Skip to content

MCP (Model Context Protocol)

MCP (Model Context Protocol) is a standard way for AI clients (like Claude Desktop) to discover and call tools.

In SmartForge, an MCP server exposes a curated set of approved functions from a specific project + system so external clients can:

  • List the tools you published (names, descriptions, input schema)
  • Call those tools with validated parameters
  • Receive structured results back (tables/JSON), just like SmartForge Cortex

This page stays intentionally short on “what is MCP” and focuses on what it does in SmartForge and how to use it end-to-end.


MCP acts as a standard protocol layer between external AI clients and SmartForge’s governed enterprise functions:

  1. Discovery: AI clients automatically discover available tools with complete schema metadata
  2. Validation: All tool calls pass through SmartForge’s existing governance, permissions, and validation layers
  3. Execution: MCP routes requests to the appropriate function with configured connections
  4. Response: Structured results are returned in a standardized format compatible with all MCP clients

All existing enterprise controls, audit logging, and security policies apply identically to MCP calls.


Use MCP when you want SmartForge’s governed enterprise tools to be available outside the SmartForge UI:

  • Run enterprise lookups from a desktop AI client (e.g., check order status without switching apps)
  • Standardize access: one governed tool catalog for multiple clients (desktop, IDE, agents)
  • Enable automation: compose functions inside larger agent workflows

If your goal is only “chat with enterprise systems”, SmartForge Cortex is usually the fastest path. MCP is best when you need tool access in another client.


Prerequisites (what must exist before MCP will be useful)

Section titled “Prerequisites (what must exist before MCP will be useful)”

Before generating an MCP server, make sure these are already done:

  1. Project is connected to a functions repository
    • This is where your functions (code + metadata) live and are versioned.
  2. Functions are in a usable state
    • Good descriptions, correct inputs/outputs, and tested in Dev Console.
    • (Recommended) An Eval run looks good for the functions you plan to expose.
  3. Connections + credentials are configured
    • MCP tools call the same governed functions; they still need working connections.

  1. Go to MCP Servers
  2. Click Generate New MCP Server Link
  3. Select the system (for example: Enterprise)
  4. Enter a clear server name
  5. Click Generate Link

After generation, open the MCP server link to view details:

  • Server URL
  • Connections bound to this server/system
  • The tool list (your functions) that clients will discover and call

What the MCP Server Details page is telling you

Section titled “What the MCP Server Details page is telling you”

This is the contract external clients will see. Tools appear here only if they are:

  • In the selected project/system scope
  • Approved/published (depending on your org’s governance settings)
  • Valid enough to be exposed (metadata and schema are required)

The Export section is the safest source of truth for client setup. It typically provides:

  • Client-specific configuration snippets
  • Any required auth headers/tokens (if your org requires them)
  • Notes on supported transport (HTTP/SSE/stdio) for that client

Setup via custom connector (Claude example)

Section titled “Setup via custom connector (Claude example)”
  1. Open claude.ai → Profile → Settings → Connectors
  2. Click Add custom connector at the bottom
  3. Paste your generated MCP Server URL
  4. Click Add and complete authentication if prompted

✅ Once connected, server tools appear in the chat attachment menu (paperclip icon).you can also manage permissions from Connectors settings.

🎥 Full video walkthrough with connection demo