I'll just start with what happened.
I was debugging a data issue on a client's org. Normally that means: open the org in a browser, navigate to the record, check field values, open Developer Console, write SOQL, run it, copy results, paste into my notes, think about it, write another query, repeat. Probably 15 minutes of tab-switching before I even start solving the problem.
With MCP connected, I described the issue to Claude, it queried the records directly, pulled the field values, checked the related records, and showed me the data flow. Same investigation, maybe 3 minutes. That's when MCP went from "interesting tool" to "I'm not going back."
What MCP Is
MCP (Model Context Protocol) is an open protocol, originally from Anthropic, now under the Agentic AI Foundation (part of the Linux Foundation). It defines a standardized way for AI tools to connect to external systems. An MCP server exposes tools (functions the AI can call) and resources (data it can read). An MCP client (Claude Desktop, Claude Code, Cursor) connects to those servers and gives the AI access to their capabilities.
For Salesforce, an MCP server translates Salesforce API operations into tools Claude can call. When I ask "what fields are on the Financial_Account__c object?" Claude doesn't guess. It calls the Describe endpoint on the actual org and gives me the live answer.
MCP is not a Salesforce product (though Salesforce has announced native MCP client support in Agentforce, currently in pilot). It's a protocol layer that makes existing APIs accessible to AI tools.
By early 2026, MCP had been adopted by OpenAI, Google, Microsoft, and Salesforce, and donated to the Agentic AI Foundation under the Linux Foundation. That kind of convergence on an infrastructure standard almost never happens. Investing in MCP-based integrations is a safe bet.
Setup
The Salesforce CLI team maintains an official MCP server at salesforcecli/mcp on GitHub. That's the one I use.
Install the server. It runs as a Node.js process. Clone the repo, install deps, configure the connection.
Authenticate. It uses your existing Salesforce CLI auth. If you've already run sf org login web for your target org, the MCP server picks up that auth context. No separate credential setup.
Point your MCP client to it. In Claude Code or Claude Desktop, add the server in your MCP config. Once connected, Claude sees the available tools and can invoke them during your conversation.
Whole thing takes about 15 minutes if you're already familiar with the Salesforce CLI.
How I Actually Use It
Schema exploration without opening Setup
I work on a lot of client orgs. I don't have every object and field memorized. Instead of navigating to Object Manager and clicking through tabs, I just ask Claude what's on the object. It calls the Describe API and returns the field list with types, labels, and relationships.
This comes up constantly with managed packages or FSC orgs where there are dozens of packaged objects you didn't build. "What custom fields are on FinServ__FinancialAccount__c?" Real answer. From the real org. In seconds.
SOQL that writes and runs itself
I describe what data I need in plain English. Claude writes the SOQL, runs it, and returns the results. If something's off (wrong filter, missing relationship), I describe the fix and Claude adjusts and re-runs.
This kills the loop of: write SOQL in Developer Console, copy the results, paste into chat, ask for analysis, get suggestions, go back to Dev Console. The AI sees the data directly. One conversation, not four tools.
Debugging with the actual records
I mentioned this in the intro. When I'm troubleshooting, Claude can query the specific records involved, inspect field values, check related objects, and trace the data. Instead of me describing the problem in abstract terms ("there's an Account with a null field that should have been populated by the trigger"), the AI looks at the data and can tell me exactly what happened.
For issues that depend on specific data states (a trigger that fires correctly for 99% of records but fails on one edge case), this saves a ton of time.
Understanding what's deployed
"Show me all the triggers on the Opportunity object." "What Apex classes reference the Meeting_Note__c.Status__c field?" "How many active flows are there on Account?"
Claude queries the Tooling API and compiles the info. For orgs with years of accumulated automation, this kind of quick audit is useful before making changes.
Code generation that knows your schema
When Claude generates Apex or LWC, it can reference your actual schema. Real field API names, real relationship names, real picklist values. Not placeholders you have to manually correct.
I used to spend probably 10% of my time fixing generated code where field names didn't match the actual org. That's just gone now.
The Workflow Shift
Before MCP, my setup for client work was: org open in Chrome, VS Code, Developer Console or Workbench, and constant copy-paste between all of them. Context lived in my head.
Now the AI has direct access to the org. I describe what I'm building, it pulls the context it needs, generates code against real metadata, and I deploy. The context-switching drops dramatically.
Is it a 10x improvement? No. But across dozens of interactions per day, the time compounds. And the quality improvement (code that references the correct field names the first time) compounds even more.
Where This Is Heading
Salesforce has MCP client support coming to Agentforce (in pilot now). The same MCP servers that connect Claude to my dev workflow could also connect Agentforce agents to external systems. An MCP server for your data warehouse or internal APIs becomes a universal integration layer that works with your dev tools and your production agents.
That's the real play. The protocol is converging across the industry. Building MCP servers now gives you connectors that work with basically every AI tool and platform on the market.
If you want help setting up MCP for your Salesforce dev workflow or building MCP-based integrations, let's talk.