InvenTreeMCP

An MCP (Model Context Protocol) server for InvenTree, exposed as an InvenTree plugin. It lets MCP clients (Claude, other MCP-aware agents) query InvenTree inventory data over a Streamable HTTP endpoint.

Design

Every tool is a thin wrapper around InvenTree's own REST API view classes (see inventree_mcp/proxy.py), dispatched as the authenticated caller - MCP requests go through exactly the same permission checks, filtering, and serialization as the regular REST API. Tool code never queries the Django ORM directly.

Currently read-only, covering parts, stock items/locations, part categories, purchase/sales/return/ build orders (with line items and allocations), companies, contacts, addresses, manufacturer/ supplier parts, BOM items, attachments, parameters, stock tracking history, test results, and project codes. Once write tools land, the MCP_READ_ONLY setting (see Configuration) will block them by default regardless of the calling user's permissions.

Each tool's outputSchema and filter/ordering options are derived live from InvenTree's own serializers and views (not hand-maintained), so they can't drift as InvenTree evolves. Call describe_filters(resource) to see what's available for a given resource. The set of tools an MCP client sees is also filtered to what the calling user can actually use - though every call is still permission-checked in full regardless of what was advertised.

Setup

1. Install the plugin

Install via the InvenTree plugin manager, or via pip:

pip install inventree-mcp

Then enable the plugin under Admin > Plugins, and configure its settings (see Configuration below).

2. Create a token for your MCP client

Create an InvenTree API token for your MCP client.

3. Configure your MCP client

The endpoint is <your-inventree-server>/plugin/inventree-mcp/mcp/, using Streamable HTTP transport with an Authorization: Token <token> header.

For a client that supports remote Streamable HTTP servers directly, add:

{
  "mcpServers": {
    "inventree": {
      "url": "https://<your-inventree-server>/plugin/inventree-mcp/mcp/",
      "headers": {
        "Authorization": "Token <your-api-token>"
      }
    }
  }
}

For a client that only supports local (stdio) servers, bridge it with mcp-remote:

{
  "mcpServers": {
    "inventree": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://<your-inventree-server>/plugin/inventree-mcp/mcp/",
        "--header",
        "Authorization: Token <your-api-token>"
      ]
    }
  }
}

Configuration

Under Settings > Plugin Settings:

• Require Authentication (REQUIRE_AUTH, default True): reject unauthenticated requests. Only disable for local testing.
• Read Only (MCP_READ_ONLY, default True): block all write actions via MCP, regardless of the calling user's permissions. A plugin-wide kill switch, independent of per-user roles.

Authentication

Access follows the calling user's normal InvenTree role assignments. Supported auth methods:

• An InvenTree API token: Authorization: Token <token>.
• Basic auth (username/password).
• An OAuth2 bearer token: Authorization: Bearer <token>. A scoped token (e.g. r:view:part) narrows access below the underlying user's roles - useful for issuing an agent a tightly-scoped token without creating a separate low-privilege user.

Session/cookie auth is not supported (not meaningful for a machine client).