Deep Agent MCPs: Superpowers for Your God-Tier Agent

Deep Agent now supports Model Context Protocols (MCPs), enabling seamless integration with external tools and data sources. MCP standardizes how AI Workflows interact with resources, making it easier to enhance Deep Agent's capabilities with real-time data, APIs, and services. This guide walks you through setting up and using MCPs in Deep Agent to create powerful, context-aware AI workflows.

What Are MCPs?

The Model Context Protocol (MCP) is an open standard that connects AI models, like those in Deep Agent, to external systems via a client-server architecture. Think of MCP as a universal connector, allowing Deep Agent to access tools (e.g., GitHub, databases) and resources (e.g., files, APIs) without custom integrations. With MCP, you can enable Deep Agent to perform tasks like querying databases, fetching web content, or automating workflows.

Prerequisites

• A Deep Agent account with access to the platform.
• API keys or credentials for external tools you want to integrate (e.g., GitHub, Notion).
• Access to an MCP server (community-built or custom) or the ability to set one up.

Note: You can find a list of community-built MCP servers at GitHub and mcp.so. The MCP directory platforms will guide you on how to obtain the necessary tokens or API keys to configure your servers. You can pick your favorite MCPs from these platforms and get started quickly.

Step-by-Step Guide to Using MCPs in Deep Agent

Step 1: Go to MCP Settings Config Page

1. Log in to your Abacus.AI account and navigate to the Deep Agent Homepage.
2. You may directly click on Configure MCP button.

3. Or Go to the Deep Agent Settings page via the top-right corner menu, then select MCP Server Config from the sidebar.

Step 2: Install and Configure an MCP Server

1. Choose an MCP server that suits your needs. (e.g., Create a repository in github, fetch content from the web). Note that we currently support both local servers (which can be pulled from npm or pypi) and remote servers that are already deployed and ready to use.
2. Any server which requires interaction with the users's local filesystem will now work, as we run the MCP servers in an isolated environment without access to the local filesystem.
3. Copy the server’s config JSON and paste it into the JSON Config settings page.
4. Add a new MCP server config JSON based on its transport type:
   1. Stdio: For local servers, provide the required parameters - command, args, env, etc. Example: "command": "npx", "args": ["-y", "@modelcontextprotocol/server-github"], etc.
   2. SSE: For remote servers, provide the server URL and other optional parameters. Example: http://example.com:8000/sse
5. Configure the server with necessary credentials or environment variables (e.g., API keys), and make sure you are authenticated on the remote server platform when using remote servers.
6. Deep Agent will query the server to list available tools and resources.

Example Config:

The example below shows the JSON configuration for two MCP servers: GitHub and Google Tasks.

• Each server config JSON starts with the server name ( github and google_tasks ).
• Multiple server configurations are separated by a comma.

{
  "github": {
    "command": "npx",
    "args": ["-y", "@modelcontextprotocol/server-github"],
    "env": {
      "GITHUB_PERSONAL_ACCESS_TOKEN": "<YOUR_TOKEN>"
    }
  },
  "google_tasks": {
    "url": "<REMOTE_SERVER_URL>"
  }
}

Step 3: Use MCP Tools in Deep Agent

1. In your Deep Agent chat interface, instruct the agent to use the MCP tool. For example: 'you can say this to test Playwright MCP server, e.g - Access this given website and tell me its structure using Playwright’s tools'
2. Deep Agent will display a call to the tool having the requested parameters.
3. View the tool's output in the chat, which Deep Agent can use for further tasks (e.g., summarizing content).

FAQs and Troubleshooting

Why isn't my MCP server connecting?

• Please ensure that the command or URL in config json is correct and the json format is as shown in the example config, and verify that the server is running.
• A common mistake is copying a sample config like the one below without replacing placeholder values, such as the authentication token:

"github": {
    "command": "npx",
    "args": ["-y", "@modelcontextprotocol/server-github"],
    "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "<YOUR_TOKEN>"
    }
}

• If required authentication parameters (e.g., access tokens) are missing or not correctly filled in, the MCP server will fail to connect. Always make sure to generate and insert valid credentials as per the instructions provided on the respective MCP directory or website.
• Some platforms may wrap the json config within this given structure. When copying the config, make sure to include only the server configuration JSON (i.e, <your server config json> part in below example) —not the wrapper. You may refer to the example config json above.

{
    "mcpServers": {
        <your server config json>
    }
}

Can I use multiple MCP servers?

Yes, Deep Agent supports adding up to 5 servers and supports up to 50 tools across the servers. Add each server in the Integrations settings, but limit active tools to avoid overwhelming the LLM.

How do I secure my MCP connections?

Use MCP servers from trusted sources only and try to follow OAuth authentication in remote servers.