Search for "Claude Salesforce" today and you land in the middle of four different products that all get described the same way. A sales rep, a Salesforce admin and an Apex developer can each "connect Claude to Salesforce" and end up with completely different things.
This guide sorts them out: what each option is, who it is built for, what it can and cannot touch in your org, and how to set it up without handing an AI model more access than you meant to.
The short version
| You want to… | Use | Who sets it up |
|---|---|---|
| Let sellers prep calls, review pipeline and update opportunities from Claude | Salesforce in Claude (plugin, beta) | Salesforce admin + Claude owner |
| Let Claude read and write CRM records with each user's own permissions | Salesforce Hosted MCP servers as a Claude custom connector | Salesforce admin |
| Use Claude as the model behind Agentforce agents, Agentforce Vibes or Slackbot | Claude inside Salesforce (part of the Claudeforce partnership) | Salesforce, mostly by default |
| Ask Claude about your org itself: metadata, Apex, flows, debug logs, permissions, security posture | An admin-focused MCP server, such as orgadmin.ai | Each admin or developer |
| Write and fix Salesforce code with an agent in your terminal | Claude Code plus an MCP server | Each developer |
Most teams end up with more than one. They sit at different layers and do not conflict.
Why there are suddenly so many options
On 26 August 2026, Salesforce and Anthropic announced an expanded partnership they called Claudeforce. It runs in both directions:
- Salesforce inside Claude. A plugin, Salesforce in Claude, puts CRM data and sales workflows in the Claude app.
- Claude inside Salesforce. Salesforce says Claude is a reasoning model for the Atlas Reasoning Engine, powers Agentforce Vibes and Agentforce Coworker by default, is available in Agent Builder, and is the default model for Slackbot. It also says Claude runs inside the Salesforce Trust Boundary through Amazon Bedrock, which matters to regulated industries.
The plumbing underneath is the Model Context Protocol (MCP), the open standard that lets an AI client call tools on another system. Salesforce's own hosted MCP servers, third-party MCP servers and the plugin all use it. That is why one Claude account can talk to Salesforce in several ways at once.
Option 1: Salesforce in Claude (for sellers)
What it is. A Salesforce-built plugin for Claude with 37 prebuilt sales skills: account research, call prep, deal reviews, pipeline dashboards, follow-up emails and draft opportunity updates. It also pulls context from Slack and email.
Status. Anthropic announced the beta on 15 September 2026 and says it is "available in beta on all paid Claude plans". A Salesforce admin requests access through AgentExchange. Admins connect Salesforce once for the organisation and choose which groups get the plugin. Each seller then signs in with their own Salesforce credentials.
What it can touch. Claude only reads what the seller's Salesforce permissions allow. By default it asks the seller to approve each proposed change before writing it back, and Salesforce stays the system of record.
Good fit if your goal is seller productivity: fewer clicks to prep a meeting or update a stage. Not built for admin or developer work. It does not explain why a flow failed or who holds Modify All Data.
Option 2: Salesforce's hosted MCP servers as a Claude connector
What it is. Salesforce runs MCP servers you switch on per org. You add one to Claude as a custom connector. The standard servers include:
- SObject Reads (
platform/sobject-reads): read and query records, no changes - SObject All (
platform/sobject-all): full create, read, update, delete and search - Salesforce API Context (
platform/salesforce-api-context) - Metadata Experts (
platform/metadata-experts), which Salesforce's own documentation still labels beta
How to connect it to Claude (Salesforce's Headless 360 workshop walks through this):
- In Salesforce Setup, search MCP and activate the server you want. Start with SObject Reads.
- Create an External Client App for Claude, with Claude's callback URL
https://claude.ai/api/mcp/auth_callback, and turn on JWT-based access tokens for named users. - In Claude, open Customize → Connectors → + → Add custom connector.
- Paste the server URL, for example
https://api.salesforce.com/platform/mcp/v1/platform/sobject-reads. For a sandbox, add/sandbox/before the server name:.../v1/sandbox/platform/sobject-reads. - Under Advanced settings, paste the External Client App's Consumer Key as the OAuth Client ID. Then click Connect and log in to Salesforce.
What it can touch. Every call runs as the logged-in Salesforce user, so object permissions, field-level security and sharing rules all apply. That is a strong default, but SObject All lets Claude delete records if the user can. Give people the read server first, and enable write servers deliberately.
Common errors: invalid_client_id usually means the External Client App has not finished propagating, so wait and retry. A redirect failure means the callback URL is missing. JWT errors mean the JWT access token setting is off.
Good fit if you want Claude to answer questions about, and act on, records: accounts, opportunities, cases.
