django-admin-mcp

Expose Django admin models to MCP (Model Context Protocol) clients via HTTP. Add a mixin to your ModelAdmin classes and get instant access to CRUD operations, admin actions, model history, and more.

✨ Features

• 📦 Zero dependencies — only Django and Pydantic required
• 🔐 Token authentication — secure Bearer token auth with configurable expiry
• 🛡️ Django admin permissions — respects existing view/add/change/delete permissions
• 🔒 Field filtering — control which fields are exposed via mcp_fields and mcp_exclude_fields
• 📝 Full CRUD — list, get, create, update, delete operations
• ⚡ Admin actions — execute registered Django admin actions
• 📦 Bulk operations — create, update, or delete multiple records at once
• 🔍 Model introspection — describe model fields and relationships
• 🔗 Related objects — traverse foreign keys and reverse relations
• 📜 Change history — access Django admin's history log
• 🔎 Autocomplete — search suggestions for foreign key fields

📥 Installation

pip install django-admin-mcp

Add to your Django project:

# settings.py
INSTALLED_APPS = [
    'django_admin_mcp',
    # ...
]

# urls.py
from django.urls import path, include

urlpatterns = [
    path('mcp/', include('django_admin_mcp.urls')),
    # ...
]

Run migrations to create the token model:

python manage.py migrate django_admin_mcp

🚀 Quick Start

1️⃣ Expose Your Models

Add the mixin to any ModelAdmin. Set mcp_expose = True to expose direct tools:

from django.contrib import admin
from django_admin_mcp import MCPAdminMixin
from .models import Article, Author

@admin.register(Article)
class ArticleAdmin(MCPAdminMixin, admin.ModelAdmin):
    mcp_expose = True  # Exposes list_article, get_article, etc.
    list_display = ['title', 'author', 'published']

@admin.register(Author)
class AuthorAdmin(MCPAdminMixin, admin.ModelAdmin):
    pass  # Discoverable via find_models, no direct tools

🔒 Protecting Sensitive Fields

Use mcp_exclude_fields to prevent sensitive data exposure:

@admin.register(User)
class UserAdmin(MCPAdminMixin, admin.ModelAdmin):
    mcp_expose = True
    # Never expose sensitive fields via MCP
    mcp_exclude_fields = ['password', 'security_token']

2️⃣ Create an API Token

Go to Django admin at /admin/django_admin_mcp/mcptoken/ and create a token. Tokens can optionally be tied to users, groups, or have direct permissions assigned.

3️⃣ Configure Your MCP Client

Add to your MCP client settings (~/.claude/claude_desktop_config.json or project .mcp.json):

{
  "mcpServers": {
    "django-admin": {
      "url": "http://localhost:8000/mcp/",
      "headers": {
        "Authorization": "Bearer YOUR_TOKEN"
      }
    }
  }
}

4️⃣ Use with Your Agent

Once configured, the agent can use the tools directly:

User: What models are available in Django admin?
Agent: [calls find_models tool]

User: Show me the latest 10 articles
Agent: [calls list_article with limit=10]

User: Get article #42 and update its title to "New Title"
Agent: [calls get_article with id=42, then update_article]

🛠️ Available Tools

For each exposed model (e.g., Article), the following tools are generated:

📝 CRUD Operations

🔍 Model Introspection

⚡ Admin Actions

🔗 Relationships

🌐 HTTP Protocol Reference

For custom integrations, the MCP endpoint accepts POST requests:

# List available tools
curl -X POST http://localhost:8000/mcp/ \
  -H "Authorization: Bearer TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"method": "tools/list"}'

# Call a tool
curl -X POST http://localhost:8000/mcp/ \
  -H "Authorization: Bearer TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"method": "tools/call", "name": "list_article", "arguments": {"limit": 10}}'

💬 Example Conversations

📝 CRUD Operations

User: Create a new article titled "Getting Started with Django"

Agent: I'll create that article for you.
[calls create_article with title="Getting Started with Django"]
Created article #15: "Getting Started with Django"

User: Update article 15 to add content

Agent: [calls update_article with id=15, content="..."]
Updated article #15 successfully.

