Agentforce
Salesforce in Claude: what the 37 skills do, and how to pilot the beta
The Claudeforce plugin that puts a governed Salesforce org inside a Claude conversation goes to open beta this month, with a live demo at Dreamforce. This is the practitioner's guide: the five jobs the 37 skills cover, the hosted MCP server and per-user OAuth underneath, what a write from Claude triggers in your org, the zero-data-retention story, and a four-week pilot that ends in a decision.
VentureBeat’s headline was that Salesforce had put its entire CRM inside Claude and said you’d never need its app again. Good headline. Poor specification.
What ships is a plugin called Salesforce in Claude. Salesforce lists it as available to select pilot customers now, with an open beta expected in September 2026, gets a live demo at Dreamforce on September 15 to 17, and carries 37 prebuilt sales skills the two companies say they built together.
I covered what Claudeforce changed architecturally the week it was announced. This post is about the plugin alone, because it is the one a sales leader will ask you to turn on this month. For the short answers on cost, timing and models, see the Claudeforce questions page.
The 37 skills, sorted by the job they do
Salesforce’s product page does not list all 37 skills. It groups them into five jobs, and the grouping is more useful than a list because it tells you which reps to pilot with.
The daily briefing. Ask what needs attention today and get an action plan from the live pipeline. Reps will use this most and it will produce the least measurable value, because a pipeline summary was already a report. Its worth is in what it links to next.
Deal strategy. Stakeholder mapping, close planning and objection handling against a specific opportunity. The page’s phrasing is that you ask about any deal and get more than a health score. Reasoning over the whole account, activity history, contacts, open cases and Slack context if connected, does something a dashboard cannot.
Account growth. Account planning, lead triage and customer-health assessment. The framing is that the plugin finds the growth already sitting in your accounts, which is the promise the Agentforce sales agents make from the other direction. The difference is that here the seller is in the loop on every step instead of reviewing an agent’s overnight output.
Meeting preparation. Call prep, conversation summaries and scheduling, with the follow-up drafted. The press release names meeting prep, deal health review and pipeline review as its three flagship examples. Meeting prep is the one to measure in a pilot because it has a before and after a rep can feel.
Data management. Win-loss review, activity logging and Salesforce hygiene checks. The line that matters for admins: when it updates a record, it changes only the field it said it would. Activity logging from a chat is the skill most likely to improve your data quality, and the one most likely to generate tickets when a rep discovers what a validation rule is.
Every skill is a sales skill. Salesforce says service, marketing and commerce follow, with additional skills beginning to launch in late 2026. If you want a service desk in Claude, this is not yet that.
How it connects, and as whom
The plugin does not talk to your database. It talks to Salesforce’s hosted MCP server, the same Headless 360 MCP server I described for driving Data 360 from Claude. Apex Hours’ write-up lays out the four tools the server exposes: Discover, a semantic search across Salesforce operations; Describe, which returns one operation’s specification; Dispatch, which invokes it; and a read-only Dispatch that permits GETs only.
Claude decides which operation fits the seller’s request by searching, reading the spec and calling it. Which is the MCP pattern in general, and the reason new skills can be added without a new integration each time.
The identity model is the part to brief your security team on, because two routes exist with two different answers to “who did that”.
The hosted MCP route, which Salesforce in Claude uses, authenticates each user through OAuth against an External Client App with the mcp_api scope, so every call runs as the person asking. The Claude Tag connection inside Slack uses a client-credentials flow with a dedicated integration user, so actions there run as that user rather than the human who typed.
Both are legitimate. Only the first gives you a per-seller audit trail out of the box. Apex Hours also notes the plugin requires API version 67.0 or later, which any org on Summer ‘26 or the Winter ‘27 release already has.
Setup is one org-level connection. Salesforce’s line is no per-user setup, no new permissions model to build, no re-auditing account by account. An admin connects the plugin and it works for the team. True, and it is the reason to read the next section carefully, because “works for the team” means “works with whatever the team’s permissions already are”.
What a write from Claude triggers
The governance claim is precise: every answer and every action runs through your existing Salesforce permissions and business rules, and Claude sees only what a given user is authorised to see. In practice a write from a Claude conversation is an ordinary API write as that user.
Validation rules fire. Record-triggered Flows fire. Apex triggers fire. Sharing rules and field-level security apply on the read side, so a rep who cannot see a field in Lightning cannot have Claude summarise it either.
Three things follow for whoever runs the org.
Your permission hygiene becomes visible in a new place. A profile that has been over-permissioned for years, tolerated because nobody used the extra access, is now reachable by a model that will use whatever it is given when asked. The agent user permissions discipline I recommend for Agentforce applies to human users of this plugin for the same reason.
Your automation gets a new caller. A Flow that assumes edits arrive from the page layout, say one that reads a field the layout always populates, will now meet edits that carry only the field Claude said it would change. Correct behaviour on the plugin’s side. It will expose fragile automation on yours.
Test the plugin against your real validation rules in a preview sandbox before the beta reaches production users, the same way you test an agent.
Outbound communication has an explicit control. The product page describes a setting for whether Claude checks with you before it emails anyone outside your company, or just sends it once you are comfortable. Start with the check. Move to auto-send per rep and per skill after the pilot, and never for a skill that drafts pricing.
Privacy: zero data retention, and the safeguards arriving in the fall
Two privacy statements are in play and they are easy to conflate.
