Salesforce MCP: What It Is, How to Set It Up, and What It Can't Do

Salesforce MCP is a way to plug an AI assistant straight into your CRM, so a model like Claude can pull up an account, summarize a pipeline, or update an opportunity without you clicking through Salesforce yourself. It uses the Model Context Protocol, the open standard for connecting AI to the tools where your data lives, and as of April 2026, Salesforce ships its own official server for it.

There’s a lot of confusion around the term, though. Some people mean Salesforce’s hosted server. Some mean Agentforce reaching out to other systems. Some mean a custom server they built over a weekend. This guide sorts out what Salesforce MCP actually is, how to set it up, whether it’s secure, and the one thing it still can’t do for you. It’s built from Salesforce’s own documentation rather than a live test, because I couldn’t get access to a Salesforce org (more on that in the setup section).

What Is Salesforce MCP?

Salesforce MCP is the use of the Model Context Protocol to connect AI clients to Salesforce data and actions. In plain terms, it’s a standardized doorway that lets an AI assistant query your CRM, generate reports, and make changes, using the same protocol it would use to reach any other connected tool.

MCP itself isn’t a Salesforce invention. It’s an open standard that Anthropic introduced in November 2024 and later handed to a vendor-neutral foundation, designed to give AI models one consistent way to talk to external systems instead of a tangle of one-off integrations.

Think of it as a universal adapter: build the connection once, and any MCP-compatible client can use it.

Applied to Salesforce, MCP exposes your org’s tools so an AI client can discover and call them on your behalf. Those tools include searching records, running SOQL queries, triggering Flows, and calling Apex actions.

Is Salesforce MCP an Official Product or a General Term?

Both, which is where the confusion starts. Salesforce does offer an official, first-party server, the Hosted MCP Server, generally available since April 2026, so Salesforce MCP is a real product you can turn on, not just a concept.

But the phrase also gets used loosely for a couple of other things: community-built or custom MCP servers that wrap the Salesforce API, and Salesforce’s Agentforce reaching out to external MCP servers.

Claude’s connector directory showing the official Salesforce connector alongside community-built MCP servers like Hubbl, LeadIQ, and AgentContext The official Salesforce connector (verified) sits among community-built MCP servers in Claude’s directory, November 2026.

The distinction that matters most is whether Salesforce is acting as the server, something an AI client connects to, or as the client, an agent connecting out to other tools. Those are opposite directions, and mixing them up is the single most common mistake people make with this topic. More on that next.

Salesforce MCP vs. Agentforce: What’s the Difference?

The difference is direction. With the Hosted MCP Server, Salesforce is the server: outside AI clients like Claude or Cursor connect in to reach your CRM. With Agentforce, Salesforce is the client: its agents reach out to external MCP servers to use tools that live in other systems. Same protocol, opposite flow.

Salesforce as Server Salesforce as Client
What it is Hosted MCP Server Agentforce (native MCP client)
Direction of flow AI clients connect in to Salesforce Salesforce reaches out to other tools
Who connects to whom Claude, ChatGPT, or Cursor call your CRM An Agentforce agent calls Snowflake, Slack, etc.
Answers the question “Let me ask AI about my Salesforce data” “Let my agent act on outside systems”
Status GA (April 2026) Check release notes

Checked November 12, 2026. Sources: Salesforce developer documentation and product announcements.

Here’s why it matters. If your goal is to ask Claude about your Salesforce data, you want the Hosted MCP Server, Salesforce as server. That’s the one that went GA in April 2026. If your goal is to let an Agentforce agent pull inventory from Snowflake or post to Slack, you want Agentforce’s MCP client support, Salesforce as client, consuming outside servers.

The two features also ship separately. The Hosted MCP Server is generally available. Agentforce connects to external MCP servers through the Agentforce Registry in Setup, so check your org’s release notes for its current status.

Salesforce documents the setup for connecting Agentforce agents to external MCP servers if you’re going that route.

Salesforce also runs AgentExchange, its marketplace where partners publish trusted MCP servers that Agentforce agents can install, which is worth knowing if you’d rather adopt a vetted connector than build one yourself.

Agentforce is a big topic in its own right, and tl;dv’s explainer on what Agentforce is covers the agent side in depth. For this guide, just hold onto the server-versus-client distinction. Everything else follows from it.

The Official Salesforce MCP Servers: Hosted vs. DX

Salesforce actually ships two official MCP servers, and they serve different people. The Hosted MCP Server is for connecting AI clients to live org data; the DX MCP Server is for developers working locally. Pick based on whether you’re operating on business data or building on the platform.

