Knowledge Base Academy Community My Tickets

Home Knowledge Base Academy Community My Tickets

Fabric 8.4

Data Product as MCP Server: Overview

Starting with Fabric 8.4, every Data Product (Logical Unit) deployed is automatically exposed as an MCP (Model Context Protocol) server. This enables any MCP-compatible AI client to interact with the organization's enterprise data through a standardized, secure protocol, without requiring custom MCP integrations per data source.

Overview

The Model Context Protocol (MCP) is an open standard that defines how AI applications communicate with data sources. K2view Fabric leverages its data-product architecture to natively support MCP: each Logical Unit becomes an MCP server upon deployment, exposing its data and services as MCP resources, tools, and prompts.

Since each Logical Unit Instance (LUI) represents a specific business entity (e.g., customer #1, order #5042), the MCP server operates at the entity level — creating an on-the-fly, ad-hoc MCP server per entity instance. This follows Fabric's unique MicroDB architecture, which provides high security and data isolation by design, preventing cross-entity data leakage. In the AI world, this isolation becomes even more critical.

Notes:

System Logical Units — k2_ws and k2_ref — are automatically excluded, so that MCP servers are not created for them.

While the higher security level and isolation are achieved when the MCP server endpoint is defined at the client with the LUI instance ID (IID) as part of the URL, this is not available in many AI applications and clients — hence the attach IID tool can be used.

Key Concepts

Fabric Concept	MCP Mapping	Description
Logical Unit (LU/LUT)	MCP Server	The data product definition becomes the server
LU Tables	MCP Resources	Table metadata exposed as readable resources (markdown format)
Broadway Flows (tagged `mcp-tool`)	MCP Tools	Flows tagged with `mcp-tool` become callable tools
Broadway Flows (tagged `mcp-prompt`)	MCP Prompts	Flows tagged with `mcp-prompt` become callable prompts
Built-in Operations	MCP Tools	query, readTable, describeTables, listTables, attach
Prompts	MCP Prompts	Predefined prompts (e.g., `queryAssistant`)
LUI (entity instance)	MCP Session Scope	Each session is scoped to a specific LU + entity ID

How It Works

Automatic Server Creation

When a Logical Unit is deployed, Fabric automatically:

Creates an MCP server for the LU on each Fabric node (during the LU init process)
Registers a servlet with appropriate rewrite rules
Publishes the server, making it available to MCP clients

On redeployment, the existing MCP server is updated in-place: Broadway tools and prompts are dynamically refreshed — newly tagged flows are registered and flows whose tag was removed are unregistered. The server itself and active sessions remain running; clients do not need to reconnect. The 5 built-in tools are unaffected by redeployment.

Connection Flow

MCP Client                    Fabric (MCP Router)              MCP Server
    |                                |                              |
    |-- POST /mcp/customer/1 ------->|                              |
    |   {initialize}                 |-- Authenticate ------------->|
    |                                |-- Authorize (role check) --->|
    |                                |                              |-- Create session
    |                                |                              |   (LU=customer, IID=1)
    |<-- {Mcp-Session-Id=1234} ------|<-----------------------------|
    |                                |                              |
    |-- POST /mcp/customer/1 ------->|                              |
    |   Mcp-Session-Id=1234          |-- Authenticate ------------->|
    |   {tools/call...}              |-- Authorize (role check) --->|
    |                                |-- Route to session node ---->|
    |                                |                              |-- Execute tool
    |<-- {tool response} ------------|<-----------------------------|

Each request is authenticated and authorized. The Fabric router identifies the correct cluster node owning the session and routes accordingly.

Access Patterns

Base URL

The MCP server uses Streamable HTTP transport. The primary URL pattern is:

{fabric-url}/mcp/{data_product}/{iid}

Example:

https://your-fabric-host.com/mcp/customer/1

A single Fabric server hosts multiple MCP servers (one per deployed LU) and routes incoming requests to the appropriate one.

Alternative Access Modes

Mode	URL Pattern	Notes
Standard (recommended)	`/mcp/{data_product}/{iid}`	Entity ID in the URL path — most secure
IID as query parameter	`/mcp/{data_product}?iid={iid}`	IID extracted from query param
Without IID	`/mcp/{data_product}`	Only tools that don't require an instance will work; entity-scoped operations return errors. To associate an IID, use the `attach` command with the IID as variable
Without IID, with JWT claim	`/mcp/{data_product}` JWT claim with patten: `k2_data_product_{lu_name}: "{instance_id}"`	Fabric extracts the IID from the user’s JWT claims and set it as the MCP session IID. Read here for more information about JWT claims

Discovery Endpoint

GET /mcp is a plain HTTP endpoint (not an MCP protocol call) that returns a JSON listing of all active MCP servers and their capabilities — useful for discovering available data products before establishing an MCP connection.

GET /mcp

The response includes, for each data product server:

dataProduct — data product name
endpoint — the MCP protocol URL for this data product (e.g., /mcp/customer); this is the URL MCP clients connect to
tools — available tools with names, descriptions, and input schemas
resources — static resources (schema, tables)
resourceTemplates — dynamic resource templates (e.g., per-table metadata)
prompts — available prompts with names, descriptions, and arguments

Operational Notes

Deployment: MCP servers are created/updated automatically on LU deploy. No manual configuration is required.
Excluded LUs: System LUs (k2_ws, k2_ref) are automatically filtered out and hence no MCP servers are created for them.
Multi-node: Each Fabric node in the cluster exposes MCP servers. The router handles session affinity.
Redeployment: The MCP server is updated in-place on LU redeploy. Active sessions remain alive and clients do not need to reconnect. Only Broadway tools and prompts are refreshed.
Common tables: Reference/common tables are included in each LU's MCP server, providing access to shared lookup data alongside entity-specific data.
Session management: Sessions are managed with configurable cache size and expiration. Reconnecting clients receive new sessions, and stale sessions are cleaned up.