How to Integrate Claude AI with Salesforce Using MCP Server
A complete guide covering Headless 360, MCP Server architecture, External Client App OAuth setup, and a step-by-step walkthrough to connect Claude directly to your Salesforce org.
Table of Contents
- Introduction: The End of the Browser-First Era
- What is Salesforce Headless 360?
- What is Model Context Protocol (MCP)?
- MCP Servers in Salesforce: DX vs Hosted
- Hosted MCP Endpoints: What's Available
- Step-by-Step: Connecting Claude to Salesforce
- Real-World Use Cases
- Tips, Best Practices & Troubleshooting
The End of the Browser-First Era
For over two decades, using Salesforce meant one thing: open a browser, log in, navigate through tabs, click through screens. Every piece of data, every workflow trigger, every automation lived behind a UI. That model worked when humans were the primary consumers of the platform — but we've entered a different era.
AI agents don't log into browsers. They call APIs, invoke tools, and query data through standardized protocols. And Salesforce has responded to this shift with one of the most architecturally significant announcements in its 27-year history: Salesforce Headless 360, unveiled at TrailblazerDX 2026 on April 15, 2026.
In this post — and in my YouTube video above — I walk through exactly how to connect Claude AI to your Salesforce org using the Model Context Protocol (MCP). From understanding the architecture to completing the full OAuth setup, this is everything you need to get started.
What is Salesforce Headless 360?
Salesforce Headless 360 is a platform-wide architectural shift that makes every Salesforce capability — data, business logic, workflows, automations — accessible through APIs, MCP tools, and CLI commands, with no browser or UI required.
Before Headless 360, connecting an external agent to Salesforce meant writing your own connected app, building REST adapters, and creating separate integrations for every channel. The platform was designed for humans clicking through tabs. Headless 360 fundamentally reframes the entire stack: Salesforce is now designed to be consumed by code and agents, not just browsers.
Announced at TDX 2026: Salesforce Headless 360 has been called the most architecturally significant thing Salesforce has shipped for developers in years. Every Developer Edition org ships with Salesforce Hosted MCP Servers at no cost — you can start experimenting with AI agent integration immediately.
The Three Access Patterns
Headless 360 exposes the entire platform through three distinct access patterns. Understanding the difference matters, especially when planning implementations.
APIs
Every Salesforce capability — reading records, updating cases, triggering approvals, querying data — is now a direct API call. No human in the loop. Salesforce has had REST APIs for years, but Headless 360 brings breadth, consistency, and agent-readable documentation to the entire platform.
MCP Tools
The biggest net-new addition. An open standard from Anthropic that lets AI assistants connect to external tools without custom integration code for every new agent. Build one MCP server; any MCP-compatible agent can use it. Salesforce ships with 60+ pre-built MCP tools.
CLI Commands
The Salesforce CLI now supports agentic workflows. The sf agent topic lets you generate agent specs from plain-language descriptions, deploy them, and run evaluation tests — all from the terminal. Ideal for CI/CD pipelines and dev automation.
Key Components of Headless 360
- Data 360 — SQL-based access to your Salesforce data through the hosted MCP layer, without needing SOQL expertise.
- Agentforce Vibes 2.0 — A browser-based, cloud-hosted VS Code environment (formerly Code Builder) that launches from Setup, giving you a fully authenticated, org-aware dev environment with no local installation.
- Agentforce Experience Layer (AXL) — Define a component once; it renders natively across Slack, Mobile, ChatGPT, Claude, Gemini, Teams, and any MCP-compatible client.
- Agent Fabric — Governance and policy enforcement layer so AI agents operate within your existing business rules and compliance controls.
- AgentExchange — A marketplace bringing together 10,000+ Salesforce apps, Slack apps, and Agentforce agents, including partner MCP servers from Google, Docusign, and Notion.
What is Model Context Protocol (MCP)?
The Model Context Protocol (MCP) is an open standard created by Anthropic that enables AI assistants to securely connect with external tools, APIs, and data sources. Think of it as "USB-C for AI" — a universal adapter that standardizes how any MCP-compatible AI client communicates with any MCP server, without custom integrations.
