Claude Salesforce Security: A Governance Checklist Before You Connect

Checklist diagram of Claude, an MCP server and Salesforce, with access limits marked at each of the three layers

Claude Salesforce security is the set of controls that decide what Claude can see and change once it is connected to your org: which Salesforce user it acts as, which objects and fields that user can reach, which environment you connected and how you will review or reverse what happens. You set most of these controls before the first prompt, and each one takes a few minutes.

SF Ben's 2026 Salesforce Admin Survey found that almost 44% of respondents use AI daily or regularly, up from just under 25% the year before. Many of those admins are now deciding whether Claude should get a live connection to the org. This checklist covers what to settle first. It assumes you use a Salesforce MCP server, the connector between Claude and Salesforce. If you want the click-by-click, our setup guide for Claude and a Salesforce MCP server covers the Connectors screens in Claude Desktop and claude.ai.

Where the risk sits in a Claude and Salesforce connection

Three parties take part. Claude is the AI client where you type a request. An MCP server sits between Claude and Salesforce, offers tools such as running a query or creating a field, and signs in to your org with OAuth. Salesforce decides what the signed-in user may do. Each party gives you its own control, and a gap in one is not covered by the other two.

The checklist below follows that order: controls in Claude, controls in the connector and controls in Salesforce. Work through it once for the first org, then reuse the scorecard near the end for every org you connect later.

Step 1: Decide who may add a connector to Claude

Anthropic's custom connectors documentation says that on Team plans only Owners and Primary Owners can add custom connectors. On Enterprise plans, Owners, Primary Owners and people with a custom role that includes Manage access to Libraries can add them. Owners add the connector under Organization settings > Connectors. Members then connect from Customize > Connectors and enable it themselves, so Claude can reach only what each person's own login can reach. Anthropic's plugins article says the sidebar label Customize becomes Plugins, generally available November 9, 2026 for Free, Pro, Max, Ultra and Team, and on November 30 for Enterprise. The exception is a connector that uses a shared credential such as an API key, which gives everyone who uses it that credential's access.

Name one person who owns the Salesforce connector in your company, and prefer an OAuth login per person over a shared credential. To remove a connector, open Customize > Connectors in Claude. Team and Enterprise Owners can also remove it for the whole organization under Organization settings > Connectors.

Step 2: Choose the Salesforce user Claude will act as

The Salesforce API follows your security model. Salesforce's data access documentation states that the API respects object-level and field-level security, and that for most calls data outside the logged-in user's sharing model is not returned. Two system permissions override sharing. View All Data lets a user view all records regardless of sharing settings, and Modify All Data lets a user read, edit, delete, transfer and approve all records regardless of sharing settings. A System Administrator login usually carries both.

If you connect Claude with your own admin login, Claude inherits all of that. For anything beyond a first look, create a separate user for Claude work, start it on a minimal-access profile, and add a purpose-built permission set. Permission sets only add access, so check that no profile, permission set or permission set group assigned to that user grants more than the task needs. A workable starting point looks like this:

  • Object access: read on the objects the task involves, with create or edit added only where the task needs it.
  • Field access: field-level security only for the fields the task touches.
  • No override permissions: leave out View All Data and Modify All Data.
  • No metadata write unless the task builds: leave out Modify Metadata Through Metadata API Functions, Author Apex and Deploy Change Sets. Salesforce's Metadata API documentation says a user needs Modify Metadata Through Metadata API Functions or Modify All Data to edit metadata, and that Author Apex and Deploy Change Sets turn on the first one automatically. Taking away create, edit and delete on records leaves these in place.
  • No user administration: leave out Manage Users unless the task is user provisioning.

If you are still deciding how to structure that access, our guide to permission sets vs profiles explains the trade-offs, and the guide to least privilege for AI agents goes deeper on scoping an agent's access.

