Asking Salesforce Questions from Claude

Salesforce announced hosted MCP servers and a new wave of agent-ready tooling at TDX and Dreamforce this year. Here is a practical way to start: a read-only MCP server in a sandbox, with Claude as the client, so people can ask questions about Salesforce data and metadata from a familiar chat interface. This post covers how it works, the guardrails behind it, and where to take it next.

Salesforce has made MCP an official way for AI clients to work with the platform. Hosted MCP servers reached general availability in April 2026 for Enterprise Edition orgs and above, after a pilot and a beta. At TDX 2026, Salesforce introduced its Headless 360 direction and presented sessions on its MCP servers. At Dreamforce 2026 in September, the theme continued with bringing governed Salesforce data and actions into interfaces such as Slack and Claude.

For architects, the message is that the platform path is now official. You do not need to build a connector from scratch, and you can start small and widen access over time.

1. Starting Point

Start with a small scope in a sandbox

Start with a MCP server hosted in Salesforce. MCP lets an AI client discover and call a defined set of tools. Expose a small, read-only set, like: look up records, search, and describe the data model.

I connected Claude to it, so a user can type a plain-language question such as which fields an object has, how two objects relate, or which records match a condition. The answer comes from the live sandbox, not from the model's memory.

We use AI tools in different ways: GitHub Copilot in VS Code assists with code and metadata exploration, debugging, and development tasks, while other clients can be connected or extended through approved plugins, connectors, or MCP servers. For example, Claude can connect to Salesforce's published MCP services through a Salesforce MCP server, making selected Salesforce capabilities available to delegated administrators in natural language—without requiring them to work directly in the codebase.

This complements, rather than replaces, the developer workflow. In VS Code, Copilot can help trace Salesforce behavior across Apex, Flows, Lightning Web Components, metadata, integrations, and tests. Through an MCP connection, an approved client can help administrators investigate or perform specific Salesforce tasks using the services exposed to it. In both cases, AI can speed up exploration, explain errors, and turn findings into clear recommendations.

The key advantage is matching the tool to the user and task: developers work with source code in their IDE; delegated administrators can use approved Salesforce capabilities through a client they can access. MCP does not grant extra Salesforce privileges by itself—the authenticated user's permissions and the MCP server's exposed tools govern access. Before enabling it, we should define which services are available, whether they can make changes or are read-only, what data may be shared with the AI provider, and how usage is logged and reviewed.