For Salesforce professionals, MCP is the reason Claude can directly query your data, trigger flows, run SOQL, and invoke Apex — all through natural language. You describe what you need; Claude figures out the tool to call and how to call it.
How MCP Works in Practice
When you ask Claude "Show me all open Opportunities closing this quarter," here's what happens under the hood:
Claude receives your natural language request
Your prompt is processed by Claude. It identifies that this requires a Salesforce data query and looks up the available MCP tools exposed by the connected Salesforce MCP server.
MCP tool discovery via tools/list
Claude calls the Salesforce MCP server's tools/list endpoint, which returns all available tools with their names, descriptions, input schemas, and output schemas. Claude selects the most appropriate tool — in this case, a SOQL query tool.
OAuth-authenticated tool call to Salesforce
The MCP server authenticates via OAuth 2.0 + PKCE using your External Client App credentials, then executes the SOQL query against your Salesforce org on your behalf.
Structured results returned to Claude
Salesforce returns the query results to the MCP server, which passes them back to Claude. Claude then formats and presents the data in a meaningful, human-readable response.
Security note: MCP uses the running user's field-level and object-level security automatically via WITH USER_MODE in SOQL queries. Claude can only access records and fields that the authenticated user has permission to see.
MCP Servers in Salesforce: DX vs Hosted
Salesforce provides two official approaches to MCP connectivity. Understanding the difference helps you choose the right setup for your workflow.
Salesforce DX MCP Server
Developer Preview- Runs locally using Salesforce CLI credentials
- 60+ tools for metadata, data, testing, LWC development
- No OAuth setup required — uses existing CLI auth
- Package:
@salesforce/mcpon npm - Best for: Developers with Salesforce CLI installed
- Ideal for agentic coding workflows in Claude Code or Cursor
Salesforce Hosted MCP Servers
Generally Available · April 2026- Runs on Salesforce infrastructure — no local CLI needed
- OAuth 2.0 + PKCE via External Client Apps
- 11+ endpoints: sobject-all, flows, data-360, and more
- Requires
mcp-remotefor stdio clients - Best for: Teams without CLI or cloud-first workflows
- This is what we use to connect Claude.ai in this tutorial
Which should you use? If you're a developer with Salesforce CLI already set up, start with the DX MCP Server — it's easier to configure and offers more comprehensive development tooling. Use Hosted MCP Servers when you want to connect Claude.ai (the web interface) or need OAuth-based access without installing the CLI.
The DX MCP Toolsets
The DX MCP Server organizes its 60+ tools into focused toolsets so you can scope access and avoid overwhelming the LLM's context window:
| Toolset | Description | Key Capabilities |
|---|---|---|
| orgs | Org management | List orgs, view details, check connections |
| data | Data operations | SOQL queries, record CRUD, bulk operations |
| metadata | Metadata management | Deploy, retrieve, list metadata components |
| users | User management | Permission sets, profiles, user assignments |
| testing | Test execution | Run Apex tests, agent tests |
| lwc-experts | LWC development | Component analysis, best practices, optimization |
| code-analysis | Static analysis | Code quality checks, security scanning |
| all | All toolsets | Enables all 60+ tools (use with caution) |
Hosted MCP Endpoints: What's Available
When Salesforce Hosted MCP Servers reached General Availability in April 2026, the URL pattern also changed. The new GA endpoint pattern for production orgs is:
https://api.salesforce.com/platform/mcp/v1/platform/<server>Note for sandbox orgs: Use /v1/sandbox/<server> instead of /v1/platform/<server>. The previous beta URL pattern (v1-beta.2) has been retired.
Available Server Endpoints
Beyond the core platform, the broader Salesforce ecosystem also ships several specialized MCP servers: the Heroku MCP Server (GA) for managing Heroku apps and dynos, the MuleSoft MCP Server (GA) for designing and operating MuleSoft projects, and the Mobile MCP Server (Developer Preview) for generating native mobile app code.
Step-by-Step: Connecting Claude to Salesforce
Here is the complete setup process from the video, broken down into every step. We are connecting Claude.ai (the web interface) to a Salesforce org using the Hosted MCP Server and OAuth via an External Client App.
