MCP Integration

What is MCP?

The Model Context Protocol (MCP) allows AI agents to use ClassifAI as a tool for routing, decision-making, and content analysis.

When integrated, MCP clients can classify content, get project statistics, and submit feedback—all while conversing with users.

Note: These instructions are specifically for Claude Desktop, but ClassifAI can be integrated with any MCP client that supports the Model Context Protocol.

Quick Setup (Claude Desktop)

1. Get Your API Key

Sign up to get your API key from the dashboard.

2. Add to Claude Desktop

Add ClassifAI to your Claude Desktop MCP configuration:

macOS/Linux: Edit ~/Library/Application Support/Claude/claude_desktop_config.json

Windows: Edit %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "classifai": {
      "url": "https://api.classifai.dev/mcp",
      "headers": {
        "X-API-Key": "your-api-key-here"
      }
    }
  }
}

3. Restart Claude Desktop

Close and reopen Claude Desktop to load the MCP server.

4. Verify Connection

Ask Claude: "What classification tools do you have access to?"

Claude should list the ClassifAI tools:

• classify - Classify text/images for routing and decision-making
• classify_ground_truth - Submit feedback to improve accuracy
• classify_stats - Get project analytics and metrics

Available Tools

classify

Classify content for routing, decision-making, sentiment analysis, or content moderation.

Example conversation:

You: "Classify this customer message: 'My order hasn't arrived and it's been 2 weeks!'"

Claude: [Uses classify tool] This appears to be an urgent complaint about shipping delays. I've classified it as "urgent_shipping_issue" with high confidence. This should be routed to your shipping support team immediately.

Parameters:

• content - Array of text/image content items
• labels (optional) - Explicit classification labels
• description (optional) - Description for automatic label inference
• project_id (optional) - Reuse existing project labels

classify_ground_truth

Submit the correct label for a previous classification to improve accuracy.

Example conversation:

You: "Actually, that was correctly classified. Mark it as accurate."

Claude: [Uses classify_ground_truth tool] I've recorded that feedback. The system will learn from this to improve future classifications for your project.

Parameters:

• detection_id - ID from previous classification
• ground_truth - Correct label

classify_stats

Get analytics and accuracy metrics for a classification project.

Example conversation:

You: "How accurate is my support routing system?"

Claude: [Uses classify_stats tool] Your "support-routing" project has processed 1,234 classifications with an accuracy rate of 89% based on 150 feedback submissions. The most common category is "technical_support" (42%).

Parameters:

• project_id - Project to get statistics for

Use Cases

Customer Support Routing

Claude can automatically route customer inquiries:

You: "New ticket from john@example.com: 'I can't log into my account'"

Claude: [Uses classify tool]
This is a technical/authentication issue. I've classified it as "technical_auth"
and here's what I recommend:

1. Route to: Technical Support Team
2. Priority: High (login issues prevent product usage)
3. Suggested response: Send password reset instructions

Would you like me to draft a response?

Content Moderation

Claude can help moderate user-generated content:

You: "Check if this comment violates our policies: [comment text]"

Claude: [Uses classify tool]
I've analyzed the comment and classified it as "hate_speech" with 92% confidence.
This violates your content policy. I recommend:

1. Remove the comment
2. Send warning to user
3. Flag account for review

Should I proceed with these actions?

Sentiment Analysis

Claude can analyze customer feedback trends:

You: "Analyze today's customer reviews"

Claude: [Uses classify tool for each review, then classify_stats]
Based on 45 reviews classified today:

- Positive: 62% (mainly praising fast shipping)
- Neutral: 28% (mixed feedback on pricing)
- Negative: 10% (quality concerns on Product X)

The overall sentiment is positive. I notice an uptick in negative reviews for
Product X specifically—you may want to investigate the quality issues there.

Advanced Usage

Multi-Step Workflows

Claude can chain multiple tool calls:

You: "Process this support ticket and track the outcome"

Claude:
1. [Uses classify] Classification: "billing_issue"
2. [Provides response to customer]
3. You confirm it was helpful
4. [Uses classify_ground_truth] Feedback recorded
5. [Uses classify_stats] Updated project accuracy: 91%

Project Management

Track classification accuracy over time:

You: "Give me a weekly report on classification performance"

Claude: [Uses classify_stats for each project]

Weekly Classification Report:
├─ support-routing: 1,234 classifications, 89% accuracy
├─ spam-detection: 5,621 classifications, 94% accuracy
└─ sentiment-analysis: 892 classifications, 87% accuracy

Overall performance is strong. Spam detection has the highest accuracy due
to consistent feedback. Consider providing more feedback on sentiment analysis
to improve its accuracy.

Rate Limits

MCP requests follow the same rate limits as direct API usage:

Troubleshooting

Claude Can't See ClassifAI Tools

1. Check your claude_desktop_config.json syntax
2. Ensure the URL is correct: https://api.classifai.dev/mcp
3. Restart Claude Desktop completely
4. Check Claude Desktop logs: Help → View Logs

Rate Limit Errors

If you see "Rate limit exceeded":

• Reduce request frequency
• Upgrade to a higher tier for increased limits

Classification Errors

If classifications seem inaccurate:

• Provide feedback using classify_ground_truth
• Use consistent project_id values
• Provide more descriptive labels or descriptions
• Submit more ground truth feedback to improve accuracy

Using ClassifAI with Other MCP Clients

While these instructions are tailored for Claude Desktop, ClassifAI can be integrated with any MCP client that supports the Model Context Protocol. The MCP server is available at:

https://api.classifai.dev/mcp

When configuring other MCP clients, ensure you:

1. Set the server URL to https://api.classifai.dev/mcp
2. Include your API key in the X-API-Key header
3. Consult your MCP client's documentation for specific configuration instructions

Next Steps

• API Reference - Direct API usage
• Examples - Code samples and use cases
• Getting Started - Basic setup guide

Learn More About MCP

• MCP Documentation
• Anthropic MCP Announcement