Salesforce’s statement is that Claude runs through Amazon Bedrock inside the Salesforce trust boundary, which sits alongside the Einstein Trust Layer rather than replacing it.
Anthropic’s statement is that the plugin runs with zero data retention on Sonnet, Opus and Haiku, and that Enterprise Frontier Safeguards, rolling out in phases from the fall, keep the activity data used for misuse detection in cloud storage the customer controls, under the customer’s own keys and audit logging, with flagged signals routed to the customer rather than read by Anthropic.
For a regulated seller, the combination is stronger than either alone. The caveat comes from Anthropic itself, as reported by The Register: zero data retention is a configuration customers must confirm is in effect, not an assumption. Put “verify the retention setting on the connection” on the pilot checklist, in writing, with a named owner.
Cost, availability, and what nobody can tell you yet
Pricing is not published. Salesforce’s release says packaging is subject to change, and the plugin’s consumption is metered on the model side in a way nobody outside the pilot can quote. Apex Hours recommends a pilot before any org-wide commitment, and I agree, because the token profile of a seller who reviews forty opportunities every morning is unknown until you measure it.
The shape of an Agentforce bill, conversations or Flex Credits, does not obviously apply to a plugin that runs inside a Claude seat. Ask which contract it lands on.
Availability is pilot now, open beta in September, more skills late in the year. Beta means the skill set will change under you and stability is not promised.
It also means the beta overlaps your Winter ‘27 production weekend for most orgs. Sequence them. Debugging a validation-rule failure while both the platform and the plugin are new is a poor week.
Two things the announcement leaves alone. The default-model changes for Coworker, Vibes and Slack are live, but Claude is a default there, not the only option, and the model choice for your unattended agents is untouched. And Salesforce in Claude is a seller-in-the-loop surface. A person asks, reviews and confirms. Nobody should mistake it for a way to run an autonomous agent from Claude, so a team that wants unattended pipeline work should still be looking at Agentforce agents, with this plugin as the human’s window onto the same data.
A four-week pilot that ends in a decision
The failure mode of a September beta is a demo that impresses everyone and decides nothing. The pilot I would run:
Week one, scope and safety. Pick five reps across two segments, one heavy Slack user among them. Choose three skills to measure: meeting prep, activity logging, and one deal-strategy skill. Connect the plugin in a preview sandbox first, confirm the per-user OAuth route is the one in use, verify the zero-data-retention setting, and set outbound email to confirm before send.
Week two, automation contact. Have the five reps log activity and update opportunity fields through Claude only. Collect every validation-rule and Flow error. Each one is either a fragile automation to fix or a guardrail doing its job. Classify them, because the ratio tells you how much admin work a wider rollout costs.
Week three, value. Measure meeting-prep time per call before and during, and the count of opportunities with a logged next step. Pull the audit trail for the five users and check that every write is attributable to the person who asked for it. Ask the reps which skills they stopped using after the novelty faded. That list is more honest than a satisfaction survey.
Week four, the decision. Bring three numbers: model consumption per active rep per week, the count of automation fixes required, and the change in next-step coverage. Decide on a per-team rollout with the outbound-email control still on, or a hold until the service and marketing skills arrive. Either is a fine outcome. “We turned it on for everyone because Dreamforce was exciting” is not.
If your org is not yet in a state where a governed model can be pointed at it safely, with over-broad profiles and automation full of undocumented assumptions, the Agentforce readiness assessment covers the same prerequisites. The plugin and an agent make identical demands of your permission model.
Understanding the basics
What is Salesforce in Claude?
Salesforce in Claude is a plugin, part of the Claudeforce partnership announced on August 26, 2026, that lets a seller work their Salesforce data from inside a Claude conversation. It ships with 37 prebuilt sales skills covering daily briefings, deal strategy, account growth, meeting preparation and data management, and routes every read and write through the user’s existing Salesforce permissions.
When is Salesforce in Claude available?
It is available to select pilot customers now, with an open beta expected in September 2026 and a live demonstration at Dreamforce on September 15 to 17. Salesforce says additional skills, including service, marketing and commerce, will begin to launch in late 2026.
Does Salesforce in Claude bypass Salesforce security?
No. The plugin connects through Salesforce’s hosted MCP server using per-user OAuth, so Claude sees only what the requesting user can see and can only do what that user can do. Writes are ordinary API writes that trigger validation rules, Flows and Apex triggers as normal.
Is my data retained by Anthropic when using the plugin?
Salesforce states the plugin runs with zero data retention on Claude’s Sonnet, Opus and Haiku models, with inference served inside the Salesforce trust boundary. Anthropic’s Enterprise Frontier Safeguards, rolling out from the fall of 2026, keep misuse-detection data in storage the customer controls. Anthropic advises customers to verify the retention setting is in effect on their connection.
The plugin as a mirror
Salesforce in Claude will be judged on whether sellers keep using it after the first month, and the skills will decide that. Its second effect is the one admins will remember. It points a capable, permission-respecting model at an org and shows you exactly what that org allows. Over-broad profiles, automation that assumed a page layout, fields nobody validated: the plugin does not create any of these. It reveals them, one confirmed write at a time.
Treat the pilot as a governance audit with a sales benefit attached and you get both. Treat it as a demo and you get a demo.
If you want the pilot scoped, the permission model checked, and the automation-contact week run before your sellers see the plugin, talk to us. Or start with the Agentforce readiness assessment to see how far your org is from a safe pilot.
