Claude and Salesforce · setup, not a pilot deck
Your CRM, answerable in chat.
Switched on properly, and bounded.
Anthropic and Salesforce released Salesforce in Claude in beta on 15 September 2026: a plugin that brings accounts, opportunities and pipeline into Claude under each person’s own Salesforce permissions, with 37 prebuilt skills for the work sellers do daily. Switching it on is an administrator job on both sides, and the second half of it is permissions. These are the three sizes of that job, priced in hours like everything else here.
Independent practice, not affiliated with Anthropic or Salesforce and not a reseller of either. Every vendor fact below was read from the vendor’s own page on 19 October 2026 and carries its link.
What it actually is
A plugin in beta, not a platform change.
Anthropic released it in beta on 15 September 2026: 37 prebuilt skills, working in chat and in Claude Cowork on web and desktop, with every proposed write to Salesforce approved by the person before it happens. It runs under the signed in user, so it inherits that user’s permissions rather than bringing its own.
It is the first product of a partnership announced on 14 October 2025, which also put Slack context into Claude from 2 February 2026 on the Pro, Max, Team and Enterprise plans. Agentforce is a separate product with its own licence and is not required for any of this.
What that means in practice: a seller asks what changed on an account this week, what to prepare for a call, or which opportunities have no next step, and gets an answer from your own records instead of a request to somebody who will get to it tomorrow.
Before anything is switched on
Four things have to be true, and two of them are not mine to arrange.
Read from the setup article for organisations and from Salesforce’s own documentation for activating hosted MCP servers, both on 19 October 2026. If one of these is missing, the honest answer is that the work waits rather than starts.
Salesforce approves the beta, and nobody else can
Access is requested by your Salesforce admin through the AgentExchange listing, and the acceptance email carries the setup instructions. I cannot obtain it for you and will not pretend otherwise. Until it arrives there is nothing to configure and nothing to pay.
Your org is on the latest Sales Cloud Enterprise edition
That is what the beta states. If your org is on a different edition, you find out before any work starts rather than halfway through it.
Somebody holds Owner on the Claude side
The connector is configured by a person with Primary Owner or Owner access on your Claude organisation. The plugin itself runs on any paid plan.
Salesforce can talk to it at all
Hosted MCP servers are off by default and an administrator turns them on in Setup. That means an External Client App, the right scopes, and the consumer key and secret handed to whoever configures the Claude side. This is the part that quietly eats an afternoon.
Three packages
Hours multiplied by fifty, and the scope written before the work.
Each one is fixed scope, fixed price and has an end date. The difference between them is not speed, it is how much of the org gets read before people start asking it questions.
Switched on, in one org
Prerequisites checked against your org and your Claude plan, the External Client App created, hosted MCP servers activated, the connector configured on the Claude side, and one pilot group of up to ten people signed in and working. You get the runbook that describes exactly what was changed, and one handover session.
Switched on, and bounded
Everything above, plus the part most setups skip: a permission and field level security review on every object Claude will read and write, written down per profile, so you know what each seller exposes before they ask their first question. Then rollout past the pilot, an approval policy in writing, and a usage review two weeks later against what people actually asked for.
Fitted to how you sell
Everything above, plus your own skills and prompts built around your stages, your fields and your definitions, so answers match the way your pipeline is actually run. Includes the data hygiene those answers depend on, separate scopes per team where sales and service should not see the same records, monitoring, and thirty days of support after handover.
Not included, in any package: the Claude and Salesforce licences themselves, which you buy directly; beta approval, which Salesforce grants; and anything with its own scope, such as an implementation, a migration or rebuilding automation. Those are priced separately at the same rate.
What changes on the Monday after
The queue goes, the people stay.
The work that stops waiting is the routine half: pulling a list, researching an account, preparing for a call, writing the note, updating a field on a record somebody already has permission to edit. That work used to be a message to an administrator or to an outside contractor, and it used to come back the next day. In chat it comes back in a minute, and the person approves the write before it lands.
The work that does not change: anything touching the data model, automation, permissions, integrations or anything with a licence attached. An assistant that reads your records cannot redesign them, and a model that can write to Salesforce under a seller’s permissions makes the state of those permissions matter more, not less.
That is the honest trade, and it is why the permission review sits in the middle package rather than in a footnote.
Asked often enough to answer here
Questions
How much does it cost to connect Claude to Salesforce?
€800 to switch it on in one org with a pilot group, €2,000 to add a permission review and roll it out past the pilot, and €3,500 with your own skills built around your pipeline. Those are 16, 40 and 70 hours at €50 an hour, which is the same rate as everything else on this site.
Can you get us into the beta?
No. Salesforce approves it through the AgentExchange listing, and that decision is theirs. What I can do is have everything else ready so the day approval lands is the day it works, and tell you honestly if your edition makes approval unlikely.
Can Claude see everything in our Salesforce?
It acts as each person’s own Salesforce account, so it is bound by their object permissions, their field level security and the sharing rules. That is also the reason the middle package exists: in most orgs nobody has read those permissions in years, and the honest answer to what a seller can expose is unknown until somebody checks.
Will it change records without asking?
By default it asks the person to approve each proposed change before it is written. That default is worth keeping, and the written approval policy in the middle package says who may change it and for which objects.
Do we need Agentforce for this?
No. This is the plugin on the Claude side plus hosted MCP servers on the Salesforce side. Agentforce is a different product with its own licence, and if agents are what you actually want, the readiness assessment is the cheaper first step.
What does this replace?
The queue, not the people. Pipeline questions, account research, call prep and routine record updates stop being a request to somebody and become a question in chat. Everything that changes the data model, automation, permissions or an integration still needs a person who knows Salesforce, and that is the work I am usually called for.
Serhii Skrypnyk · Senior Salesforce Administrator · 7 Salesforce certifications · on the platform since 2018. Reviewed 19 October 2026.