MCP

Model Context Protocol (MCP) is an advanced framework that extends your AI agent's capabilities beyond what traditional tools can provide. MCPs enable agents to interact with complex systems, maintain state across interactions, and access specialized services that require multi-step operations.

What is MCP?

MCP (Model Context Protocol) is a protocol developed by Anthropic that provides a standardized way for AI agents to connect to external data sources and services. Unlike simple tools that perform single operations, MCPs can:

Maintain state: Remember information across multiple tool calls within a session
Provide complex functionality: Offer multi-step operations that would be cumbersome with individual tools
Stream data: Handle real-time data streams and long-running operations
Bundle capabilities: Package related functionality into cohesive units

Think of MCPs as "super tools" - they're more powerful and flexible than regular tools, but also more complex to set up and configure.

MCPs vs Tools

Understanding the difference between MCPs and tools helps you choose the right approach for your use case:

Aspect	Tools	MCPs
Complexity	Simple, single-purpose functions	Complex, multi-capability systems
State	Stateless (each call is independent)	Stateful (can maintain context)
Setup	Easy - just add to agent	Requires server/service setup
Use Case	Fetch data, perform single actions	Browser automation, database access, complex workflows
Examples	listAccountContacts, draftEmailToSend	Browser MCP, Database MCP, API Gateway MCP

Available MCPs in Bigmind

Bigmind currently supports the following MCPs:

Browser MCP

The Browser MCP enables agents to browse websites, interact with web applications, and extract information from the internet - similar to how a human would use a web browser.

Capabilities:

Navigate: Visit any website or web application
Interact: Click buttons, fill forms, scroll pages
Extract: Pull data from web pages (text, tables, structured data)
Screenshot: Capture visual representations of web pages
Observe: Monitor changes on web pages
Think: Reason about web page content and structure

Use cases:

Research competitor websites for pricing and features
Monitor prospect company news and updates
Extract publicly available company information
Validate website functionality and availability
Gather intelligence from industry publications

Important notes:

Browser MCP is resource-intensive and should be used judiciously
Agents are instructed to close browser sessions when done to conserve resources
Currently available only to whitelisted organizations (contact support for access)

Adding MCPs to Agents

To add MCPs to an agent:

Open the agent configuration in Settings → AI → Agents
Navigate to the MCPs tab
Click "Add MCP"
Select the MCP you want to add from the available list
Save the agent configuration

Once added, the MCP's capabilities become available to the agent during conversations.

Background Agent MCPs

Like tools, MCPs can be configured separately for background agents. When you toggle to "Background Agent" mode in the agent configuration, you can add MCPs specifically for automated workflows:

Internal MCPs: Configured separately from foreground agent MCPs
Automation-optimized: Selected for reliability and efficiency in automated contexts
No user interaction: Background agent MCPs operate autonomously

System Prompts from MCPs

When you add an MCP to an agent, the MCP can provide additional system prompt instructions that guide the agent on how to use the MCP's capabilities effectively. These instructions are automatically appended to the agent's main system prompt.

For example, the Browser MCP adds this instruction:

You MUST close the session when you are done with the task, otherwise we will be using up resources.

These MCP-provided instructions ensure agents use the capabilities responsibly and effectively.

Technical Implementation

For developers interested in how MCPs work in Bigmind:

MCP Architecture

Transport: MCPs use StreamableHTTPClientTransport for communication
Client: Bigmind uses the Vercel AI SDK's experimental MCP client
Integration: MCP tools are merged with regular tools at runtime
Lifecycle: MCP clients are created when agents are initialized

MCP Configuration

MCPs are defined in the codebase with the following properties:

ID: Unique identifier for the MCP
Name: Human-readable name displayed in the UI
Description: Explanation of the MCP's capabilities
Enabled: Whether the MCP is active
Whitelisted tenants: Optional list of organizations/users with access
System prompt: Additional instructions for agents
Client factory: Function that creates the MCP client connection

Best Practices

Use sparingly: MCPs are powerful but resource-intensive. Only add them when simpler tools won't suffice
Test thoroughly: MCP operations can be complex - test extensively in the agent testing interface
Monitor usage: Keep an eye on how often MCPs are being called and optimize if needed
Provide context: In your system prompt, give agents clear guidance on when to use MCP capabilities
Handle failures: MCPs can fail (network issues, resource constraints) - ensure your agent can handle failures gracefully

Future MCPs

Bigmind is continually expanding MCP support. Future MCPs may include:

Database MCPs: Direct database query capabilities
API Gateway MCPs: Unified access to multiple APIs
File System MCPs: Advanced document processing and storage
Custom MCPs: Build your own MCPs for specialized use cases

Contact your Bigmind representative if you have specific MCP requirements for your organization.

Troubleshooting

MCP not available:

Check if the MCP is enabled for your organization
Some MCPs require whitelisting - contact support for access
Verify you have the necessary permissions in your organization

MCP operations failing:

MCPs depend on external services - check network connectivity
Browser MCP: Ensure the target website is accessible and not blocking automated access
Review error messages in the chat interface for specific failure reasons

Agent not using MCP:

Verify the MCP is added to the agent in the MCPs tab
Check that your request requires the MCP's capabilities
Add specific instructions in your system prompt about when to use the MCP
Test with direct requests that clearly need the MCP (e.g., "Visit example.com and tell me what their pricing is")