Pilot Scope
Salesforce’s hosted MCP servers are in Pilot as of April 2026, with pricing still TBD. Three MCP servers exist today for development: the DX MCP Server (Developer Preview), plus others for Agentforce and Flow scenarios. The Pilot designation means feature-complete enough to use against scratch orgs and developer sandboxes, but not yet under SLA for production. Customer enrollment runs through your Salesforce account team; there’s no self-serve enrollment in Setup yet. Expect GA timing in Winter ‘26 release window with pricing announced 60 days ahead.
Salesforce DX MCP Server
Runs common developer tasks via natural language from Claude Desktop, Cursor, or other MCP clients: deploy code, create scratch orgs, run tests, query org metadata. Replaces shell commands for routine flows. Setup is two steps: install the server (npm install -g @salesforce/mcp-server-dx) and register it with your MCP client. Claude Desktop config:
{
"mcpServers": {
"salesforce-dx": {
"command": "sf-mcp-dx",
"args": ["--target-org", "my-scratch-org"]
}
}
}
Once registered, Claude responds to commands like:
> Run all Apex tests in my QA org and tell me which ones failed
[MCP] Executing: sf apex run test --target-org qa --code-coverage --result-format json
[MCP] 3 of 247 tests failed:
- OpportunityTriggerTest.testBulkInsert (line 45: assertion failed)
- AccountServiceTest.testMergeContacts (line 89: timeout)
- QuoteCalculatorTest.testRounding (line 112: precision mismatch)
The DX MCP Server exposes 28 tools as of April 2026, covering deploy, retrieve, scratch org lifecycle, test execution, metadata query, and Apex anonymous execution.
MuleSoft Anypoint Bridge
Anypoint can now convert any API or existing Mule application into an MCP server. Bridges your enterprise API estate to MCP-compatible agents without bespoke wrappers. The conversion happens through a new “Expose as MCP” option in Anypoint Studio — the bridge auto-generates MCP tool definitions from the API’s OpenAPI spec, applies your Anypoint API Manager policies (rate limits, auth, throttling), and exposes the result as an MCP endpoint. A REST API documented with OpenAPI 3.x becomes an MCP server in roughly 15 minutes including policy attachment and testing. The same auth and audit logs apply whether the API is called by a traditional integration or an MCP-mediated agent.
Governance
Admins get a central registry of MCP-defined tools with policy enforcement — rate limits, access controls, audit. Don’t let agents connect to MCP servers unreviewed; enterprise hygiene applies here as elsewhere. The registry lives at Setup > MCP > Connected Servers and tracks tool inventory, policy bindings (rate limits, allowed users), audit logs, and vendor trust posture. Production connections require admin approval and security review; sandboxes allow self-serve connection for testing.
What Changed in 2026
MCP went from a Anthropic-only protocol in late 2024 to a multi-vendor standard with adoption from Salesforce, Microsoft, Google, and most major SaaS vendors by Q1 2026. Salesforce’s Pilot is part of that broader inflection. The strategic implication: MCP becomes the integration protocol of choice for agent-mediated work, gradually displacing point-to-point REST integrations for use cases where an LLM is in the loop. REST APIs don’t go away; they become one transport that MCP servers wrap.
Common Failure Modes
Connecting MCP servers without rate limits in place. An overeager agent can blast a downstream API with hundreds of calls per minute, blowing past throttling and getting your service banned. Always set per-tool rate limits in the registry. Second: trusting the OpenAPI-to-MCP auto-generation without review. Generated tool descriptions from API specs are often vague (“returns customer data”) and confuse the LLM about when to call them. Hand-edit critical tools’ descriptions to match agent reasoning patterns. Third: forgetting that MCP tool calls are still Apex/API calls under the hood — they consume governor limits and API call quotas like any other integration.
What to do this week
Enroll one developer in the DX MCP Server Pilot through your account team. Install it locally, register it with Claude Desktop, and run one end-to-end task (test execution or scratch org creation). The 30-minute experiment surfaces both the value and the gaps before you scope wider rollout.