On this page
- What does the Salesforce hosted MCP server give Claude?
- Which editions get it, and does it cost anything?
- How do I connect Claude to Salesforce, step by step?
- How do I test that the connection is safe?
- What can Claude read and change, and what does it not know about my org?
- How do I stop Claude from writing to production by accident?
- What can go wrong, and how do I undo it?
- Do I need anything on top of the Salesforce server, and when?
- Sources
- Common questions about the Salesforce MCP server for Claude
The Salesforce MCP (Model Context Protocol) server for Claude is a Salesforce-hosted endpoint. It lets Claude read and change your org as the signed-in user, and it has been generally available since 29 April 2026. Start with the read-only sobject-reads server and test every write in a sandbox.
Most setup guides stop at "connected". This one adds a two-prompt test that proves Claude cannot write, five rules that keep writes out of production, and the undo for a bad write.
Clientell AI, the AI Salesforce admin, gives Claude a map of your org on top of that server. Clientell AI MCP works in Claude Code, Claude Cowork, Codex, Cursor and VS Code, so you pick the client. New to Claude Code? Read Claude Code for Salesforce.
What does the Salesforce hosted MCP server give Claude?
MCP is an open standard that lets an AI tool call outside systems. A hosted MCP server is a Salesforce-managed endpoint that exposes your org's data, flows, Apex actions and queries to any AI client that speaks MCP. Salesforce hosts the server, runs the sign-in and enforces your permissions, so there is nothing to install (Salesforce Developers Blog, 29 April 2026).
Every call runs as the user who signed in. Field-level security (FLS, which controls which fields a user can see), object permissions and sharing rules still apply.
The standard servers are inactive by default. You turn on only the ones you want. Four of them work on records:
Swipe sideways to see all columns
| Server | What Claude can do | Use it when |
|---|---|---|
sobject-reads | Read records and run queries. No writes. | You want answers, reports and audits. |
sobject-mutations | Create and update records. No delete. | You test writes in a sandbox. |
sobject-deletes | Delete records only. | A controlled cleanup for a trusted user. |
sobject-all | Read, create, update and delete. | A sandbox you can throw away. |
Salesforce lists these in its servers reference. A SalesforceTrails guide from 7 June 2026 names three more: salesforce-api-context, metadata-experts and data-cloud-queries. You can also build custom servers from invocable Apex actions, autolaunched flows and Apex REST endpoints.
Which editions get it, and does it cost anything?
Every Enterprise Edition org and above gets it, per the general availability post of 29 April 2026.
Billing may change. On 22 September 2026 Salesforce Ben reported that successful MCP and API calls from registered agents will use Flex Credits, with no published rate yet. Check Salesforce's current terms before you plan around cost.
How do I connect Claude to Salesforce, step by step?
Setup has three parts. Salesforce's step-by-step walkthrough (26 May 2026) keeps the exact screens current. This is the order it follows:
- Activate a server. In Setup, search for MCP Servers, open the Salesforce Servers tab, click the server name and click Activate. Copy the server's API Name and its Server URL.
- Create an External Client App (ECA). This is the OAuth (open authorization) app that lets Claude sign in without sharing your password. In Setup, search for External Client App Manager and click New External Client App. Check Enable OAuth, add the scopes
mcp_apiandrefresh_token, offline_access, and check Require Proof Key for Code Exchange (PKCE) and Issue JSON Web Token (JWT)-based access tokens for named users. Uncheck Require secret for Web Server Flow and Require secret for Refresh Token Flow. Save the app, then copy the Consumer Key from the app's OAuth Settings. - Add the server to Claude. For Claude Desktop, set the callback URL to
https://claude.ai/api/mcp/auth_callbackand add a custom connector. For Claude Code, set the callback URL tohttp://localhost:38000/callbackand runclaude mcp addwith the server URL and your Consumer Key.
Then limit who can use the app. In the app's policies, restrict access to pre-authorized users and assign one permission set, for example a set named MCP Client User. Without this, any user in your org can connect to the ECA (Salesforce Developers Blog, 30 June 2026).
When you finish, run /mcp in Claude Code and choose Authenticate. A successful connection shows the server as "connected". If a server does not show up, check that you activated it, because standard servers start inactive.
Three choices matter more than the clicks:
- Activate
sobject-readsfirst. Leave every write server off. - Use the sandbox server URL first. For a sandbox, Salesforce swaps
/v1/platform/for/v1/sandbox/platform/in the URL. - Make one ECA per client. Salesforce recommends a dedicated ECA for Claude, another for ChatGPT, and so on, so you can audit each one (Salesforce Developers Blog, 30 June 2026).
Want the same org map in every client? The Clientell AI MCP install steps cover Claude Code, Claude Cowork, Codex, Cursor and VS Code, one config block each.
How do I test that the connection is safe?
Run two prompts right after you connect sobject-reads:
- Ask: "Using Salesforce, count the Opportunity records with
IsClosedset to false." Expected result: a count that matches the same filter in a Salesforce report. - Ask: "Set the Amount on one Opportunity to 1." Expected result: Claude says it has no tool that can write. If it writes, a write server is active, so deactivate it in Setup.
If prompt 1 fails with an authentication error, the ECA callback URL is the first thing to check.
What can Claude read and change, and what does it not know about my org?
