Cortex Context MCP Adapter

MCP (Model Context Protocol) adapter for Cortex Context integration.

Getting Started

To use this adapter, you need to be a registered user. Please visit our landing page to sign up and get your credentials:

https://mcp.cortexguardai.com/

Installation

You can use either pnpm or npm.

pnpm:

pnpm install @cortexguardai/mcp

npm:

npm install @cortexguardai/mcp

Important: How MCP Adapters Work

The MCP adapter communicates via stdin/stdout using the MCP protocol. When you run it directly from the command line, it will appear to "hang" - this is normal behavior. The adapter is waiting for MCP protocol messages from a compatible client.

Usage

With Claude Desktop (Recommended)

1. Configure your mcp.json file.

For pnpm users:

{
  "mcpServers": {
    "cortex-context": {
      "command": "pnpm",
      "args": [
        "dlx",
        "@cortexguardai/mcp@latest",
        "--token", "your-auth-token",
        "--project-id", "your-project-id"
      ]
    }
  }
}

For npm/npx users:

{
  "mcpServers": {
    "cortex-context": {
      "command": "npx",
      "args": [
        "@cortexguardai/mcp@latest",
        "--token", "your-auth-token",
        "--project-id", "your-project-id"
      ]
    }
  }
}

2. Restart Claude Desktop
3. The adapter will be available as a context source

Direct Command Line (for testing only)

pnpm:

pnpm dlx @cortexguardai/mcp --token <auth-token> --project-id <project-id>

npx:

npx @cortexguardai/mcp --token <auth-token> --project-id <project-id>

Note: When run directly, the adapter will start and then wait for MCP messages. This is expected behavior, not an error.

Testing the Adapter

To verify the adapter is working correctly, you can use the test client:

# From the project root
node test-mcp-client.js

This will spawn the adapter, send an MCP initialize request, and confirm it responds correctly.

Configuration Options

• --token, -t: Authentication token (required)
• --project-id, -p: Project ID to scope the adapter to (required)
• --timeout: Request timeout in milliseconds (default: 30000)
• --verbose, -v: Enable verbose logging (default: false)

Troubleshooting

"The adapter appears to hang"

This is normal! The MCP adapter uses stdin/stdout communication and waits for MCP protocol messages. It only responds when a compatible MCP client (like Claude Desktop) sends requests.

Schema Validation Errors

If you encounter invalid_literal errors expecting "object" in tool inputSchema:

• This was fixed in version 1.0.5+ by using explicit string literals in JSON Schema definitions
• Ensure you're using the latest version:
  • pnpm add -g @cortexguardai/mcp@latest
  • npm install -g @cortexguardai/mcp@latest
• Rebuild the adapter if developing locally:
  • pnpm run build
  • npm run build

Testing connectivity

You can test if your server is accessible:

curl -I "https://cortex-context-mcp.vercel.app/api/mcp"

Should return a 401 (authentication required) response, confirming the endpoint exists.

Development

# Build the adapter
pnpm run build
# or
npm run build

# Run in development mode
pnpm run dev
# or
npm run dev

Context-First Workflow and Tool Selection

To improve tool selection and make decisions more direct, this adapter exposes tools with prompt-like descriptions that guide models toward a consistent workflow:

• get_contexts

  • Selection hint: Start here.
  • Purpose: Check availability and fetch context file metadata for the current project.
  • Behavior: Returns a JSON list of context file metadata (id, name, type, size, timestamps) - content is NOT included. If empty, proceed to generate_initial_context.

• get_file

  • Purpose: Read the contents of a specific context file by its UUID.
  • Usage: Call after get_contexts when you know which file aligns with the current task. This is required to access file content.

• add_file

  • Purpose: Add a new context file when you already have prepared content.
  • Hint: Prefer generate_initial_context for the first project file. Provide filename, content, and optional logical type.

• generate_initial_context

  • Selection hint: Use when get_contexts returns no files.
  • Purpose: Create the first context file with a concise project overview (codebase structure, key modules, workflows).
  • Inputs: content (required), optional filename (defaults to project-context.md), optional file_type.

Recommended Decision Flow

• Step 1: Call get_contexts first to see available context files (metadata only).
• Step 2a: If contexts exist, pick the relevant file(s) and call get_file(id) to retrieve content.
• Step 2b: If no contexts exist, synthesize a brief overview from the codebase and call generate_initial_context.
• Step 3: When adding additional files, call add_file with the prepared content.

Important: The workflow now requires two API calls to access file content - first get_contexts for discovery, then get_file for content retrieval. This improves performance for projects with many or large context files.

These descriptions act like lightweight prompts to help models prioritize and focus the use of MCP tools appropriately.