django-silk-mcp

An MCP server that exposes django-silk profiling data as tools for any MCP-compatible AI coding assistant. Enables query-level investigation and optimization directly from your conversation — N+1 detection, EXPLAIN ANALYZE, over-fetch analysis, and Python profiling without leaving your editor.

Installation

pip install django-silk-mcp
# or
uv add django-silk-mcp

Configuration

1. Add to INSTALLED_APPS

INSTALLED_APPS = [
    ...
    "silk",
    "django_silk_mcp",
]

2. Configure your MCP client

Expose the MCP server as a URL inside your Django app. Works on both WSGI and ASGI with no extra infrastructure.

Add to your root urls.py:

urlpatterns = [
    ...
    path("silk/", include("silk.urls", namespace="silk")),
    path("silk/mcp", include("django_silk_mcp.urls")),
]

This serves the MCP endpoint at /silk/mcp. The MCP server runs as part of your Django app — no separate process needed, it shares the same port as your dev server. Then configure your MCP client:

Claude Code:

claude mcp add silk-mcp --transport http http://localhost:8000/silk/mcp

Cursor: Go to Settings → Toos & MCPs → New MCP Server and add a new server with the following configuration:

{
  "mcpServers": {
    "silk-mcp": {
      "url": "http://localhost:8000/silk/mcp"
    }
  }
}

The same URL works with any MCP-compatible AI tool.

Security

Warning: The MCP endpoint uses AllowAny with no authentication by default, because MCP clients (Claude Code, Cursor, etc.) connect over plain HTTP and do not send credentials. This is intentional for local development but must not be exposed in production.

Silk profiling data includes raw SQL query strings, which can contain business data at runtime (user IDs, emails, search terms, etc.). Anyone who can reach the Django host can read this data through the MCP endpoint.

Recommended: Only mount in DEBUG mode:

# urls.py
from django.conf import settings

if settings.DEBUG:
   urlpatterns = [
     path("silk/", include("silk.urls", namespace="silk")),
     path("silk/mcp", include("django_silk_mcp.urls")),
   ]

Tools

Tool	Purpose
get_most_expensive_endpoints	Rank all endpoints by average DB time
get_request_time_breakdown	DB% vs Python% — confirm where time is spent
get_duplicate_queries	Detect N+1 patterns
get_query_sources	Which code lines triggered each query
get_overfetched_fields	SQL columns vs serializer fields — .only() candidates
explain_slow_queries	EXPLAIN (ANALYZE, BUFFERS) on slow SELECTs
get_request_queries	All SQL for a specific request
get_slow_queries	Slowest individual queries across all requests
get_slow_requests	Slowest HTTP requests
compare_requests	Before/after comparison with burst deduplication
get_python_profiles	@silk_profile decorated block timings
get_cprofile_hotspots	cProfile hotspots — no decorators needed

Usage

# Hit an endpoint first so Silk has data
curl http://localhost:8000/api/your-endpoint/

# Then ask your AI assistant
# "What are the slowest endpoints?"
# "Find N+1 queries in /api/your-endpoint/"
# "Run EXPLAIN on the slow queries for /api/your-endpoint/"