Governing Salesforce MCP Servers with MuleSoft Omni Gateway
Overview
Salesforce ships four hosted MCP Servers that any AI agent can call:
| MCP Server | Capabilities |
|---|---|
| SObject All | Full CRUD operations + SOQL queries across all Salesforce objects |
| SObject Reads | Read-only access to Salesforce data — safer for reporting agents |
| Data 360 | Unified customer data including cross-cloud identity resolution |
| Tableau Next | Analytics and KPI queries, chart data, and dashboard summaries |
These servers are powerful — and that’s exactly the problem. Without governance, any AI agent with valid credentials can read or modify arbitrary Salesforce data, exfiltrate customer PII, exhaust your LLM token budget with unconstrained query calls, and leave no audit trail for cost attribution or compliance.
MuleSoft Omni Gateway (formerly Omni Gateway) is the enforcement point. It sits between the AI client and the Salesforce MCP Server and runs the MCP policy stack — intercepting every tool call before it reaches Salesforce and every response before it reaches the agent.
Architecture
[AI Agent / Claude Desktop]
│ MCP calls
▼
[Omni Gateway — MCP policy stack]
├── MCP Support (protocol)
├── MCP PII Detector
├── MCP Global Access
├── MCP ABAC (Cedar)
├── MCP Payload Optimization
└── Agent Connection Telemetry
│
▼
[Salesforce MCP Server]
├── SObject All
├── SObject Reads
├── Data 360
└── Tableau Next
The gateway is the only component your AI agents talk to. They never connect directly to the Salesforce MCP Server URLs. This means:
- Policy changes take effect immediately — no agent reconfiguration needed
- Credentials stay in the gateway — agents present gateway credentials, not Salesforce credentials
- Every tool call is logged — the Agent Connection Telemetry policy captures client identity, tool name, token counts, and latency for every request
What you will build
In this workshop you will:
- Register the Salesforce SObject All and Data 360 MCP Servers in Anypoint Exchange and API Manager
- Create Omni Gateway API instances that front both servers
- Apply the full 6-policy MCP governance stack
- Configure Cedar-based ABAC to restrict which tools each agent can call
- Measure and reduce token costs using MCP Payload Optimization and Tool Mapping
- Verify governance is working end-to-end from a real MCP client
Time breakdown
| Section | Duration |
|---|---|
| Prerequisites | 15 min |
| Connect (find, register, configure gateway) | 45 min |
| Govern (apply policies + cost attribution) | 60 min |
| Test and Verify | 20 min |
| Summary | 5 min |
| Total | ~3 hours |
Continue to Prerequisites →