Step 3: Connect a sandbox before production

Run the first week of requests against a sandbox. A wrong field, flow or permission set costs nothing there, and you can review each change before it gets near users. The sandbox best practices guide covers sandbox types and how to move a change to production. Cirra AI does not run a sandbox-to-production pipeline. If you connect a production org, a change Claude makes there takes effect immediately, so set every Access Rules category for that org to Read Only until you are ready to work in it. One Cirra AI account can hold several orgs, so you can add production later without starting over.

Step 4: Check what Salesforce authorizes when you add the org

With Cirra AI, the Salesforce authorization happens in your Cirra AI account, before Claude is involved. When you add an org there, you sign in to Salesforce through OAuth and Salesforce asks you to approve the access the app requests. Sign in as the user you chose in Step 2, because that login decides which permissions the connection carries. Read the access Salesforce lists before you approve it. Cirra AI never sees your password.

Some MCP servers ask an admin to register an External Client App in the org first. That includes Salesforce's own hosted MCP servers: the setup overview says creating the app requires System Administrator or equivalent permissions and calls for Proof Key for Code Exchange (PKCE) and JWT-based tokens. If your org reviews new connected apps, send this one through the same review.

Step 5: Connect Claude and set limits in the connector

When you click Connect in Claude, a browser tab opens with the Cirra AI MCP Server - Authorization Request. This screen asks you to let Claude use your Cirra AI account, and it lists only openid, email, profile and offline_access. It says nothing about Salesforce access, because you settled that in Step 4. Click Allow Access once you recognize the request as your own.

Then set limits inside the connector. Cirra AI Access Rules let you limit what it can read and write, down to a single object or field. The product groups org content into four categories (Metadata, User Data, Personal Setup Data and Business Data), and you set each one to None, Read Only or Read and Write. Building a validation rule needs Metadata access, and you can set Business Data to None for that work. Team governance settings also let you control which org types can connect.

Step 6: Decide how Claude asks before it acts

Claude lets you control tool permissions per connector. Anthropic's documentation advises clicking Allow always only when you use a server and tool you trust to run unsupervised. It also describes a Search and tools menu where you can turn off tools a conversation does not need, and it advises disabling tools that can take write actions in Research mode, because Research can call connector tools without further approval.

For a Salesforce connection, that translates into a simple habit: skip Allow always for any tool that writes to your org until you have watched it work on a sandbox, and switch off write tools in conversations that only need answers.

Step 7: Treat text inside your org as untrusted input

A connected assistant reads records, and records hold text that customers, partners and web forms wrote. A case description or lead comment can contain sentences that read like instructions to an AI model. Security teams call this prompt injection. You cannot filter every record, so limit what a bad instruction could do.

Keep Business Data on Read Only or None when the task is a metadata task. Avoid combining a request to summarize customer-written text with write access in the same conversation. Read the proposed change before you approve it, especially when it was triggered after Claude read a batch of records.

Step 8: Make every change visible and reversible

Cirra AI documents every change it makes. Salesforce keeps its own record as well. The Setup Audit Trail retains 180 days of setup changes, shows the 20 most recent entries in the browser and lets you export the rest as a CSV file from View Setup Audit Trail in Setup. Add a one-line log of your own with who asked, what changed, which org and how you tested it.

Practice the exit before you need it. In Setup, type "OAuth" in Quick Find and open Connected Apps OAuth Usage to see each app with live user tokens. From there you can revoke one user's access, revoke everyone's or block the app, which ends its current sessions and stops new ones. If your org connected Cirra AI through an External Client App instead, open Apps > External Client Apps > OAuth Usage in Setup, which offers similar controls. Individual users can also revoke an app themselves: in personal settings, open Advanced User Details and find the OAuth Connected Apps section.

Step 9: Test as the user and ship through your normal process

Log in as someone with the same permissions as the people who will use the change. A test as System Administrator proves little, because that login sees everything. Once the result behaves correctly in the sandbox, deploy it through whatever your team already uses, whether that is change sets, DevOps Center or another pipeline. If production is also connected, set it to Read Only in Access Rules until you are ready, because a change Claude makes there takes effect immediately. The same review that applies to a colleague's change applies to Claude's.

A pre-connection scorecard

Use this table for each org before you connect it. Every row should have an answer you can say out loud.

