How to Connect Claude to Salesforce

TL;DR

Claude connects to Salesforce through any Salesforce MCP server using the Connectors feature, in Claude Desktop, Claude Web, or Claude Code.
You’ll need Claude Desktop, Claude Web, or Claude Code, a Salesforce MCP server and its connector URL, and Salesforce System Administrator access (or a user with the permissions you want Claude to have).
These steps work with any Salesforce MCP server. Cirra AI is used for the screenshots below and covered in detail at the end.

What You Need

What you need

Claude Desktop. Claude Code and Claude Web (claude.ai) also support MCP connections, using the same underlying setup.
A Salesforce MCP server and the connector URL it gives you.
Salesforce System Administrator access, or a user with the specific permissions you want Claude to operate under.

Connect Claude to a Salesforce MCP Server

Claude Desktop

Open Claude Desktop Settings (macOS: Claude → Settings, Windows: File → Settings).
Navigate to the Connectors section.
Click “Add custom connector.”
Enter a Name for the connector and the Server URL your Salesforce MCP server gives you.
Click Add, then Connect, and complete the authorization flow for your Salesforce org.
Start a new conversation and describe a Salesforce task to confirm the connection is live.

Claude Web (claude.ai)

Open the “Connectors” settings. Navigate to claude.ai/settings/connectors.
Click “Add custom connector.”
Enter a Name for the connector and the Server URL your Salesforce MCP server gives you, then click Add.
Click Connect and complete the authorization flow for your Salesforce org.
An authorization request opens in a new browser tab (it may be hidden behind other windows). Review the requested permissions and click “Allow Access” to authorize the connection.
Start a new conversation and describe a Salesforce task to confirm the connection is live.

Verify the Connection

Ask Claude something specific to your org, such as “What custom objects exist in my Salesforce org?” If it returns your org’s actual custom objects, the connection is working. If it gives a generic answer about Salesforce object types in general, the connection did not initialise. Check that the connector shows as Connected and try again in a new conversation.

What Claude Can Do Once Connected

This depends entirely on which Salesforce MCP server you connect, not on Claude itself. An operational-only server lets Claude read and update records, run SOQL queries, and invoke automations you’ve already built. A server with metadata write access lets Claude also create objects and fields, build Flows, and manage permission sets, directly from conversation.

Troubleshooting

Connector shows as disconnected. Re-check the Server URL for typos, then reconnect from Settings → Connectors.

Claude gives generic answers instead of org-specific ones. This usually means the connection didn’t initialise for that conversation. Confirm the connector is marked Connected and start a fresh conversation.

Authorization fails during setup. Confirm you’re authorizing with a Salesforce user that has the permissions the connector needs, and that your org allows the OAuth flow the server uses.

Frequently Asked Questions

Does Claude support MCP natively?

Yes. Claude Desktop, Claude Web (claude.ai), and Claude Code all support MCP connections natively, through the Connectors feature. No third-party plugin or browser extension is required.

Is this the same as Agentforce?

No. Agentforce is Salesforce’s own AI agent platform, built natively into Salesforce. Connecting Claude via a Salesforce MCP server is an external connection: Claude reaches into your org through the server and operates within your existing permissions. The two are complementary, not competing.

Can I connect Claude to more than one Salesforce org?

It depends on the Salesforce MCP server. Cirra AI supports connecting multiple Salesforce orgs, with a multi-team hierarchy that lets consulting partners and BizOps teams manage them from a single place rather than one connection per org.

Is my Salesforce data used to train Claude?

No. Cirra AI does not use customer data to train, fine-tune, or improve any foundation or shared machine-learning model, and contractually prohibits its subprocessors from doing so. Data flows through the MCP connection to answer your request, it is not retained for model training.

Using Cirra AI as Your Salesforce MCP Server

Cirra AI is the Salesforce MCP server used in the screenshots above. Unlike Salesforce’s own hosted MCP server, the Cirra AI Salesforce Admin MCP Server writes to the metadata layer directly in conversation: creating objects and fields, building Flows, and managing permission sets and page layouts.

Governance is built into every connection: granular, category-level access controls (Metadata, User Data, Personal Setup Data, Business Data, each set to None, Read Only, or Read and Write), with actions logged in an exportable audit trail. A multi-team hierarchy also lets consulting partners and BizOps teams manage more than one connected Salesforce org from a single place, rather than one connection per org.