Search “Salesforce MCP server” and you will land on guides, forum threads, and vendor comparisons that use “Salesforce MCP” to mean two different products. Salesforce ships two official MCP servers under that umbrella: Hosted MCP Servers and the DX MCP Server. They share a vendor and a protocol. They do not share an audience, a setup process, or a job.
If you are trying to figure out which one applies to you, or you are getting conflicting answers from different docs pages, here is the direct comparison: what each one is, who it is actually built for, and where the confusion usually comes from.
The Short Answer
Hosted MCP Server is Salesforce’s cloud-based, OAuth-connected server. Any Salesforce user can connect an AI client to it and work with records, run existing flows, and pull report-style analytics, all within their own permissions. DX MCP Server is a local, CLI-based server built for developers already working in Salesforce DX, covering metadata deployment, testing, and code-level tasks.
Neither one is built to sit in a normal conversation and create a new custom field, build a flow from scratch, or restructure a permission set as a live, no-deploy change. That gap matters, and it is covered near the end of this piece.
Salesforce Hosted MCP Server, Explained
Hosted MCP Servers run entirely on Salesforce’s own infrastructure. There is nothing to install locally. You connect an AI client to it by creating an External Client App in Setup and authenticating with OAuth 2.0 plus PKCE, then any MCP-compatible client, Claude, ChatGPT, or similar tools, can call it.
It reached General Availability in April 2026, and it requires Salesforce Enterprise Edition and up. The endpoint catalog is broad:
- sObject Reads, Mutations, Deletes, and All, covering record-level create, read, update, and delete
- Invocable Actions, exposing Apex @InvocableMethod actions as MCP tools
- Flows, for running flows that already exist in your org
- Data 360, for unified data queries
- Prompt Builder, exposing existing prompt templates
- Tableau Next, for semantic model discovery and analytics
- API Catalog and Custom Servers, for teams building their own tool wrappers on top of Salesforce APIs
That is a genuinely capable set of tools for operational work. Any Salesforce user, not just admins, can ask an AI client to update records, trigger a flow, or explain what a report shows, all scoped to whatever that user could already see and do in Salesforce.
What it will not do is create or modify metadata. It cannot build you a new object, add a field, write a validation rule, or construct a flow from nothing. It works with what is already configured in your org. Configuring the org is a different job.
Salesforce DX MCP Server, Explained
The DX MCP Server is the opposite shape. It runs locally, using your existing Salesforce CLI authentication, and it is aimed squarely at developers already working inside a Salesforce DX project. There is no OAuth flow to configure. You install it as an npm package, point it at an authorized org alias, and it exposes more than 60 tools across toolsets including metadata, data, testing, users, and Lightning Web Component development.
Because it is open source and works with the CLI directly, DX MCP Server runs against any org edition, including free Developer Edition orgs. It is genuinely free to use.
Metadata deployment is where DX MCP Server pulls ahead of Hosted MCP, on paper. It can retrieve metadata as local source files, let an AI assistant edit those files, and deploy them back to an org. That is a real capability. It is also a deploy-pipeline model, not a live, in-conversation change. The work happens in a local project folder, gets tested, and then gets pushed, the same shape as a developer’s existing workflow, just with an AI assistant driving parts of it.
Side by Side
| Aspect | Hosted MCP Server | DX MCP Server |
|---|---|---|
| Runs where | Salesforce’s own infrastructure, cloud | Locally, via Salesforce CLI |
| Built for | Any Salesforce user, operational tasks | Developers already using sf CLI |
| Setup | External Client App, OAuth 2.0 + PKCE | npm install, authorize an org alias |
| Editions | Enterprise Edition and up | Any edition, including free Developer Edition |
| Can create or modify metadata live, in conversation | No | Through a local deploy pipeline, not live |
| Best for | Records, existing flows, report-style analytics | Metadata deployment, Apex testing, LWC work |
Why People Mix These Two Up
A few reasons this keeps happening. Both are Salesforce’s own products, both use “MCP” in the name, and Salesforce’s own documentation for each links out to the other, which reads as more overlap than actually exists. The beta-to-GA transition added another layer: Hosted MCP’s URL pattern changed when it went GA in April 2026, and older guides referencing the retired beta path get cited alongside current DX MCP docs as if they describe the same thing.
The most common specific mix-up is the --orgs flag. It is a DX MCP Server flag, used locally to specify which authorized CLI orgs a session can touch, including a value that allows all of them. It gets described in some secondhand write-ups as a Hosted MCP multi-org feature. It is not. Hosted MCP’s scope is set through OAuth permissions on a single connected org, not a CLI flag.
What Neither One Does
Both of these are real, useful, Salesforce-built products, and this is not a case against either one. But if what you actually want is to describe a change in plain English, in a normal chat, and have an admin-level configuration change happen right there, a new field, a rebuilt page layout, a new validation rule, governed by your existing permissions and without a local project or deploy step, neither Hosted MCP nor DX MCP is built for that. That is a third category: purpose-built Salesforce admin MCP servers, like Cirra AI’s, that sit alongside Salesforce’s own two and are scoped specifically to metadata configuration work.
Which One Do You Actually Need?
A quick way to sort it:
- You want to query or update records, or run a flow that already exists, without installing anything: Hosted MCP Server.
- You are already working in a Salesforce DX project with the CLI installed, and you want an AI assistant helping with metadata deployment, Apex tests, or LWC work: DX MCP Server.
- You want an AI assistant to configure your org, creating objects, fields, flows, or permission sets, conversationally and without a deploy pipeline: a purpose-built admin MCP server.
FAQ
Can I use Hosted MCP and DX MCP Server at the same time?
Yes. They are separate servers with separate scopes, and there is no conflict in having both connected, one for operational record and flow work, one for development tasks.
Does Hosted MCP Server require Enterprise Edition?
Yes. Hosted MCP Servers require Salesforce Enterprise Edition and up. DX MCP Server has no such requirement and works with any org edition, including Developer Edition.
Can Hosted MCP Server create custom fields or objects?
No. It works with records, existing flows, and existing reports and analytics surfaces. It does not create or modify metadata.
Is DX MCP Server free?
Yes. It is open source and free to use, though it operates against Salesforce orgs, which still require whatever Salesforce licenses your org already needs.
Where does Agentforce fit into this?
Agentforce is Salesforce’s agent platform, and it is a different thing again: it can act as an MCP client itself, meaning Agentforce agents can call MCP servers, including third-party ones, rather than being one. It is not a substitute for either server covered here.
The Takeaway
Hosted MCP and DX MCP Server solve different problems for different people, and knowing which is which saves you from configuring the wrong one. If your actual goal is conversational org configuration rather than record operations or code deployment, that is a distinct category worth understanding on its own terms.
Learn more about the full landscape of Salesforce MCP options, including where Cirra AI’s admin-focused server fits, on the Salesforce MCP Server pillar page.
