Quick answer: To connect Claude to Salesforce MCP, enable a Salesforce Hosted MCP Server in Setup, create an External Client App for OAuth, and add a custom Salesforce Claude connector in Claude (web, desktop, Claude Cowork, or Claude Code) that points at the server URL. As of April 2026, the Salesforce Hosted MCP Servers are generally available for Enterprise Edition orgs and above so bringing Salesforce and Anthropic’s Claude together no longer requires custom API code. Every request runs under the logged-in user’s own permissions.
Last updated: June 2026 · ~12 min read
@salesforce/mcp server (best for developers in Claude Code), and third-party MCP connectors (Windsor.ai, CData, and others).sobject-reads server first, validate in a sandbox, and only graduate to write-enabled servers like sobject-all once you trust the workflow.Salesforce Claude MCP is the use of the Model Context Protocol to connect Anthropic’s Claude with a Salesforce org so Claude can securely query and act on CRM data using the logged-in user’s permissions. It replaces brittle, custom-coded API integrations with a single standardized protocol that any MCP-compliant AI client can speak.
The Model Context Protocol is an open standard drafted by Anthropic and since adopted by OpenAI, Google, Microsoft, and Salesforce that defines how AI applications connect to external tools and data. It is often described as “USB-C for AI”: instead of building a separate connector for every system an AI agent needs to touch, you expose one MCP server and any compatible client can plug in.
In Salesforce terms, an MCP server is a managed endpoint that exposes your org’s logic and assets objects, fields, SOQL queries, Flows, and Apex actions as structured tools an AI client can call. Salesforce handles the hosting, authentication, and permission enforcement. Claude is the host/client: it sends a structured request, the Salesforce MCP server validates the user and returns only authorized data, and Claude turns the result into a natural-language answer.
With Salesforce Claude MCP, Claude can:
Bringing Salesforce and Anthropic’s Claude together solves a problem that is mostly about human bandwidth, not Salesforce itself. Sales reps, service agents, and executives all know the data exists they just can’t get to it fast enough without building reports, clicking through list views, or asking an admin.
When you connect Claude to Salesforce MCP, people interact with the CRM in natural language and get answers (and actions) in seconds. Instead of navigating dashboards, a user can simply ask:
The business impact lands in five areas:
| Benefit | What changes |
|---|---|
| Faster decisions | Answers come from live CRM data in seconds, not from manually built reports. |
| Higher productivity | Reps and agents complete lookups, summaries, and follow-ups without leaving their AI assistant. |
| Better CX | Instant access to full customer history enables more personalized, lower-handling-time service. |
| Developer velocity | Claude generates and reviews Apex, SOQL, LWCs, and Flow documentation as a development copilot. |
| Lower integration cost | A standardized MCP layer removes much of the custom API plumbing you’d otherwise maintain. |
There isn’t one path there are four, and choosing the right one is the single most important decision in your rollout. Here’s how they compare.
| Method | Best for | Where Claude runs | Setup effort |
|---|---|---|---|
| Salesforce Hosted MCP | Most business + analytics users | claude.ai (web), Claude Desktop, Claude Cowork | Low – enable a server, add a connector |
Local @salesforce/mcp | Developers writing/deploying code | Claude Code, Cursor, VS Code | Medium – runs on your machine |
| Agentforce native MCP client | In-Salesforce agentic workflows | Inside Salesforce (Agentforce is the host) | Admin – configured via Agent Builder |
| Third-party MCP connectors | No-code analytics & multi-source | claude.ai via a vendor’s MCP server | Low – vendor-managed |
A few notes on each:
@salesforce/mcp runs on your own machine and is the better fit for developer workflows this is the Claude Code Salesforce MCP path for reading metadata, deploying, and running anonymous Apex.The flow is deliberately simple and security-first. Five steps:
The critical design detail: MCP is itself a safety mechanism. The server defines exactly which operations are exposed, so Claude can’t call an API that hasn’t been explicitly turned on. Structured tool calls replace open-ended API access – a meaningfully smaller attack surface than handing an AI raw API credentials.