User: Delete article 15

Agent: [calls delete_article with id=15]
Deleted article #15.

⚡ Admin Actions

User: Mark articles 1, 2, and 3 as published

Agent: [calls action_article with action="mark_as_published", ids=[1,2,3]]
Marked 3 articles as published.

When an admin action returns a Django HttpResponse / StreamingHttpResponse (typical for CSV/TSV/PDF downloads), action_<model> serializes the body instead of str(response):

{
  "success": true,
  "action": "export_csv",
  "affected_count": 2,
  "result": {
    "type": "file",
    "encoding": "utf-8",
    "content_type": "text/csv; charset=utf-8",
    "filename": "export.csv",
    "content_disposition": "attachment; filename=\"export.csv\"",
    "size": 10,
    "status_code": 200,
    "content": "a,b\r\n1,2\r\n"
  }
}

Text-ish content types (text/*, application/json, …) decode using the Content-Type charset (or response.charset, default UTF-8) and return "encoding": "utf-8" with Unicode text in content. If that decode fails (unknown charset or invalid bytes), the payload falls back to "encoding": "base64" like binary downloads. Binary payloads use "encoding": "base64" with ASCII base64 in content.

Large downloads are rejected once the body exceeds MCP_ACTION_MAX_FILE_BYTES (default 5 MiB) to protect MCP transports.

📦 Bulk Operations

User: Set status to "archived" for articles 10-15

Agent: [calls bulk_article with operation="update", ids=[10,11,12,13,14,15], data={"status": "archived"}]
Updated 6 articles.

User: Delete all draft articles from last month

Agent: [calls list_article to find drafts, then bulk_article with operation="delete"]
Deleted 12 draft articles.

🔗 Exploring Relationships

User: Show me all comments on article 42

Agent: [calls related_article with id=42, relation="comments"]
Found 8 comments on article #42...

User: What changes were made to article 42?

Agent: [calls history_article with id=42]
Change history for article #42:
- 2024-01-15: Changed title (admin)
- 2024-01-10: Created (admin)

🔍 Model Discovery

User: What can I manage through MCP?

Agent: [calls find_models]
Available models:
- article (5 tools: list, get, create, update, delete)
- author (5 tools: list, get, create, update, delete)
- category (5 tools: list, get, create, update, delete)

User: What fields does article have?

Agent: [calls describe_article]
Article fields:
- id (AutoField, read-only)
- title (CharField, max_length=200, required)
- content (TextField, optional)
- author (ForeignKey to Author, required)
- published (BooleanField, default=False)
- created_at (DateTimeField, auto)

🔐 Security

🏗️ Two-Level Exposure

Models with MCPAdminMixin are automatically discoverable via the find_models tool, allowing the agent to see what's available. To expose full CRUD tools directly, set mcp_expose = True:

# Discoverable via find_models only
class AuthorAdmin(MCPAdminMixin, admin.ModelAdmin):
    pass

# Full tools exposed (list_article, get_article, etc.)
class ArticleAdmin(MCPAdminMixin, admin.ModelAdmin):
    mcp_expose = True

🔑 Token Authentication

• 🎫 Tokens are created in Django admin
• 👤 Tokens can be associated with a user, groups, or have direct permissions
• 🚫 Tokens without any permissions have no access (principle of least privilege)
• ⏰ Token expiry is configurable (default: 90 days)
• 🗑️ Revoke tokens by deleting them in admin

🛡️ Permission Checking

All operations respect Django admin permissions:

If a token lacks permission, the operation returns an error.

📋 Requirements

✅ Supported Django Versions

Django 3.2 · 4.0 · 4.1 · 4.2 · 5.0

📄 License

MIT