Back to all posts
AI

How to use MCP with Apify Actors

Learn how to expose over 4,500 Apify Actors to AI agents with Claude and LangGraph, and configure MCP clients and servers.
by
Percival Villalva
Percival Villalva

Anthropic’s Model Context Protocol (MCP) has quickly become one of the hottest topics in AI engineering circles. Despite all the buzz, the core idea behind MCPs is actually pretty straightforward, and far from new in software engineering. The goal is to create a standardized way for AI apps and agents to connect with external tools and data sources, similar to how REST APIs structure backend communication or how Language Server Protocols (LSPs) let IDEs talk to language-specific tools.

As simple as it sounds, MCP facilitates the building and integration of powerful AI-powered apps that benefit developers, tool creators, end users, and enterprises, making the whole AI integration process much more scalable and accessible.

How MCP fits into Apify

Apify is uniquely positioned to take full advantage of MCP, as it offers over 4,000 public tools, available as Actors on Apify Store, that can be exposed to LLM applications through an MCP server. To better understand how Apify fits into the MCP architecture, take a look at the diagram below.

MCP + Apify architecture

From the diagram, you can see that Apify wears a few different hats in the MCP architecture, and this is only possible because of the flexibility provided by the Apify Actors. Here is how to interpret this diagram:

MCP clients

An MCP client is essentially the part of an AI system that knows how to discover and call tools using the MCP format. It might sound fancy, but you’re probably already using one daily, like the Claude desktop app, the Cursor IDE, or the Apify Tester MCP Client.

MCP servers

An MCP server exposes tools, resources, or prompts to an LLM client in a way that the client can understand and interact with.

Practically anything can be an MCP server if it follows the protocol. Apify can expose its Actors through the Apify MCP Server, Slack could offer chat and workspace actions, and GitHub could wrap its API for things like creating issues or reviewing pull requests. If it can describe its capabilities and handle structured tool calls, it can plug into the MCP ecosystem.

Tools (Apify Actors)

Tools are external functions or capabilities, like APIs, scripts, or automations, that an LLM can call to get things done.

Herein lies the true value of Apify: MCP makes it possible to expose thousands of public Actors as tools that the LLM can use. That means giving an LLM the ability to scrape structured data from nearly any website and automate workflows like sending emails, downloading YouTube videos, or searching the web.

Plus, you’re not limited to existing tools: you can build your own Actors and make them accessible to LLMs through the MCP protocol.

Resources (Datasets and Key Value Stores)

In MCP, resources are data objects that tools can read from or write to. Basically, they’re how tools share or persist information across calls.

In the case of Apify, resources map directly to things like Datasets and Key-Value Stores. A Dataset can hold structured data, like the results of a web scraping run, while a Key-Value Store is useful for storing config files, inputs, or intermediate results.

By exposing these through MCP, LLMs can not only call tools but also interact with the data those tools generate, making workflows more dynamic and stateful.

Prompts

Finally, we have prompts, which act as predefined text templates that provide guidance to the LLM on how to use a specific tool or handle its output. They act as built-in instructions or examples that help the model understand when and how a tool should be used within a larger task.

For example, a prompt might tell the LLM how to phrase a query, what kind of input to provide, or how to interpret the results it receives. By exposing prompts through the MCP server, developers can fine-tune the model’s behavior dynamically, without modifying the LLM itself, making tool usage more reliable, contextual, and flexible.

Now that we’ve covered how Apify fits into the MCP ecosystem, let’s jump into how you can actually use Apify Actors as tools through MCP in practice.

How to use MCP with Apify Actors: Tester MCP Client

In this section, we’ll use Apify’s Tester MCP Client, which is the easiest way to test and validate how a client interacts with tools.

We’re starting with this client because it gives us a clear view of each step in the MCP architecture we discussed earlier, helping you understand how everything works together to give your LLM its capabilities.

After that, we’ll look at how to set up the same functionality in the Claude Desktop App or any other MCP-compatible client.

To get started, open the Tester MCP Client and hit Try for free. You don’t even need an LLM API key, if you’re on Apify’s free plan, you get $5 in credits every month, which the client will use automatically. Once you’re logged into your Apify account, you’ll see a UI that looks like this:

Tester MCP Client - Initial prompt

Here, you’ll already notice one of the MCP components we talked about earlier: the prompt, specifically the system prompt. This initial prompt sets the overall context, behavior, or persona for the AI’s responses. The Tester MCP Client lets you customize this prompt before starting the server, so feel free to tweak it however you like. Once you’re ready, just click “Start.”

After that, you’ll be taken to the Log tab, where you can see the system prompt that was sent to the LLM along with a generated URL that links to the client’s chat UI.

System prompt and client chat UI

For this test run, I’ll ask the LLM to create a list of the top 10 cafes in Boston, ranked by ratings and number of reviews.

Notice that I’m not telling it which tool to use or how to complete the task. Instead, the client will automatically check which tools are available from the server - in this case, all 4,500+ Actors on Apify Store - and choose the one that best fits the request.

Tester MCP Client demo

The LLM ended up selecting one of the Google Maps Scraper Actors to complete the task. You can clearly see which tool it used in the UI. After a couple of minutes, it returned an up-to-date and fairly accurate list of the top cafes in Boston - something that wouldn’t have been possible without access to external tools.

Tester MCP Client output

Using Actors with any MCP Client with Actors MCP Server

You might have noticed that while using the Tester MCP Client, we didn’t mention one of the key parts of the MCP architecture: the server.

That’s because the Tester MCP Client is already preconfigured to connect to the Apify MCP Server, so there’s no setup required on your end.