This is the most common way to connect Claude to Salesforce MCP. It takes roughly 15–20 minutes the first time. Do it in a sandbox or Developer Edition org first – never wire a brand-new Salesforce Claude connector straight into production.
sobject-reads (read-only) for safety.⚠️ Sandbox vs. production URLs differ. Sandbox/scratch server URLs contain
/sandbox/in the path. Grabbing the wrong one is the single most common cause of a silent connection failure.
https://claude.ai/api/mcp/auth_callbackmcp_api, refresh_token, offline_access.The
mcp_apiscope is purpose-built: it grants access to MCP only, not to your existing REST APIs. That separation is a feature it keeps the AI integration narrowly scoped.
Salesforce — Prod (read-only)) and paste the Server URL from Step 1.Before you start prompting, open the connector in Claude and set approval rules per tool. A sensible default:
get_user_info, find, soqlQuery, getObjectSchema, getRelatedRecords) → Always allowcreateSobjectRecord, updateSobjectRecord, updateRelatedRecord) → Ask every timeReads are safe to automate; writes deserve a human in the loop.
Confirm Claude only surfaces data the connected user is allowed to see, then test the same prompts as a lower-privilege user. If permissions are correctly inherited, the lower-privilege user should get a narrower result set automatically.
Tip: If Claude answers from general knowledge instead of calling the connector, tell it explicitly: “Use the Salesforce SObject Read connector to answer this.”
Developers usually want Claude inside their IDE workflow. There are two routes to connect Claude to Salesforce MCP from Claude Code.
Option A Hosted MCP from Claude Code. Same Hosted MCP servers as above, configured via the command line. The OAuth callback differs: Claude Code uses http://localhost:38000/callback (used when you only have an enterprise-provisioned Anthropic API key). You register the server with a name following the convention salesforce-<server-api-name> (for example, salesforce-sobject-all).
Option B Local @salesforce/mcp. For pro-code work, the local server is the stronger fit. It wraps the Salesforce DX tooling and exposes capabilities like deploying metadata, reading the schema, and running anonymous Apex as MCP tools ideal for Claude Code, Cursor, and VS Code.
A practical rule of thumb: use Hosted MCP for data and analytics work, and the local server for development and deployment work.
One reason this integration is powerful: you don’t get one generic server you get a catalog of purpose-built ones, each with a clear scope. Activate only what each team needs.
| Server | What it exposes | Typical use |
|---|---|---|
sobject-reads | Read-only schema, SOQL, search, related records | Safest starting point; analytics & lookups |
sobject-all | Everything in reads plus create / update / delete | “All-in-one” for full data workflows (~90% of demos) |
flows | Trigger Salesforce Flows | “When X happens, run our standard playbook” |
invocable-actions | Run invocable Apex actions | Domain-specific business logic |
data-cloud-sql | Query Data Cloud via SQL | Unified, cross-source customer data |
tableau-next | Tableau analytics access | Embedded BI questions |
prompt-builder | Salesforce prompt templates | Reusable, governed prompt workflows |
custom-servers | ISV/partner-built tools (Apex-backed) | Industry-specific schemas and actions |
For most teams, the pattern is: sobject-reads for the sales/ops floor, sobject-all for a small trusted group, and flows once you’re ready to let Claude kick off automations.
This is where the GA architecture earns enterprise trust. Six things are true by design:
mcp_api scope grants MCP access without exposing your existing REST APIs a clean, narrow blast radius.Best-practice security checklist:
sobject-reads (read-only) before enabling any write server.Pull stale-deal lists, rank pipeline by value, and surface single-threaded accounts then have Claude draft personalized follow-ups. With a Salesforce Claude Cowork setup on the desktop, a rep can run these multi-step workflows in plain English without opening Lightning. Example: “List open opportunities closing this quarter with no activity in 14 days, grouped by rep, ranked by value, so I can coach this week.”
Generate instant summaries of a customer’s case history to cut average handling time, and draft responses grounded in real account context.
Analyze campaign performance and segmentation conversationally, and connect Data Cloud (via data-cloud-sql) for cross-source customer insight.
Get pipeline summaries and forecasts without touching a dashboard. A finance team can prompt: “Compare this quarter’s closed-won opportunities against GL revenue and flag discrepancies over $10,000.”