The Hosted MCP Server is the one most Salesforce MCP searches are after. It’s a Salesforce-managed endpoint, generally available since April 2026 for Enterprise Edition orgs and above, that exposes your data, SOQL queries, Flows, Apex actions, Data 360, and Prompt Builder to any MCP-compatible client. It authenticates through OAuth 2.0 with PKCE via External Client Apps, runs inside Salesforce’s own security perimeter, and enforces the permissions you already have.

The Salesforce DX MCP Server is a different animal. It runs locally, uses your Salesforce CLI credentials, and ships more than sixty development-focused tools for tasks like deploying metadata, running Apex tests, and building and testing Lightning Web Components. It’s open-source and needs no OAuth setup, though some of its tools are still marked non-GA.

If you’re an admin or a salesperson, you want hosted. If you’re a developer living in Cursor or Claude Code, DX is built for you.

Feature Hosted MCP Server DX MCP Server Agentforce (as MCP Client)
Who it’s for AI users connecting clients to live CRM data Developers building on the platform locally Teams whose agents must reach outside systems
Direction Client connects in Local developer tooling Salesforce reaches out
Status GA (April 2026) GA and non-GA tools Check release notes
Authentication OAuth 2.0 + PKCE Salesforce CLI credentials OAuth 2.0
Where it runs Salesforce infrastructure Locally Within Salesforce
Edition / access Enterprise Edition and above Developer / CLI Agentforce

Checked November 12, 2026. Sources: Salesforce developer documentation and product announcements.

Which Salesforce Editions Support MCP?

The Hosted MCP Server requires Enterprise Edition or above. That’s the short answer, and it’s worth knowing before you plan a rollout: Salesforce makes it available to every Enterprise Edition org and above, so Professional Edition and below aren’t included.

The DX MCP Server sits at the developer tier and depends on the Salesforce CLI rather than an edition gate. Salesforce also lets admins assemble custom hosted servers in Setup, with no code, from existing tools and Apex REST endpoints. Building and hosting your own MCP server on top of the Salesforce APIs is another route, but you take on the hosting and security work in exchange.

How to Set Up Salesforce MCP (Connect Claude, ChatGPT, or Cursor)

Setting up the Hosted MCP Server comes down to four things: an eligible org, an External Client App, the right scopes, and an MCP client to point at it. Salesforce’s own setup documentation walks through the exact click path, but here’s the shape of it so you know what you’re signing up for.

I tried to test this setup myself for this guide but couldn’t get access to a Salesforce org. I’d heard it was notoriously awkward to log in to Salesforce for the first time, but didn’t expect to not be able to get in at all. Leave yourself extra time for that step before you start.

First, confirm your org is Enterprise Edition or above and that you can create an External Client App. In Setup, you’ll enable the standard MCP servers and create the client app that authorizes access.

The scopes to configure are the MCP server access scope plus a refresh-token scope, with PKCE enabled. That’s what lets the connection stay authorized without stashing a long-lived secret in a config file. Once authorized, the client talks to a Salesforce-hosted endpoint, so there’s no local server to babysit.

Then you connect your AI client. Claude, ChatGPT, and Cursor all speak MCP (Salesforce has tested each one), so once the endpoint is live you authorize the client the same way you’d authorize any OAuth app.

The Connect Salesforce dialog with OAuth Client ID and Client Secret fields for linking Claude to a Salesforce org The Connect Salesforce dialog in Claude, November 2026.

If you’re wiring up Claude specifically, tl;dv’s walkthrough on connecting Claude via MCP covers the client side, and our roundup of the best MCP servers is a useful map of what else you can plug in alongside Salesforce.

What Authentication Does Salesforce MCP Use?

The Hosted MCP Server uses OAuth 2.0 with PKCE, authenticating each request against a real Salesforce user’s identity. This is the part that keeps it safe: the AI client doesn’t get a blanket key to your whole org. It acts as the user who authorized it, and Salesforce enforces that user’s object, field, and record-level permissions on every call.

If a rep can’t see a certain opportunity in the interface, the AI acting as that rep can’t see it through MCP either. You can rotate refresh tokens, and revoking a user’s Salesforce session ends their MCP session too.

What Can You Do with Salesforce MCP?

On the read side, Salesforce MCP lets an AI client search accounts, contacts, leads, opportunities, and cases, summarize an account’s history, run SOQL queries, and answer plain-language questions about your pipeline. Instead of building a report, you ask which enterprise deals slipped a stage this month, and the model queries the org and answers.

Real prompts look like this: summarize the Initech account and its last three opportunities, create a follow-up task on the Globex renewal for Friday, or list every open case for this contact.