Claude can do what you can do, through the server you picked. When Claude updates a record, your name sits in the audit trail. If your permissions do not allow an operation, Claude cannot do it either.
The server does not carry your team's history. It returns records and metadata, such as a Flow definition or a ValidationRule formula. It does not say why that flow exists, who asked for it, or what broke the last time someone changed it. Claude starts each session from the data, not from the decisions behind it.
For example, ask Claude why Opportunity.StageName flips after a save. It can read the flows. It cannot tell you which one your team meant to keep.
Clientell AI MCP covers that gap. Its Context Graph indexes your objects, fields, flows and permissions. You describe a change in plain words, and your client drafts it against your real org.
How do I stop Claude from writing to production by accident?
Salesforce describes three security layers, plus logging. Each layer has its own job, and a gap in one is not covered by another (Salesforce Developers Blog, 30 June 2026).
Swipe sideways to see all columns
| Layer | What it controls | What to set |
|---|---|---|
| Authentication | Who is making the request | Users sign in through the ECA with the authorization code flow. |
| Authorization | What that user may reach | Grant the mcp_api scope only, and add an app policy that limits the ECA to one permission set. |
| Permission controls | Object and field access | Sign in as a user whose profile and permission sets fit the job. |
Salesforce does not support a shared service account for these sessions. Each session ties to one named user, which is why the audit trail stays honest.
A safe default has five rules:
- Connect only
sobject-readsto production. - Connect
sobject-mutationsto a sandbox only, using the sandbox server URL. - Sign in as a user with only the access the job needs, such as read on
OpportunityandAccount. Never use a System Administrator. - Read the change before you approve it. In Claude Code, choose approval for each call to a write tool such as
sobject-mutations, and do not allow it for the whole session. - Keep a way back. Export the records before a bulk change (Data Export under Setup), and test the undo in the sandbox first.
Clientell AI follows the same routine. It scans your org read-only first, tests changes in a sandbox, and asks before it changes anything in production. On the Clientell AI Platform, deployments also have rollback. See security.
Try it in Clientell AI: "Using the Clientell AI MCP, list every flow that writes to Opportunity.Amount in read-only mode, and explain in plain words what each one does."
Book a call with an expert, or start the 14-day trial, no card.
What can go wrong, and how do I undo it?
Most failures are access problems, not bugs. Three show up first:
Swipe sideways to see all columns
| Symptom | Likely cause | Fix |
|---|---|---|
| The server is missing in Claude | The server is still inactive in Setup | Activate it under MCP Servers, then reconnect. |
| Sign-in loops or fails | The callback URL on the ECA does not match the client | Use the Claude Desktop URL or the Claude Code URL exactly as above. |
| A write is refused | Your user lacks edit access to the object or field | Fix the permission set, not the server. |
For a bad write, your name is in the audit trail, so you can find what changed. Restore the field values from your Data Export file with Data Loader (an update on the record Id). Never run a bulk write in production before it has run in a sandbox. If you plan to run several agents against one org, read what Salesforce has said about registering AI agents first.
Do I need anything on top of the Salesforce server, and when?
Not to read records or run queries as yourself. Salesforce's server does that well, and it is the better choice when you need Salesforce-native custom tools built on your own Apex or flows. Add a layer when one of these is true:
- Five admins use Claude on the same org, and you want one shared record of which flows and fields each change touched.
- Writes need a reviewer, not just a confirmation click.
- Metering changes. If calls start costing credits, every lookup of the Flows on
Opportunitycosts a call. A tool that already holds a map of your org may need fewer calls.
Clientell AI MCP costs $249 a month for 1 org and 10 people, with +$25 a person after that. Book a call with an expert, or start the 14-day trial, no card. See pricing for Clientell AI Platform at $500 a month, or read Clientell AI vs Claude Code.
Common questions about the Salesforce MCP server for Claude
Is Salesforce's MCP server free?
Salesforce includes it for every Enterprise Edition org and above. Salesforce Ben reported on 22 September 2026 that Flex Credits will apply to successful agent calls, with no published rate (Salesforce Ben). Check Salesforce's current terms.
Does it work with Claude Code and Claude Desktop?
Yes. Salesforce documents both in its Claude walkthrough. Claude Desktop connects through a custom connector, and Claude Code connects with claude mcp add.
Can Claude delete records?
Only if you connect sobject-deletes or sobject-all, and only records your signed-in user can delete. Use sobject-reads in production.
Does it work in a sandbox?
Yes. Salesforce documents a sandbox server URL. Keep write servers there.
Does Clientell AI MCP replace Salesforce's server?
No. It runs next to it. Salesforce's server gives Claude the data. Clientell AI gives your client the org map and drafts changes, in Claude Code, Claude Cowork, Codex, Cursor or VS Code.
Sources
Salesforce documentation
- Salesforce Hosted MCP Servers Are Now Generally Available
- Connect Claude with Salesforce Hosted MCP Servers
- How to Secure Salesforce Hosted MCP Servers
- Hosted MCP servers reference
Industry research
- Salesforce Hosted MCP setup for Claude (SalesforceTrails)
- Flex Credits for agentic MCP and API calls (Salesforce Ben)
Last reviewed October 2026
Last updated: October 22, 2026



