On this page
The short version
Salesforce MCP is the set of hosted Model Context Protocol servers Salesforce runs so AI clients such as Claude and ChatGPT can query, search and update CRM records under each user's own permissions. RevSure, the agentic GTM platform whose Digital Workers run on the Full Funnel Data Graph, works next to it: its workers read Salesforce data joined with marketing, ad and intent data, and write results back to Salesforce.
In late October, the Salesforce administrator at a market data spend management company, who also maintains the company's field mappings in our platform, used a routine sync with our team to share some news. The company's legal counsel had finally approved plans to "connect Salesforce to Claude," along with its call recording tool, after a stretch in which there had been a lot of talk about "not being able to connect those things because of like customer data and stuff." His next thought was the connection he expected to add after that: "probably I would imagine be looking at like the RevSure MCP and stuff like that."
The approval was the slow part. Salesforce's documentation puts the waits at up to 30 minutes for a new app registration and up to 2 minutes for a server to switch on. The decisions that take longer come after the connection works: which servers a team switches on, and which questions belong to the CRM's own server rather than to a layer that also holds the data the CRM never sees.
What is Salesforce MCP?
Salesforce MCP most often means Salesforce Hosted MCP Servers, which Salesforce describes as giving "AI agents a secure, governed way to interact with Salesforce data and automation." Each server offers a set of tools over the Model Context Protocol, the open standard our guide to MCP explains, and an AI client such as Claude, ChatGPT, Cursor or a custom agent connects to it with per-user OAuth. Salesforce made the hosted servers generally available on April 29, 2026, for "every Enterprise Edition org and above."
Standard server. A server Salesforce builds and maintains. An administrator switches it on, and its list of tools is fixed.
Custom server. A server an org assembles from its own logic. Salesforce's overview lists Apex invocable actions, Apex REST methods, Flows and Named Queries among the things it can expose as tools "without writing connector code."
Tool. One operation a server offers, such as running a query or fetching related records. The AI client's model decides which tool to call and with what arguments.
People searching for Salesforce MCP also find community-built and third-party servers that wrap Salesforce's APIs. This guide follows Salesforce's own hosted servers.
Which Salesforce MCP servers can a GTM team use?
Salesforce's standard servers reference splits record access by what a server may do, so read, write and delete rights come from the choice of server as well as from each person's profile.
The SObject Reads server is the one most revenue teams start with, and its six tools show how an assistant works through a question. SObject Reads offers getObjectSchema, which "Returns Salesforce schema information optimized for LLM consumption"; soqlQuery, which runs a SOQL query; find, a text search across several objects at once; getUserInfo; listRecentSobjectRecords; and getRelatedRecords, which follows relationships from a parent record to its children. The guidance on soqlQuery is to "Always include a WHERE clause and a LIMIT clause."
That design means the model reads the org's schema and writes its own SOQL for each question. It is a capable way to reach records, and it puts every choice of field, filter and date range into the query the model writes on that run.
How do you connect Salesforce MCP to Claude?
The setup needs a System Administrator, or someone with equivalent permissions, on the Salesforce side and a person who can add connectors on the Claude side. Salesforce documents it in five steps.
- Register Claude as an External Client App. In Setup, search for External Client App Manager and create an app with the callback URL https://claude.ai/api/mcp/auth_callback, the scopes "Access MCP servers (mcp_api)" and "Perform requests at any time (refresh_token)", and JWT-based access tokens for named users turned on. Salesforce says the app "can take up to 30 minutes to become available."
- Switch on the servers. In Setup, search for MCP Servers under API Catalog and enable the ones the team needs. A server can take up to 2 minutes to become active, and "If you don't enable your MCP servers, clients will not be able to connect, even with correct authentication."
- Add a custom connector in Claude. Under Customize, then Connectors, choose Add custom connector, give it a name and enter the server URL. For SObject Reads in production that is https://api.salesforce.com/platform/mcp/v1/platform/sobject-reads.
- Paste the consumer key. In the connector's Advanced settings, paste the consumer key from the External Client App into OAuth Client ID, then add the connector.
- Connect and sign in. Each person clicks Connect and signs in to Salesforce as themselves, so the assistant sees only what that person could already see.
Our guide to Claude connectors covers the Claude side for any vendor, including who on a Team or Enterprise plan has to enable connectors first.
What happens when a revenue team starts asking questions?
The first questions are usually ones the CRM answers well, such as open opportunities over a value threshold or the deals a rep owns in a given stage. The interesting moment comes when the answer is checked against something people already trust.
At a test automation software company, the demand generation manager described her main use of RevSure as "connecting the MCP, and then being able to get like a quarter to date view, and seeing like where the numbers are moving. And then I'm just getting that validated by some of our Salesforce reports." Moments later she summed up the arrangement as "using sales force as like the the source of truth." The habit is sound. Salesforce holds the record of what was created and booked, and an answer from any assistant earns its place in a meeting when it ties back to the CRM report the team already signs off on.
At an AI IT service management company, a VP of revenue operations about six weeks into the job told our team that he had been querying our platform through Claude, and "that's been helpful to get behind, uh, some of the more complex analysis I wanted to do." Complex analysis is where a CRM-only connection asks the most of the model, because a horizontal MCP server makes the LLM reconstruct what records mean, how the GTM motion works, and what the data says. Through SObject Reads, a question about pipeline by segment becomes a SOQL query the model writes from the schema it just read, and a colleague asking the same thing an hour later gets a query of their own. Our comparison of MCP and APIs explains why that choice moves from reviewed code into the model on every run.