However, if you want to give other MCP clients the same capabilities we just demonstrated, you’ll need to connect them to the Apify MCP Server yourself. Fortunately, thanks to the standardized MCP protocol, this is now a pretty straightforward process.

Start by going to the Actors MCP Server, and you will see the UI below:

Actors MCP Server UI

In the input section, you’ll see that it offers more granular control over which tools (Actors) are exposed to the client. This is useful because, as a user, you might not have access to all the Actors on Apify Store, or you may want to limit the available tools to improve the LLM’s predictability and response accuracy.

To add an Actor, simply click the “Add” button and paste its ID into the input field. You can find the Actor ID right below the title on its Apify Store page, for example, apify/actors-mcp-server is the ID for the Actors MCP Server.

Actors MCP Server on Apify Store

Finally, paste the MCP configuration below into the MCP client you want to connect to your Apify Actors. Make sure to replace the placeholder with your actual Apify API token:

{
     "mcpServers": {
           "actors-mcp-server": {
                 "command": "npx",
                 "args": ["-y", "@apify/actors-mcp-server"],
                 "env": {
                        "APIFY_TOKEN": "your-apify-token"
                 }
           }
     }
}

For Claude Desktop, you’ll typically find the claude_desktop_config.json file in one of the following locations, depending on your operating system:

  • On macOS: ~/Library/Application\\ Support/Claude/claude_desktop_config.json
  • On Windows: %APPDATA%/Claude/claude_desktop_config.json
  • On Linux: ~/.config/Claude/claude_desktop_config.json

After reloading your Claude application, you should see the Actors defined in the Actors MCP Server listed as available tools.

Actors defined in MCP Server

Remember, MCP was created to standardize how tools are made available to clients. That means the process we just followed will work the same way for any other MCP-compatible client. All you need to do is locate the client’s configuration file and paste in the JSON MCP configuration.

Using Apify Actors with LangGraph MCP Adapters

Finally, several AI agent frameworks are starting to adopt the MCP protocol and provide built-in support for accessing MCP servers directly from your code.

In the example below, we’re using LangGraph to initialize an MCP client, connect to the Actors MCP Server, and generate a list of the top 10 cafes in Boston using the available Actor tools. Here’s how it’s implemented:

import dotenv
import os
import asyncio
from pathlib import Path

from langgraph.prebuilt import create_react_agent
from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client
from langchain_mcp_adapters.tools import load_mcp_tools
from langgraph.prebuilt import create_react_agent
from langchain_openai import ChatOpenAI

# Load environment variables from .env file
dotenv.load_dotenv()

# Make sure to set your APIFY_TOKEN in the environment or .env file
APIFY_TOKEN = os.getenv("APIFY_TOKEN")
if APIFY_TOKEN is None:
    raise ValueError("APIFY_TOKEN environment variable must be set.")

# Initialize the model
# Note: You can use any model from langchain_openai
model = ChatOpenAI(model="gpt-4o")

# Initialize the server parameters
# Note: You can use any server parameters from mcp
server_params = StdioServerParameters(
    command="npx",
    args=["-y", "@apify/actors-mcp-server"],
    env={
        "APIFY_TOKEN": APIFY_TOKEN
    },
)

# Initialize the MCP client, load tools, and create an agent
async def main():
    async with stdio_client(server_params) as (read, write):
        async with ClientSession(read, write) as session:
            # Initialize the connection
            await session.initialize()

            # Get tools
            tools = await load_mcp_tools(session)

            # Create and run the agent
            agent = create_react_agent(model, tools)
            agent_response = await agent.ainvoke(
                {
                    "messages": "Give me a list with the top 10 cafes in Boston."
                }
            )

            # Get the response content
            response_content = agent_response["messages"][-1].content
            print(agent_response["messages"][-1].content)

            # Save the response to a markdown file
            output_dir = Path("results")
            output_dir.mkdir(exist_ok=True)
            filename = f"agent_response.md"
            file_path = output_dir / filename

            # Write to file
            with open(file_path, "w") as f:
                f.write(response_content)

            print(f"\\nResults saved to {file_path}")

# Run the async main function
asyncio.run(main())

As you can see, the server configuration uses the same parameters we used when setting it up in Claude Desktop.

server_params = StdioServerParameters(
    command="npx",
    args=["-y", "@apify/actors-mcp-server"],
    env={
        "APIFY_TOKEN": APIFY_TOKEN
    },
)

Also note that in this example, we’re using GPT-4o instead of Claude. Although MCP was introduced by Anthropic, it’s an LLM-agnostic protocol and works with any frontier model that supports it.

Once the code runs, the LLM’s response will be saved to a file named agent_response.md.

Agent response

Conclusion

The introduction of the Model Context Protocol marks a major step forward in the evolution of AI tooling. It brings much-needed standardization and pushes the ecosystem closer to maturity. And if Apify alone can expose over 4,500 tools through a single MCP server, imagine the scale and variety of tools that will become available as MCP adoption continues to grow.

This article is just a starting point for using Apify Actors to improve your AI applications. The space is evolving at a breakneck pace and we're committed to match this pace by continuously pushing new solutions, so we encourage you to keep exploring on your own. In particular, check out more code examples of the Actors MCP Server implementation using different transport methods (like SSE and Stdio) on GitHub:

And don’t forget to join the Apify Discord to connect with other AI developers and stay up to date with what we’re building!

Learn more about MCP and AI agents

Percival Villalva
Percival Villalva
Developer Advocate on a mission to help developers build scalable, human-like bots for data extraction and web automation.
On this page

Build the scraper you want

No credit card required

Start building

Related articles