Option 3: Claude inside Agentforce, Slack and Agentforce Vibes
You do not "connect" this one. Under Claudeforce, Claude is a model choice (and in several places the default) inside Salesforce products: Agentforce agents built in Agent Builder, Agentforce Vibes for development, and Slackbot.
The question here is not setup but cost and control. These run on Salesforce's metering, not your Claude subscription. Agentforce usage draws on Flex Credits, and Agentforce Vibes has its own allowances. If you are weighing that against using your own AI subscription, see Agentforce, Einstein or your own AI key? and Is Agentforce Vibes free?.
Option 4: An MCP server for admins and developers
Options 1 and 2 are about records. Much of the work admins and developers do is about the org itself:
- "Which Apex classes reference
Discount__c?" - "Why did last night's deployment fail?"
- "Who can Modify All Data, and through which permission set?"
- "Explain this debug log and tell me which query is slowest."
- "Run a security review on the sandbox and summarise the critical findings."
Record-level MCP servers are not built for those questions. This is the gap the orgadmin.ai MCP server fills. It exposes the investigation surface of an org: object and field descriptions, metadata search, Apex, LWC and flow source, SELECT-only SOQL, debug logs, past deployment failures, and the same security review and tech debt engines the web app uses.
Connecting it to Claude takes about a minute:
- Copy the server URL from Settings → MCP access in orgadmin.ai (
https://orgadmin.ai/api/mcp). - In Claude (web, desktop or mobile), open Connectors (under Customize or Settings, depending on your Claude version) → Add custom connector, paste the URL, click Add, then Connect.
- Sign in to orgadmin.ai and choose scopes on the consent screen. Read-only is ticked by default.
You do not need a Salesforce External Client App. Your orgs are already connected in orgadmin.ai, and you choose per org whether the connector gets full access, read-only or nothing. That way an agent can read production while it builds in a sandbox.
On a Claude Team or Enterprise plan, an owner adds the connector once under Admin settings → Connectors. Each person then connects with their own orgadmin.ai account and gets their own access, never a shared one.
The guardrails that matter for admin work:
- Scopes hide tools. A read-only grant does not just refuse write calls. The write tools are never shown to Claude at all.
- Deploy is not in the tool catalog. Claude can read why a deployment failed and edit the fix in a Development Studio project, but a human ships it.
- It is cut off the moment you say so. Revoking the connection or deactivating the account stops the next call, even in the middle of a conversation.
- It runs on your Claude subscription. The model reasoning over your org is the one you already pay for in Claude, so a review driven over MCP spends none of your orgadmin.ai AI allowance.
The step-by-step guide covers Claude, Claude Code, ChatGPT and Codex, including what a working connection looks like.
Option 5: Claude Code for Salesforce development
If you write Apex, LWC or flows, Claude Code in your terminal or IDE is the natural home. It can read your local SFDX project and, through MCP, your actual org, so it stops inventing field names. We covered the setup in detail in Connect Salesforce to Claude Code and Codex with MCP. In short:
claude mcp add --transport http --scope user orgadmin https://orgadmin.ai/api/mcp
Then run /mcp in Claude Code and authenticate. You can add Salesforce's hosted servers the same way next to it.
Which should you choose?
- Sales team wants AI in their day: Salesforce in Claude. Let the plugin's approve-each-change default stand.
- Ops or analysts want to ask questions of CRM data: Salesforce Hosted MCP, SObject Reads only.
- You're building customer-facing agents: Agentforce, with Claude as the model.
- Admins and developers want an assistant that understands the org: an admin MCP server such as orgadmin.ai, in Claude or Claude Code.
These combine cleanly. One Claude account can have the Salesforce plugin, a Salesforce hosted server and the orgadmin.ai connector at the same time, and Claude picks the tool that fits the question.
A safety checklist before you connect anything
- Start read-only. For every connector, read access first and write access as a separate, deliberate decision.
- Connect sandboxes before production. Both Salesforce's hosted servers (
/sandbox/URLs) and orgadmin.ai (per-org access) make this easy. - Know whose permissions apply. Salesforce's servers run as the signed-in user. Check what a typical user can actually see before you roll it out.
- Keep humans on the irreversible steps. Deletes, deploys and mass updates should need a person's click.
- Know how to revoke. Make sure someone knows how to disconnect a connector in Claude and revoke the OAuth app in Salesforce or orgadmin.ai.
Claude and Salesforce now work together at almost every layer, from the seller's daily brief to the model inside Agentforce. The fastest way to get value without surprises is to match each connection to one job, start it read-only, and widen access only when you have seen it work.