If the Salesforce DX MCP Server is the tool that helps developers ship faster, Hosted MCP Servers are the tool that helps business teams use AI against real Salesforce data without a six-month integration project.
They entered Beta in late 2025, and they are — quietly — the most important piece of the MCP story for non-developers.
What They Actually Are
A Hosted MCP Server is an MCP endpoint that Salesforce runs for you. You don’t install anything. You don’t write any integration code. You point an MCP-speaking client (Claude.ai, ChatGPT desktop, an internal assistant, an Agentforce agent) at a Salesforce-issued URL, you go through a normal OAuth flow using a connected app, and the client lights up with a curated set of MCP tools that wrap specific Salesforce APIs.
It’s “MCP-as-a-Service” for the Salesforce platform. The API mapping, rate limits, audit, and authentication are all Salesforce’s problem — your job is to decide which servers to turn on and who gets OAuth access.
The Four Toolsets Available in Beta
Each of these four toolsets maps to a business use case that was previously a custom integration:
1. Core Platform — “Ask the CRM anything”
sObject CRUD, SOQL, SOSL, schema discovery. An internal AI assistant can look up any record, summarize a pipeline, or draft an update note.
2. Data Cloud — “What’s really happening with this customer?”
SQL over Data Cloud entities means unified profile queries, segment insights, and the marketing-data side of the house become conversational.
3. Tableau / Tableau Next — “Explain this chart”
Semantic models, dashboards, and natural-language data analysis. An exec can ask Claude “why did EMEA bookings dip in week 14” and the AI can pull the right dashboard into context.
4. Revenue Lifecycle Management — “Price this deal”
Quoting, ordering, configuration, and the next-generation pricing engine become MCP tools. Sales ops teams get conversational access to the pricing engine without waiting on custom screens.
How Setup Actually Works
The full setup flow at a glance
- Admin enables the Hosted MCP Server in Setup. (The beta is opt-in; not every edition has access yet.)
- Admin registers a connected app with the right OAuth scopes for the toolsets you want exposed.
- End user installs
mcp-remotelocally (this is a small npm package that bridges stdio-based clients like Claude Desktop to the remote endpoint). - End user adds the server URL to their MCP config.
- First run opens a browser for OAuth — the usual Salesforce consent screen appears.
- Tokens are cached locally and the MCP tools show up in the client.
From the end user’s perspective, it’s “click approve, ask a question.” From the admin’s perspective, it’s normal connected-app governance.
The OAuth scopes on the connected app are the real permission boundary. An MCP client can’t do anything the connected app isn’t allowed to do. This is the same mental model you already use for any integration — treat the hosted MCP server exactly like you’d treat a partner integration, because that’s what it is.
Business Use Cases That Actually Pay Off
After a few months watching customers pilot these servers, the patterns that deliver real value fall into four buckets.
Pattern 1 — Sales Ops Research
Account executives and sales operations teams spend a meaningful chunk of their week answering the same questions: “What’s the last touch on this account?” “Has this customer churned product X before?” “What’s the current open pipe for EMEA mid-market?”
Wiring the Core Platform toolset into the team’s AI assistant turns this from a 15-minute CRM deep-dive into a 30-second question.
Pattern 2 — Marketing Segment Discovery
Data Cloud’s SQL toolset lets marketing strategists interrogate segments without waiting for an analyst. “Show me the customers who bought in Q4 2025 but haven’t re-engaged.” “What’s the overlap between newsletter subscribers and trial users who never converted?”
Pattern 3 — Executive Dashboard Narration
The Tableau Next toolset is the one that surprises people. Execs don’t want a new dashboard — they want someone to read the existing dashboard and tell them the story. That someone is now the AI assistant.
Pattern 4 — Quote-to-Cash Acceleration
The Revenue Cloud toolset shortens the tail of deal configuration. A seasoned rep can describe a deal in plain English, get a draft quote, iterate on it, and hand it to CPQ for finalization.
The pattern I’ve seen work: pick one toolset (almost always Core Platform), pick one team (almost always Sales Ops), pick one connected app scope (read-only on Account, Opportunity, and Contact). Ship that in a week. Prove the hours saved. Then expand.
What It Is Not
Hosted MCP is not a replacement for building your own integrations when you need:
- Tools that wrap your own services (your ERP, your data lake, your custom pricing engine)
- Write semantics beyond what the beta exposes
- Deterministic behaviour that the agent can’t route around — in that case, Apex + Flow + explicit APIs stay the right answer
Salesforce has signalled that custom MCP server configurations are on the roadmap for GA. Until then, the beta is excellent for the curated toolsets and worth pairing with your own MCP servers when you need more.
Real-World Scenario
Problem: A sales enablement team turned on the Hosted MCP Server for their whole field organization, using a connected app that had broad read/write scopes because that was the quickest way to unblock everyone.
Three weeks later, an AE asked their AI assistant to “clean up duplicate contacts for Acme.” The AI did exactly that — on production — and deleted the wrong set.
Root cause: The connected app had write scope on Contact.
Fix: Treat the connected app exactly like an integration user. Read-only scopes for everyone by default. Write scopes granted per-user, per-object, per-business-case, with audit. The MCP surface has no more privilege than the connected app — so reducing the connected app’s scope is the move.
What to Watch For
- GA timing — targeted for 2026, with custom MCP server configurations on the roadmap
- Pricing — TBD at time of writing; keep an eye on the platform pricing page
- Expanded toolsets — new toolsets are being added through the beta period
- Partner MCP servers — 30+ partners committed to listing, including AWS, Box, Cisco, Google Cloud, IBM, Notion, PayPal, Stripe, and Teradata
If your team could wire one of these four toolsets into their AI assistant tomorrow, which one would save the most time — and why?
How did this article make you feel?
Comments
Salesforce Tip