# The Best MCP Servers of 2025 Source: https://www.youtube.com/watch?v=kRxdsv1ZY1E Recap page: https://rapidrecap.app/video/kRxdsv1ZY1E Generated: 2025-12-23T13:32:47.128+00:00 --- ## Quick Overview The video showcases various Model Context Protocol (MCP) servers available in 2025, emphasizing how dynamic tool selection via "code-mode" tools significantly reduces context window bloat by only loading necessary tools, saving tokens and improving agent efficiency across platforms like GitHub, Notion, Google Cloud services (Maps, BigQuery, GCE, GKE), Supabase, and Shadc'n UI. **Key Points:** - The Model Context Protocol (MCP) saw significant adoption by 2025, with new models like Claude Opus 4.5 and Gemini 3 being released, driving the need for efficient tool integration. - Blink, an AI app builder, allows users to connect to various MCP servers, including integrations for Notion, Shadc'n UI, Google services, and Supabase. - Code-mode tools, introduced to address token inefficiency from loading all available tools, allow agents to dynamically select and use only the necessary tools, saving thousands of tokens per turn. - Specific new remote MCP servers announced by Google include those for Google Maps (Grounding Lite), BigQuery, Google Compute Engine (GCE), and Google Kubernetes Engine (GKE). - Open-source MCP servers demonstrated include those for Notion (providing tools for managing pages and databases), Shadc'n (for UI components), and GitHub (for repository management). - The Supabase MCP server enables agents to natively query database schemas and execute SQL operations without moving sensitive data into the LLM context window, ensuring governance. - Tools like `mcp-find` and `mcp-add` facilitate discovery and connection to trusted MCP servers dynamically, solving the friction point of tool discovery. ![Screenshot at 00:04: The screen displays the announcement of "Gemini 3" alongside Anthropic's "Introducing Claude Opus 4.5," setting the context of advanced AI models driving the need for better tool integration.](https://ss.rapidrecap.app/screens/kRxdsv1ZY1E/00-00-04.jpg) **Context:** The video provides an update on the evolution of the Model Context Protocol (MCP) ecosystem in a hypothetical 2025 setting, focusing on how LLM agents are becoming more powerful by reliably accessing and using external tools and data. The demonstration centers on the capabilities of an AI platform (Blink/Claude Code) to integrate and leverage these specialized MCP servers for development tasks, code generation, and data manipulation. ## Detailed Analysis The video projects the state of AI agent development in 2025, highlighting the maturity of the Model Context Protocol (MCP) ecosystem. It showcases advancements in LLMs like Claude Opus 4.5 and Gemini 3. A key focus is on improving tool efficiency using "code-mode" tools, which dynamically select only the required MCP tools (e.g., 2 tools out of 6 available) to prevent context window bloat, saving significant tokens (up to ~9,900 per turn compared to sending 10,000+ tokens for tool definitions). Several MCP servers are demonstrated: Google announced fully managed remote servers for Google Maps, BigQuery, GCE, and GKE, enabling agents to perform real-world queries and infrastructure management securely. Open-source servers for Notion, Shadc'n UI components, and GitHub are shown in action, allowing agents to search, create, update, and manage data directly within those platforms using natural language commands. For instance, the Notion MCP provides tools for read-only searches and destructive actions like creating pages. The Supabase MCP allows agents to query databases and manage projects directly, bypassing security risks associated with moving sensitive data into the context window. The core mechanism enabling this efficiency is dynamic tool selection, where agents use commands like `mcp-find` and `mcp-add` to manage their toolset on the fly via an MCP Gateway, ensuring agents remain secure within sandboxes while accessing vast capabilities. ### AI Model Advancements - Claude Opus 4.5 introduced - Gemini 3 released, requiring reliable tool work with data - Context usage shown consuming 21.0k tokens for MCP tools in one session ### Tool Efficiency Solution - Code-mode enables agents to select only 1 or 2 necessary tools from all available tools, saving hundreds of thousands of tokens per turn - This contrasts with older methods where tool definitions were sent every turn ### Google's New Remote MCPs - Google announced fully-managed remote MCP servers for Google Maps (Grounding Lite) for real-world queries - BigQuery MCP for reasoning over enterprise data without security risks - GCE and GKE MCPs for autonomous infrastructure and container management ### Open-Source MCP Integrations - Notion MCP provides 14 tools for reading, creating, editing, and managing notes/databases - Shadc'n MCP enables agents to search and install UI components directly from registries like 'cult-ui' - GitHub MCP offers 40 tools for repository management like creating pull requests ### Database Interaction - Supabase MCP allows agents to execute SQL operations, manage schema, and handle project creation/cost estimation without moving sensitive data to the LLM context window ### MCP Management - Dynamic tools allow agents to use `mcp-find` and `mcp-add` to discover and connect to servers in the MCP Catalog (e.g., 272 servers available in Docker Desktop) securely. ![Screenshot at 00:03: Anthropic announcement of Claude Opus 4.5, highlighting the ongoing advancement of AI models necessitating better tool management.](https://ss.rapidrecap.app/screens/kRxdsv1ZY1E/00-00-03.jpg) ![Screenshot at 00:04: Google's announcement of Gemini 3, emphasizing the need for AI to work reliably with tools and data to act as a true agent.](https://ss.rapidrecap.app/screens/kRxdsv1ZY1E/00-00-04.jpg) ![Screenshot at 00:17: Demonstration of connecting various web tools \(Asana, Gmail, Notion, etc.\) to Claude via its connector system.](https://ss.rapidrecap.app/screens/kRxdsv1ZY1E/00-00-17.jpg) ![Screenshot at 02:52: Comparison chart showing token savings achieved by using code-mode \(1 tool call\) versus traditional methods \(6 separate tool calls, scattered results\).](https://ss.rapidrecap.app/screens/kRxdsv1ZY1E/00-02-52.jpg) ![Screenshot at 03:04: The Docker Desktop MCP Toolkit Catalog displaying numerous available MCP servers, including Context7, Grafana, and DuckDuckGo.](https://ss.rapidrecap.app/screens/kRxdsv1ZY1E/00-03-04.jpg)