AreaQuestion to answerWhere to check
Connector ownerWho added the connector, and does each person sign in with their own OAuth login?Claude: Organization settings > Connectors for the owner who added it, and Customize > Connectors for each member
Salesforce userIs it a separate user with a purpose-built permission set, no View All Data or Modify All Data, and no metadata write permissions the task does not need?Setup: Users and Permission Sets
EnvironmentIs the first connection a sandbox, and is any connected production org set to Read Only?The org list and Access Rules in your Cirra AI account
Connector limitsIs each data category set to the least access the task needs?Cirra AI Access Rules
Tool approvalsDo write tools still ask before they run?Claude tool permissions
Change recordCan you reconstruct who changed what last week?Setup: View Setup Audit Trail, plus your own log
Exit planDo you know how to revoke or block the connection today?Setup: Connected Apps OAuth Usage, or Apps > External Client Apps > OAuth Usage

Where Cirra AI fits

Cirra AI is the first commercial Salesforce Admin MCP Server. It is hosted, so there is nothing to install. You authenticate your org with OAuth and paste the MCP URL into Claude (Desktop, Code and Cowork) or another MCP-compatible client such as ChatGPT or Cursor. You describe the change, and Cirra AI makes it as the connected user, within that user's permissions, and documents it.

Cirra AI reads and writes your org's metadata layer, including objects, fields, page layouts, record types, validation rules, flows, Apex and permissions, and it queries and updates data. The limits from Steps 2 to 5 apply to all of it: the connected user's permissions set the ceiling, and Access Rules can lower it further.

Cirra AI supports Salesforce Enterprise Edition and up. The Free plan includes 100 credits a month and needs no credit card, which is enough to run a first request against a sandbox. Start Free Trial when you are ready, or Learn More about the Salesforce Admin MCP Server.

Menu names in Claude and Salesforce follow current documentation and can change between releases, so confirm each path in your own org before you rely on it.

Frequently asked questions

Is it safe to connect Claude to Salesforce?

It is safer when you limit the damage a wrong request can do. Use a separate Salesforce user with a narrow permission set, connect a sandbox first, set connector limits to the least access the task needs, keep write tools on ask-first and know how to revoke the connection. Cirra AI runs as the connected user, can do only what that user can do and documents every change.

What can Claude do in my Salesforce org?

It depends on the MCP server you connect and the Salesforce user behind the connection. Salesforce's API respects object-level and field-level security, and sharing rules limit which records come back. A server with metadata write access, such as Cirra AI, lets Claude also create fields, flows and permission sets within that user's permissions.

Should I connect Claude to a sandbox or to production first?

Connect a sandbox first. Reading tasks are low risk anywhere, and building tasks are easier to review where a mistake costs nothing. Move changes to production through your normal deployment process after you have tested them as a real user. If you connect production as well, changes Claude makes there take effect immediately, so keep that org on Read Only in Access Rules until you are ready.

Can I make Claude read-only in Salesforce?

Yes, but a permission set alone is not enough. Permission sets only add access, so start the connected user on a minimal-access profile. Then check that no profile, permission set or permission set group assigned to that user grants create, edit or delete on records, or a permission that lets the user edit metadata: Modify All Data, Modify Metadata Through Metadata API Functions, Author Apex or Deploy Change Sets. In Cirra AI, also set each data category to Read Only in Access Rules. Both limits apply, and the narrower one wins.

How do I disconnect Claude from Salesforce?

Remove the connector in Claude under Customize > Connectors. To cut off the Salesforce side as well, open Connected Apps OAuth Usage in Setup and revoke the tokens or block the app. If your org connected through an External Client App, use Apps > External Client Apps > OAuth Usage instead.

Which Claude apps work with Cirra AI?

Cirra AI works with Claude (Desktop, Code and Cowork), ChatGPT, Cursor, Gemini Enterprise, OpenAI Codex and VS Code through GitHub Copilot, plus any MCP-compatible AI client. It supports Salesforce Enterprise Edition and up.

Share the Post:

Related Posts