It’s not just sales, either. Support teams can surface a customer’s open cases and history before a call, and RevOps can interrogate pipeline changes without waiting on a dashboard to refresh.

On the write side, it can create and update records, log activities, set follow-up tasks, and draft updates back into Salesforce, the work that usually eats a rep’s afternoon. This is the same territory as the AI sales agents that now sync notes into the CRM automatically.

Because writes actually change your system of record, the sane pattern is to keep them behind a confirmation step rather than letting a model fire updates unattended.

A few honest caveats. MCP is request-response inside a conversation, so it’s excellent for on-demand questions and one-off actions and poor for anything that needs to run on a schedule or react to an event.

It won’t watch your pipeline. It won’t fire when a deal changes stage. And it only knows what’s already in Salesforce, which, as any sales manager will tell you, is rarely the full story of what happened on the call. We’ll come back to that.

Can Salesforce MCP Read and Write CRM Data?

Yes to both. The Hosted MCP Server supports read and write operations, from querying records to creating and updating them. The important qualifier is governance, not capability. Just because an AI client can write to Salesforce doesn’t mean it should do so freely.

The recommended setup is least privilege: start with the read-only SObject Reads server, limit the client app to specific profiles and permission sets, and require human approval before any write lands. Capability is easy; restraint is the part you have to design.

Is Salesforce MCP Secure?

Salesforce MCP is as secure as the permissions and guardrails you put behind it. The protocol inherits Salesforce’s existing controls rather than bypassing them.

Because the Hosted MCP Server runs inside Salesforce’s security perimeter and acts as an authenticated user, object, field, and record-level security all still apply. That’s a strong foundation, and it’s why the hosted route is safer than a custom server holding credentials in a file somewhere.

The risk that’s genuinely new is prompt injection. An AI client reading your CRM is also reading whatever text lives in it, notes, email bodies, case descriptions, and any of that could contain instructions crafted to hijack the model. OWASP lists prompt injection first in its Top 10 for LLM applications, because models can’t reliably tell trusted instructions from untrusted content.

In a CRM context, that means you should treat record text as untrusted, keep write tools separate from read tools, and put a human in the loop for anything destructive.

The practical playbook is short: least-privilege access, client apps limited to the right profiles and permission sets, read-only by default, confirmation on writes, and an audit trail of what the AI queried and changed.

What Are the Risks of Giving AI Write Access to Salesforce?

The main risks are wrong-record updates, over-permissioned servers, and the confused-deputy problem, where an AI with broad access is tricked into acting beyond what the user intended.

A model can misread a request and update the wrong opportunity. A server scoped too generously can reach data the task never needed. And an injected instruction can turn a helpful assistant into an unwitting accomplice.

None of these are reasons to avoid MCP; they’re reasons to scope it tightly. Grant the narrowest permissions that get the job done, log everything, and never let write access run without a confirmation gate. Excessive agency is the failure mode, so design against it from the start.

Where Salesforce MCP Stops (and Meeting Intelligence Picks Up)

Here’s the honest limit: Salesforce MCP can only work with what’s already in Salesforce, and it only works when you’re actively asking. It’s a doorway, not a worker. Nothing watches your accounts, nothing fires when an opportunity moves, and nothing runs while the chat is closed.

For a lot of CRM questions and answers, that’s completely fine. For keeping the CRM accurate in the first place, it’s a gap.

That gap matters because the most valuable detail about a deal usually isn’t in a picklist. It’s in what the customer said on the call. Salesforce’s own State of Sales research finds the average seller spends just 40% of their time selling. Roughly 60% of the week goes to non-selling work, much of it manual data entry.

An AI you have to prompt after every call doesn’t fix that; it just moves the typing into a chat window. That’s the gap agentic AI for sales is trying to close.

This is the part tl;dv handles automatically. tl;dv records, transcribes, and summarizes calls across Zoom, Google Meet, and Microsoft Teams, plus any other call through its desktop app, then pushes the insights straight into your CRM through its Salesforce integration (Pro plan and above), with no one opening a chat and no one typing up notes. The conversation becomes CRM data on its own.

Salesforce isn’t the only vendor shipping an MCP server, either. Zoom MCP takes the same approach for meetings, and the same reactive limitation applies. Google’s own Workspace MCP server is still in developer preview and doesn’t cover Meet yet, so Google Meet MCP options are community-built for now.

And if you do want to query that meeting data from an AI client, tl;dv ships its own MCP server, on the Pro plan and above, so Claude or Cursor can pull transcripts, highlights, and action items the same way Salesforce’s server exposes records.

