MCP Server

MadeWithStack publishes a hosted MCP server at /mcp. It exposes the directory as a small, stable tool surface for agent runtimes that support native MCP tool calling.

Endpoint

• MCP endpoint: https://www.madewithstack.com/mcp
• Transport: Streamable HTTP
• Protocol version: 2025-11-25
• Discovery manifest: /.well-known/agents.json

Supported capabilities

This v1 server is intentionally narrow:

• tools only
• no prompts
• no resources
• no sampling
• no elicitation
• no session management
• no SSE stream endpoint

GET /mcp and DELETE /mcp return 405. Use POST /mcp for MCP JSON-RPC requests.

Initialization flow

1. Send initialize with protocol version 2025-11-25.
2. Send notifications/initialized.
3. Call tools/list or tools/call.
4. Include the MCP-Protocol-Version: 2025-11-25 header on post-initialization requests.

Tool inventory

The hosted MCP server currently exposes these tools:

• list_tools
• search
• list_products
• get_product
• submit_product

Those tools stay parallel to the public REST API. Input names and structured outputs are intentionally close to the existing /api/v1 contracts so clients can move between REST and MCP without rewriting business logic.

submit_product example

Call submit_product when an agent has a valid product homepage, submitter email, stack metadata, and review evidence ready.

{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tools/call",
  "params": {
    "name": "submit_product",
    "arguments": {
      "name": "AgentFlow",
      "url": "https://agentflow.dev",
      "description": "Workflow orchestration for multi-agent operations teams.",
      "email": "founder@agentflow.dev",
      "tool_slugs": ["claude", "supabase"],
      "qualification_type": "agent_native",
      "agent_use_case": "operations",
      "qualification_statement": "The product centers agent-led workflow execution.",
      "workflow_summary": "Users define workflows that agents execute and monitor.",
      "agent_tools_used": ["Claude", "Supabase"],
      "supporting_links": ["https://agentflow.dev/docs"]
    }
  }
}

A successful call returns a normal tool result with the public submission payload in structuredContent.

{
  "jsonrpc": "2.0",
  "id": 2,
  "result": {
    "isError": false,
    "structuredContent": {
      "success": true,
      "slug": "agentflow",
      "receiptSent": true,
      "badge_opt_in": false,
      "review_status_url": "https://www.madewithstack.com/api/v1/products/agentflow?email=founder%40agentflow.dev",
      "owner_claim": {
        "email": "founder@agentflow.dev",
        "status": "unverified",
        "verification_email_sent": true
      },
      "claim_status": "submitted",
      "claim_eligibility": "agent_directory",
      "next_action_code": "UNDER_EDITORIAL_REVIEW"
    },
    "content": [
      {
        "type": "text",
        "text": "Submission accepted for agentflow. The listing is now pending editorial review."
      }
    ]
  }
}

Error model

• MCP protocol errors are reserved for malformed JSON-RPC requests, unsupported methods, and bad protocol negotiation.
• Tool failures return normal tool results with isError: true.
• Structured error payloads include error_code, message, and retry_after_seconds when applicable.

Rate limits and trust model

• Read tools inherit the current anonymous catalog and search limits.
• submit_product inherits the public submission limits and duplicate protection rules.
• Accepted submissions enter manual editorial review. MCP submission does not bypass curation or publish instantly.

Security notes

• HTTP requests with invalid Origin values are rejected.
• Pending review status remains gated by submitter email on get_product.
• Treat the directory as a trust-first system: a listing is reviewed before publication, and tool access does not weaken editorial control.