access.

  9. Claim: Notion MCP installs with a single command and needs authentication only on first setup. 06:25 Sources: Notion MCP getting-started guide, makenotion/notion-mcp-server Verdict: Confirmed — hosted server uses one-time OAuth at mcp.notion.com; the open-source server needs an integration token instead.

Tools, Papers & Standards Mentioned

Follow-up Questions

  1. How do the three context-bloat solutions — Docker’s Dynamic MCP, Cloudflare’s Code Mode, and Anthropic’s code-execution-with-MCP — compare in measured token savings and security model, and which is winning adoption in mainstream clients?
  2. For documentation MCPs, does Context7’s full-document injection actually degrade agent output quality versus Ref’s snippet-only approach (“context rot”), or is the difference mostly cost?
  3. What are the security and blast-radius implications of giving an agent write access via Supabase/Notion MCPs (schema drops, page deletion), and what scoping/read-only configurations do the official servers support?

Sources