Prerequisites
- A Salesforce org (Developer Edition is free and works perfectly — Hosted MCP is included at no cost)
- A Claude.ai account on a Pro, Team, Max, or Enterprise plan
- System Administrator profile in Salesforce (required to configure External Client Apps and MCP Settings)
Step 1 — Enable MCP Server in Salesforce Setup
Before creating anything, you need to ensure the MCP Server feature is enabled in your Salesforce org.
Navigate to Setup
Click the gear icon in the top-right corner of your Salesforce org and select Setup.
Search for "MCP" in Quick Find
In the Setup search box (Quick Find), type MCP. You will see MCP Settings in the results.
Enable MCP Server
Open MCP Settings and toggle the setting to Enabled. Save your changes. This activates the Salesforce Hosted MCP infrastructure for your org.
Step 2 — Create an External Client App
For Claude to authenticate against your Salesforce org, you need to create an External Client App. This is the OAuth 2.0 configuration that grants Claude permission to connect.
Go to External Client App Manager
In Setup, search for External Client App Manager in Quick Find and open it.
Click "New External Client App"
Click the New External Client App button in the top-right corner.
Fill in the Basic Details
Provide a Label (e.g., "Claude MCP Integration"), a Name (auto-populated), a Description, and a Contact Email. These are for your own reference.
Enable OAuth and Configure Settings
Toggle Enable OAuth to ON. Then configure the following settings exactly as shown:
| Setting | Value | Notes |
|---|---|---|
| Callback URL | https://claude.ai/api/mcp/auth_callback | This is the Claude-specific callback URL — do not change it |
| OAuth Scopes | mcp_api, refresh_token | Also labeled as "Access MCP servers" and "Perform requests at any time" |
| JWT-based access tokens | Enabled | Toggle this ON |
| PKCE | Required | Mandatory — do not leave as optional |
| All other security options | Unchecked | Leave defaults unless you have specific requirements |
Save the External Client App
Click Save. Note that External Client Apps can take a few minutes to propagate. If the Consumer Key isn't immediately visible, wait 2–3 minutes and refresh.
Copy the Consumer Key
Once saved, go to Settings → Consumer Key and Secret within the app. Copy the Consumer Key — this is your OAuth Client ID that Claude will use to initiate the authentication flow.
Important: Check Setup → Session Settings → "Force relogin after" and set it to None for agent-facing External Client Apps. If your org uses Login IP Ranges on the profile, agent requests coming from cloud-hosted AI services like Claude will fail, so ensure cloud IP ranges are permitted.
Step 3 — Activate Your MCP Server
Navigate to MCP Servers in Setup
In Quick Find, search for MCP Servers. You will see a list of available server endpoints.
Verify and Activate
Check that your desired servers (e.g., sobject-all for full sObject access) appear in the list. Toggle them to Active. A server must be active before any MCP client can connect to it.
Note Your MCP Endpoint URL
Your production MCP endpoint follows the pattern:https://api.salesforce.com/platform/mcp/v1/platform/sobject-all
You will need this URL when connecting from Claude.
Step 4 — Connect Claude to Salesforce
Now we connect Claude.ai to your Salesforce org using the MCP server URL and Consumer Key you've configured.
Open Claude.ai and go to Settings
Log into your Claude.ai account. Click your profile icon and navigate to Settings → Integrations (or the Connectors/MCP section, depending on your Claude plan).
Add a New MCP Server / Custom Connector
Click Add MCP Server or Add Custom Connector. You will be prompted to enter the server URL and OAuth Client ID.
Enter the Salesforce MCP Server URL
Paste the MCP endpoint URL from your Salesforce org, for example:https://api.salesforce.com/platform/mcp/v1/platform/sobject-all
Enter the OAuth Consumer Key
Paste the Consumer Key copied from your Salesforce External Client App. This is the client identifier Claude uses to initiate the OAuth flow.