Point both at the same assistant and you get the full picture: CRM context from Salesforce, and the conversation that actually explains it from tl;dv. Start with tl;dv free and see what your calls capture that the picklists miss.

The Bottom Line

Salesforce MCP is real, official, and genuinely useful. The Hosted MCP Server lets AI clients read and write your CRM through an open standard, with your existing permissions intact.

Just keep the two things straight, Salesforce as server with Hosted MCP versus Salesforce as client with Agentforce reaching out, scope it tightly, and remember it’s reactive by design. It answers when asked; it doesn’t run on its own.

For the automatic layer that captures what happens on your calls and keeps Salesforce current without anyone lifting a finger, that’s where meeting intelligence like tl;dv fits alongside it: one for the conversation, one for the record.

Поширені запитання

Is Salesforce MCP free?

There's no separate add-on fee on top of your Salesforce subscription: the Hosted MCP Server is available across every Enterprise Edition org and above. You're authorizing an AI client against your existing org and permissions, not buying a separate connector. AI clients and any custom hosting are separate costs.

Does Salesforce MCP require Agentforce or Data Cloud?

Salesforce doesn't list either as a requirement. The Hosted MCP Server is available across Enterprise Edition orgs and above, and any compatible AI client, such as Claude, ChatGPT, or Cursor, can connect to it. Agentforce is Salesforce's agent platform and uses MCP in the opposite direction, as a client consuming external servers.

Can Salesforce MCP access custom objects?

Yes. The hosted sObject tools work with standard and custom objects, and a custom-built server can expose organization-specific fields and logic. Access is still bounded by the authenticating user's permissions.

Can Claude or ChatGPT connect to Salesforce via MCP?

Yes. Any MCP-compatible client, including Claude, ChatGPT, and Cursor, which Salesforce has tested, can connect to the Hosted MCP Server once it's enabled and you've authorized the client through OAuth.

Can Salesforce MCP trigger workflows or automations?

Not on its own. MCP is request-response inside an AI chat: it has no triggers, and nothing runs when the conversation is closed. For event- or schedule-driven work, you need Salesforce automation like Flows and Apex, or a dedicated workflow tool.

Salesforce MCP vs. the Salesforce API, which should you use?

Use MCP when you want an AI client to discover and call approved tools at runtime. Use the REST, Bulk, or Apex APIs for fixed, repeatable system-to-system integrations. MCP is for agent-driven work; traditional APIs are for pipelines.

Matt Leppington

Matt Leppington — англійський фриланс-автор, який пише про віддалену роботу, продукт, UX, продажі, подорожі й технології.

TL;DV ДЛЯ ПРОДАЖІВ

Зустрічайте своїх AI-агентів з продажу

Більше про агентів з продажу зі штучним інтелектом

Наш штучний інтелект відстежує прийняття playbooks і вимірює, наскільки ефективно обробляються заперечення, надаючи менеджерам інформацію для тренінгів, щоб підвищити ефективність роботи з клієнтами. Він підбиває підсумки незліченних зустрічей і виділяє ключові дискусії. Це чудово підходить і для представників - автоматизує підготовку до зустрічей, оновлення CRM та подальші електронні листи

Агент оновлення CRM

Легко записуйте, транскрибуйте, обробляйте та діліться інсайтами в конкретних полях вашої CRM (Hubspot, Salesforce, Pipedrive тощо!).

Спеціаліст з аналізу телефонних дзвінків Агент

Аналізуйте дзвінки, щоб виявити тенденції, заперечення або успішні техніки.

Відстежувач пунктів дій

Витягуйте та відстежуйте поставлені завдання та дедлайни на нарадах, щоб забезпечити підзвітність.

Подальші дії з агентом електронної пошти

Автоматично записуйте та розробляйте подальші дії на основі обговорень на зустрічі.

Агент конкурентної розвідки

Використовуйте інформацію про конкурентів під час зустрічей, щоб виявити тенденції та стратегії.

Агент з навчання та підвищення кваліфікації

Записуйте навчальні сесії та створюйте резюме або основні моменти для подальшого використання.

Запізнюєтеся на наступну зустріч?

Надішліть замість себе tl;dv!

  • Приєднується до вашої зустрічі, навіть якщо ви не можете!
  • Записує та транскрибує понад 30 мовами.
  • Робить нотатки за зручною вам структурою.
  • Створює підсумки ваших зустрічей.
  • Автоматично оновлює ваші улюблені інструменти (CRM, Slack, Notion тощо).
  • Покращує результати зустрічей за допомогою оцінювальних карток ефективності.
Українська