Django Admin MCP

django-admin-mcp

[PyPI version] [PyPI downloads] [Python versions] [Django] [Tests] [codecov] [Documentation] [License] [MCP Registry] [Glama MCP server]

Add a mixin to your ModelAdmin and MCP clients get CRUD, admin actions and relationship traversal, inside Django's existing permissions. Only Django and Pydantic as dependencies.

[Meet django-admin-mcp: one mixin turns a ModelAdmin into MCP tools, every call goes through the admin's permissions, validation and audit log, one token per user, a conversation with the agent, and how to get started]

_{One mixin, the request's path through the admin, a token, a conversation with the agent, then getting started. 50 seconds · 1080p · 60 fps · original soundtrack. Watch the film.}

Why not just give the agent database access?

A database connection hands the agent raw tables. The Django admin is where your project already encodes who may touch what, how records are validated, and what the sanctioned bulk operations are. This package puts the agent behind that layer instead of around it:

• Permissions — every call goes through ModelAdmin.has_*_permission(). A token can only do what its linked user can do, and starts with no access until you grant it.
• Validation — writes run full_clean() and your save_model() / save_related() hooks, so the agent cannot save a record the admin form would reject.
• Audit history — every create, update, delete and action is written to Django's LogEntry under the token's user. You can see what the agent changed, and the agent can read the history too.
• Admin actions — "publish", "refund", "export CSV" and your other registered actions become tools, including two-step confirmation flows. The agent uses the operations you designed rather than inventing SQL.
• Scoped exposure — mcp_fields / mcp_exclude_fields keep password hashes and secrets out of the agent's view; get_queryset() still limits which rows it can see.

How it compares

A generic Django MCP server or a hand-rolled FastMCP wrapper asks you to define tools, schemas and permission checks yourself, one function at a time. This package reads the ModelAdmin classes you already maintain and generates the tools from them, so field choices, querysets, validation, actions and permissions stay in one place and the agent inherits every change you make to the admin. It also ships as a plain Django app with no MCP SDK or extra HTTP stack.

Installation

pip install django-admin-mcp

# settings.pyINSTALLED_APPS = [    'django_admin_mcp',    # ...]
# urls.pyurlpatterns = [    path('mcp/', include('django_admin_mcp.urls')),    # ...]

python manage.py migrate django_admin_mcp

Quick start

1. Expose a model. Set mcp_expose = True to generate tools for it. Models with the mixin but without the flag are discoverable through find_models only.

from django.contrib import adminfrom django_admin_mcp import MCPAdminMixinfrom .models import Article, Author
@admin.register(Article)class ArticleAdmin(MCPAdminMixin, admin.ModelAdmin):    mcp_expose = True                    # list_article, get_article, create_article, ...    mcp_exclude_fields = ['internal_notes']    list_display = ['title', 'author', 'published']
@admin.register(Author)class AuthorAdmin(MCPAdminMixin, admin.ModelAdmin):    pass                                 # discoverable, no direct tools

2. Create a token in the admin at /admin/django_admin_mcp/mcptoken/. Grant it exactly the permissions and groups the agent needs. The linked user caps what the token can do and is the identity its changes are logged under.

3. Point your MCP client at it, for example in a .mcp.json for Claude Code:

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

4. Ask the agent.

User: Show me the latest 10 articlesAgent: [calls list_article with limit=10]
User: Mark articles 1, 2 and 3 as publishedAgent: [calls action_article with action="mark_as_published", ids=[1, 2, 3]]
User: What changed on article 42?Agent: [calls history_article with id=42]

Tools

For each exposed model (here Article) the server generates:

Plus find_models to discover everything exposed, MCP prompts for common workflows, and read-only resources at models:// and data:// URIs. The endpoint speaks JSON-RPC 2.0 over HTTP with GET /mcp/health/ for health checks. Full reference in the docs.

Security in brief

• Tokens look like mcp_<key>.<secret>; only a salted hash of the secret is stored.
• Each token carries its own permissions and groups, intersected with its user's permissions. Deactivating the user disables its tokens.
• Expiry is configurable; revoke a token by deactivating or deleting it in the admin.
• Sensitive values are redacted from log messages, and large action downloads are capped by MCP_ACTION_MAX_FILE_BYTES (default 5 MiB).

See Permissions and Token Management for details.

Requirements

Python 3.10 to 3.14, Django 3.2 to 6.0, Pydantic 2. Tested against every Django release in that range.

Support the project

If this saved you time, a star on GitHub helps others find it. Issues and PRs are welcome; see the contributing guide.

[Star History Chart]

[PyPI Download History]

License

MIT