Authenticate with Salesforce
Click Connect. Claude will redirect you to a Salesforce OAuth login screen. Log in with your Salesforce credentials and click Allow to grant Claude permission. You will be redirected back to Claude.ai.
Verify the Connection
Once connected, the Salesforce connector should show as active in Claude's Settings. Start a new conversation and try a prompt like: "List the 5 most recently modified Accounts in my Salesforce org." Claude will use the MCP tool to query your live data.
🎉 You're connected! Claude can now query your Salesforce data, trigger Flows, invoke Apex actions, and interact with your org — all through natural language, with no browser navigation in Salesforce required.
Real-World Use Cases
Once Claude is connected to your Salesforce org, the possibilities for AI-powered interactions are extensive. Here are some of the most impactful use cases you can explore immediately.
Data Queries in Natural Language
Ask Claude "Show me all Opportunities closing this quarter over $50k" and get a clean, formatted response from live Salesforce data — no SOQL required.
Trigger Flows and Automations
Invoke Salesforce Flows and Invocable Actions from Claude. Automate approval submissions, case escalations, or custom processes through conversation.
Developer Productivity
Use Claude Code with the DX MCP server to deploy metadata, run Apex tests, create custom fields, and get org-aware code suggestions without leaving your IDE.
Pipeline Intelligence
Expose Apex business logic through custom MCP tools. Claude can run pipeline health analyses, deal risk scoring, and product-line breakdowns with a single prompt.
Multi-Channel Agentforce
With Agentforce Experience Layer, AI agents using MCP can render the same workflow as a WhatsApp card, a Slack button, and a voice response — built once, deployed everywhere.
Data 360 & Tableau Next
Use the data-360 and tableau-next MCP endpoints to run SQL-based analytics and access Tableau dashboards directly through Claude conversations.
Tips, Best Practices & Troubleshooting
Best Practices
- Scope your toolsets: When using the DX MCP Server, only enable the toolsets you need (e.g.,
data,metadatarather thanall). This reduces context window usage and keeps Claude focused on the right tools. - Use explicit org aliases: Avoid
ALLOW_ALL_ORGSin production. Specify only the orgs needed using their aliases for better security. - Wait after creating External Client Apps: They can take 2–5 minutes to fully propagate in Salesforce. If authentication fails immediately after creation, wait and try again.
- Test with a Developer Edition first: Every Developer Edition org ships with Hosted MCP Servers at no cost. Use this to validate your setup before applying to a production org.
- Monitor API usage: Enterprise Edition orgs default to 100,000 API calls per day. Agentic workflows can consume these faster than traditional integrations — implement rate limiting in your agent configurations.
- Write descriptive
@InvocableVariableannotations: If you're building custom Apex MCP tools, the description text in your annotations becomes the documentation Claude reads. Write them with the agent as your audience.
Common Issues and Fixes
| Issue | Likely Cause | Fix |
|---|---|---|
| OAuth authentication fails | IP restrictions on the org profile | Check Login IP Ranges on the profile — cloud-hosted AI services need to be permitted |
| Consumer Key not visible after saving | External Client App still propagating | Wait 2–5 minutes and refresh the page |
| MCP server not visible in Setup | MCP Settings not enabled | Go to Setup → MCP Settings and toggle to Enabled |
| Claude can't see Salesforce tools | Server not set to Active | In Setup → MCP Servers, verify the server status is Active |
| Session timeout issues | Force relogin setting is too short | Set Setup → Session Settings → Force relogin after to None for agent-facing apps |
| Old beta URL not working | Using retired v1-beta.2 URL pattern |
Update to new GA pattern: /v1/platform/<server> |
Supported MCP Clients: Claude (Desktop, Web, Code), Cursor, Postman, and ChatGPT (developer mode) are all officially listed as supported clients for Salesforce Hosted MCP Servers. The JSON-based configuration format is similar across all clients.
Watch the Full Tutorial
See every step in action — from enabling MCP Settings in Salesforce to the first live query in Claude.
▶ Watch on YouTubeSubscribe for more Salesforce, Agentforce, AI, and automation content.
.png)
.png)
.png)
.png)
.png)
.png)
.png)
.jpg)
0 Comments