Some questions also need data that never lands on a Salesforce record, such as anonymous website visits or the ad touches before an opportunity opened. Our MCP server is built for those. Its tools understand GTM entities, stages, handoffs and metrics, then return computed intelligence instead of raw records. Lifecycle, funnel, handoffs, buying groups and GTM operating logic are encoded in the context layer. The two servers do different jobs, which is why the administrator in the opening story expected to add our connection alongside the CRM.
What does a digital worker do with Salesforce data?
A digital worker is a managed AI role that owns a repeatable GTM job end to end. It differs from an assistant a person prompts because someone is accountable for its output. Your team sets the strategy and approves exceptions, and our GTM engineers manage the digital worker day to day and are accountable for the completion and quality of the work it delivers. An embedded GTM pod is the team RevSure assigns to run a customer's digital workers. The pod tunes the workstream and reviews the quality of what it delivers. We cover how a digital worker differs from an agent in a separate post.
Salesforce sits at both ends of a digital worker's job. On the way in, the worker reads Salesforce records through the context layer, already resolved to accounts and joined with marketing automation, ad and intent data. The Digital RevOps Analyst, for example, works through the MCP tools Get Funnel Health, Get Pipeline & Forecast, Get Account/Territory Context and Get Cohort Metrics, plus CRM and BI writeback actions. On the way out, writebacks push resolved identities, scores, owners and fields back into Salesforce and HubSpot, on the records your team already works, and every write is reversible, with an audit trail. Reps and any assistant connected to Salesforce MCP then read those values on the record itself.
The split of work between the worker and the people around it is written down job by job.
The administrator in the opening story was building exactly this kind of standing rule in Salesforce that week: at-risk notifications for deals without the right contact roles attached, with no activity for 45 days, or still in an early stage with a close date 30 days out. Rules like those are the material of the Track Pipeline and Detect & Resolve Anomalies jobs, with a person deciding what happens to each flagged deal. Every recommendation still requires a human to approve, adjust, or override it before it is acted on.
How should admins govern Salesforce MCP for GTM?
Salesforce's servers reference says "All servers enforce per-user authentication and respect Salesforce's standard security model," and that "field-level security, object permissions, and sharing rules apply to every tool call." That inheritance is the foundation, and a few choices on top of it decide how much an assistant can change.
- Start with SObject Reads. It cannot change records, so the first weeks of use carry no write risk while the team learns what it asks.
- Treat writes as a separate decision. SObject Mutations creates and updates without deleting, and SObject Deletes stands apart. Switch them on for named use cases, not by default.
- Keep approvals in the client. Claude asks a person to approve tool calls, and Anthropic's help center advises choosing Allow always only "when using a server and tool that you trust to run unsupervised." Write tools are the ones to leave on approval.
- Use custom servers for org-specific logic. Standard servers' tool sets cannot be modified. A custom server can expose a Flow or Named Query the RevOps team already maintains, so an assistant calls the team's logic instead of rebuilding it.
Our guide to MCP security for enterprise AI agents covers the controls security and legal teams ask about before any MCP connection goes live, the same review the market data company's counsel ran. The four MCP tool primitives post explains why governed actions are a different kind of tool from data reads, and MCP for GTM covers what the protocol leaves to each server.
What the approval did not cover
Legal's yes at the market data company covered a connection: whose data could reach Claude, and under what permissions. Salesforce MCP will now let every person in that company ask the CRM questions in their own words, and each person's assistant will write its own query. The admin who spent the same week writing at-risk rules into Salesforce has already started on the part no approval covers, deciding which questions get one shared definition and who answers for it when the number is wrong.
FAQs
What is Salesforce MCP?
Salesforce MCP usually refers to Salesforce Hosted MCP Servers: servers Salesforce runs so AI clients such as Claude, ChatGPT and Cursor can work with Salesforce data over the Model Context Protocol. Standard servers cover reading, creating and updating, and deleting records, plus Data 360, Tableau Next and beta servers. Each person signs in with OAuth and keeps their own Salesforce permissions.
Which Salesforce orgs can use hosted MCP servers?
Salesforce announced general availability on April 29, 2026, for every Enterprise Edition org and above. A System Administrator, or someone with equivalent permissions, registers the AI client as an External Client App and switches on the servers under API Catalog in Setup. Some servers, including Headless 360, Metadata Grounding and CRM Analytics, are labeled beta.
How do you connect Salesforce MCP to Claude?
An administrator creates an External Client App with Claude's callback URL and the mcp_api and refresh_token scopes, then enables the servers in Setup. In Claude, under Customize and Connectors, a person adds a custom connector with the server URL and pastes the app's consumer key as the OAuth Client ID. Each user then connects and signs in.
Can Salesforce MCP update records?
Yes, if the right server is switched on. SObject Reads is read only, SObject Mutations creates and updates records without deleting them, SObject Deletes handles deletes alone, and SObject All covers every operation. Field-level security, object permissions and sharing rules apply to every tool call, and Claude asks a person to approve tool calls unless they choose to always allow a tool.
Does Salesforce MCP replace a GTM context layer?
No. Salesforce MCP reaches what is stored in Salesforce, and the model works out what the records mean on each question. A context layer such as ours joins CRM records with marketing automation, ad, website and intent data, encodes funnel and handoff definitions once, and returns computed answers. Teams can connect both and check answers against Salesforce reports.
How do RevSure's Digital Workers use Salesforce?
Our Digital Workers read Salesforce data through the context layer, joined with the rest of the GTM stack, and write resolved identities, scores, owners and fields back to Salesforce records, with every write reversible and audited. Our GTM engineers manage each worker day to day, and the customer's team sets strategy and approves exceptions.