Use Claude Code with the local server to read metadata, write and review Apex, optimize SOQL, build LWCs, document Flows, and run tests a development copilot rather than a replacement.
Accelerate solution design, documentation, and implementation planning, and stand up read-only Claude connectors as a fast, low-risk client demo.
@salesforce/mcp server if your primary users are developers working in Claude Code, Cursor, or VS Code who need metadata, deployment, and Apex execution.Many organizations run more than one: read-only Hosted MCP for the business floor, local server for the dev team, and a third-party connector for cross-source RevOps reporting.
@salesforce/mcp server if your primary users are developers working in Claude Code, Cursor, or VS Code who need metadata, deployment, and Apex execution.Many organizations run more than one: read-only Hosted MCP for the business floor, local server for the dev team, and a third-party connector for cross-source RevOps reporting.
The “Salesforce + Claude” conversation has moved from theory to production. With Hosted MCP servers generally available, your own org becomes a secure, governed MCP endpoint, and the decision to connect Claude to Salesforce MCP takes minutes no custom middleware, no service accounts, no new security model. The result is conversational access to CRM data and actions that respects every permission you’ve already configured.
The winning rollout pattern is consistent: start read-only, validate in a sandbox, scope access with permission sets, keep writes human-approved, and expand server-by-server as trust grows. Teams that adopt this discipline get the productivity upside of AI-powered CRM without compromising governance.
Planning a deployment? Working with a certified Salesforce partner can help you scope the right servers, configure OAuth and permission sets correctly, and roll out Claude safely across sales, service, and development teams.
It’s the use of the Model Context Protocol an open standard originally created by Anthropic to connect Claude with a Salesforce org so Claude can securely query and act on CRM data under the logged-in user’s permissions, without custom API code.
Yes. Salesforce Hosted MCP servers became generally available in April 2026 for Enterprise Edition orgs and above, and Salesforce publishes official documentation for connecting Claude (web, desktop, and Claude Code).
Yes, when configured correctly. Claude runs as the authenticated Salesforce user, so CRUD, field-level security, and sharing rules all apply, every action is auditable under that user’s name, and the mcp_api OAuth scope keeps access scoped to MCP only.
Yes if you activate a write-capable server like sobject-all and the connected user has permission. With the read-only sobject-reads server, Claude can only read. Setting write tools to “Ask every time” keeps a human in the loop.
No. The Hosted MCP path is admin-driven: activate a server in Setup, create an External Client App, and add a custom connector in Claude. Developers benefit most from the separate local @salesforce/mcp server for Claude Code.
Hosted MCP is Salesforce-managed and best for data/analytics work from claude.ai or Claude Desktop. The local server runs on your machine and is best for developer tasks in Claude Code reading metadata, deploying, and running Apex.
Agentforce is Salesforce’s in-platform agent framework, where Salesforce is the host consuming external tools. Connecting Claude via Hosted MCP works the other direction it brings your Salesforce data into Claude, so you don’t have to log into Lightning for everyday lookups.
Standard objects like Accounts, Contacts, Leads, Opportunities, Cases, Activities, and Campaigns, plus your custom objects all subject to the connected user’s permissions.
The Salesforce Claude connector is the custom connector you add inside Claude that points at a Salesforce Hosted MCP Server URL and authenticates via OAuth. Once connected, Claude can query and act on your Salesforce data using the logged-in user’s permissions.
Yes. The same Hosted MCP connector works in Claude Cowork (Anthropic’s agentic knowledge-work app). A Salesforce Claude Cowork setup lets reps run multi-step CRM workflows lookups, summaries, drafting, and logging activities in plain English from the desktop.
For Claude Code, you can connect to a Hosted MCP server via the command line (callback http://localhost:38000/callback) or run the local @salesforce/mcp server. The local server is the better Claude Code Salesforce MCP path for developer tasks like reading metadata, deploying, and running anonymous Apex.
Cloudespacio, headquartered in India, is a prominent independent Salesforce and leading consulting firm, dedicated to prioritizing client satisfaction.
Copyright © 2025 All Rights Reserved. Designed by Navpatra.