MCP Tools: Plug in and Play

Introduction

MCP (Model Context Protocol) is making it dead simple to connect your LLMs and agents to real-world data, APIs, and tools. Best described as the 'USB interface for AI". With a growing ecosystem of tools and a vibrant community, MCP is revolutionizing how we interact with AI and vice-versa. Let’s dive into the coolest MCP tools, their unique features, and why this ecosystem is blowing up and where it still needs some polish. Photo by Anthropic: What is MCP?

What Are MCP Tools, Really?

Released by Claude's maker, Anthropic in November 2024, think of MCP tools as little adapters or plugins that let your AI talk to the outside world—fetching docs, querying databases, scraping the web, you name it. They all speak the same "language" (the MCP protocol), so you can mix and match them with your favorite LLMs, agent frameworks, or even your own scripts. They're lightweight, open, and run anywhere: locally, in Docker, or in the cloud. Plus, with features like Sampling ( Human in the loop i.e. edit the prompt before it's sent to LLM), usage tracking and Cursor/LLM Development, MCP tools are setting a new standard for AI integration. No need to re-invent the wheel by writing another search tool that is already easily accessible. Let us go through some of the best currently available MCP tools to understand them better. Photo by Daily Dose Of Data Science

Why Is This So Cool?

• Instant Upgrades: Add new skills to your AI in seconds by skipping the messy part.
• Privacy First: Run it locally if you're paranoid about giving AI more data to train on.
• Mix and Match: Swap out tools, try new ones, or build your own using diverse data sources. It’s all modular.
• Community Vibes: Tons of open-source MCP tools are popping up for everything from databases to 3D modeling.

Where to find them?

The take off in MCP adoption introduced the chance for a lot of providers to become the de-facto choice as a MCP directory. Some of the most popular currently are -

• MCP.so
• Smithery
• PulseMCP
• Docker MCP Catalog Released only last week! April 22, 2025

Check them out to explore the latest and most popular tools for your use-cases. Note that MCP tools may be run locally run or remotely, some of the providers may require API keys and charge accordingly.

Some of the Best MCP Tools Right Now

1. Context7 by Upstash

What it does: Context7 is your LLM's personal librarian. It grabs the latest, version-specific docs and code examples for libraries like React, Next.js, and Tailwind—so your AI doesn't spit out outdated or hallucinated code. A huge issue solved caused by LLM knowledge cut-offs and fast updating frameworks that continuously deprecate their methods. Why it rocks:

• No more "that function doesn't exist" errors.
• Plays well with Cursor, Claude Desktop, Windsurf, and more.
• Low chance to hallucinate non-existent API's
• Free for personal use.

Practical use -

"Add RAG capability to my application using langchain. Use context7 for latest docs"

2. Sequential Thinking by Anthropic

What it does: Helps the LLM to plan long-term through reflective problem solving. Why it rocks:

• Modular thinking for complex problems.
• Refines the plan iteratively.

Practical use -

"Add support for different 'Projects' in my application so user can have multiple configurations saved for their work. You will need to update both the backend and frontend containers. Use sequential-thinking to create a plan before implementation."

3. Brave Search by Brave

What it does: Provides ability to surf the web. Why it rocks:

• LLm's can access up-to-date real-time information.
• Local search mode to find businesses and services around.
• Custom filters can be applied.

Practical use -

"What is happening in Victoria, BC this weekend? Use brave search."

4. Playwright Automation by Microsoft

What it does: For all my front-end tech bros out there - enable your LLM's to directly interact with your web pages. Why it rocks:

• Makes your vibe-coding sessions more productive with automated LLM testing. No more screenshotting or copying your console outputs.
• Works with no vision models.
• General purpose web navigation and extraction.

Practical use -

"Investigate issues with the particle effect on the index page using Playwright."

5. Filesystem by Anthropic

What it does: Gives LLM's access to your local Filesystem for operations. Why it rocks:

• Only allows operations withing specified directories.

Practical use -

"Create a python script to find latest AI articles, save it as ai-news-finder.py to my Filesystem"

Ready to Plug In?

If you're using IDE'S like Windsurf, Codeium or VSCode, look for the MCP icon in your AI window. You can search for tools in the directories your editor supports to easily add them, else find a tool that you like and manually add it to config file -

{
  "mcpServers": {
    "context7": {
      "command": "npx",
      "args": ["-y", "@upstash/context7-mcp@latest"]
    }
  }
}

Or spin up a Docker container:

docker build -t context7-mcp .

and update config file to

{
  "mcpServers": {
    "Сontext7": {
    "autoApprove": [],
    "disabled": false,
    "timeout": 60,
      "command": "docker",
      "args": ["run", "-i", "--rm", "context7-mcp"],
      "transportType": "stdio"
    }
  }
}

MCP servers can also be used directly in your AI applications. Let us explore how we can use them using Google's Agent Development Kit by importing the MCPToolset library as follows -

from google.adk.agents.llm_agent import LlmAgent
from google.adk.tools.mcp_tool.mcp_toolset import MCPToolset, StdioServerParameters

async def create_agent():
  """Gets tools from MCP Server."""

  tools, exit_stack = await MCPToolset.from_server(
      connection_params=StdioServerParameters(
          command='npx',
          args=["-y",
                "@modelcontextprotocol/server-google-maps",
          ],
          env={
              "GOOGLE_MAPS_API_KEY": google_maps_api_key
          }
      )
  )

  # Make sure GOOGLE_API_KEY is defined in .env or you can run a local model using Ollama and connect using LiteLLM
  agent = LlmAgent(
      model='gemini-2.5-flash-preview-04-17',
      name='maps_assistant',
      instruction='Guide the user as they are navigating the map using the tools provided.',
      tools=tools,
  )
  return agent, exit_stack


root_agent = create_agent()

Problems in the ecosystem

• Clunky to set up. Complex and fragmented installation process for different editors, frameworks.
• Inadequate security and authentication on public MCP servers.
• Untrusted MCP servers may introduce malware or use prompt injection.

What’s Next for MCP?

The ecosystem is moving fast. Expect:

• One click installations - arrival of Docker is already changing this.
• More “plug and play” tools for everything from databases to music generation.
• Better security and sandboxing.
• Dynamic tool discovery, your AI can find and use new tools on the fly.
• Even tighter IDE integrations— debugging, querying, and automating right from your code editor.
• More standardization across the lifecycle of MCP tools.

---

References:

• Claude MCP Blog – “What is MCP?”
• Hugging Face Blog – “TOP 11 Essential MCP Libraries”
• Model Context Protocol – “Introduction”
• Docker Blog – “Introducing MCP Catalog and Toolkit”
• Smithery – “Model Context Protocol Registry”
• GitHub – “upstash/context7”
• Pulse MCP
• Docker Hub – “mcp Catalog”
• Daily Dose Of Data Science – “Visual Guide to Model Context Protocol”
• GitHub – “Model Context Protocol Servers: Sequential Thinking”
• GitHub – “Model Context Protocol Servers: Brave Search”
• GitHub – “microsoft/playwright-mcp: Playwright MCP server”
• GitHub – “Model Context Protocol Servers: Filesystem”