What Claudeforce actually is

Salesforce and Anthropic announced Claudeforce on August 26, 2026: an expanded partnership bringing Claude’s reasoning together with Salesforce’s data, workflows, and governance. If you’ve only seen the headline, or a secondhand summary of it, here’s what it actually says, and what it doesn’t yet.

---

The two halves

Claudeforce isn’t one product, it’s two integrations running in opposite directions.

Salesforce in Claude. A plugin you connect to Claude that gives it access to your Salesforce data, workflows, and business logic through a secure connection. Sellers can ask Claude to prep for a meeting, review a deal’s health, or update a pipeline record, and Claude reasons over live CRM data to do it, without opening Salesforce itself. It ships with 37 prebuilt sales skills, though only a handful are named anywhere public so far: meeting prep, deal health review, pipeline review, plus composing emails and updating records.

Claude in Salesforce. Claude runs as the reasoning model inside Salesforce’s own tools: the Atlas Reasoning Engine, Agentforce Vibes, and Agentforce Coworker, where it’s the default model, and Agent Builder, where it’s a selectable option alongside others. For regulated industries, this runs through Amazon Bedrock inside Salesforce’s own trust boundary rather than a third-party API call, which is the detail that matters if your org is in financial services or healthcare.

---

How the connection actually works

The plumbing underneath is called Headless 360, a Salesforce architecture that exposes the platform’s data and actions through the Model Context Protocol (MCP), the same open standard covered elsewhere on this site. Claude connects to your org through a hosted MCP server rather than a custom integration.

The part worth understanding, not just trusting: every action Claude takes still runs through your org’s existing permission model. Profile and permission set restrictions, field-level security, sharing rules, validation rules, and transaction security policies all still apply. Claude doesn’t get broader access than the user connecting it already has, and it can’t bypass a validation rule any more than a human clicking through the UI could.

---

What’s confirmed and what isn’t

Confirmed, from the official announcement: the two-directional integration exists, 37 skills ship with Salesforce in Claude, Claude is the default model for Agentforce Vibes and Coworker, and Salesforce in Claude is available to select pilot customers now, with open beta expected in September 2026.

Not public yet: the full list of 37 skills (three are named, the rest aren’t), pricing (token cost is tied to a separate Anthropic contract and hasn’t been quantified anywhere confirmable), and which Salesforce editions or license tiers get access by default.

Anything you read claiming a specific dollar cost, a full skills list, or a guaranteed general-availability date is going further than the official sources currently support.

---

Why this isn’t just Agentforce with a new name

Claude has been available inside Agentforce since late 2025. What Claudeforce changes is mostly administrative rather than algorithmic: instead of individual teams standing up their own connections, one admin connects the org once, centrally, and access is scoped from there to each user’s existing permissions. See the comparison guide for the fuller breakdown of what’s actually new versus rebranded.

---

When it breaks

Someone tells you Claudeforce is generally available. As of this writing it isn’t. Pilot customers and an open beta expected in September 2026 are not the same as a shipped, supported product every org can turn on. Confirm current status before planning around it.

A secondhand summary states a specific productivity number as if it’s Claudeforce’s own result. The widely cited “8.1 million hours” figure is Salesforce’s internal Slackbot usage data, a related product, not a measured result of Claudeforce itself, and it’s a vendor-reported number nobody outside Salesforce can audit. Treat vendor stats about a different product as context, not proof.

You can’t find documentation for a specific skill someone mentioned. Only three of the 37 skills are named in public sources as of this writing. If a skill you’ve heard about isn’t one of those three, you’re working from an unconfirmed claim, not a documented feature.

---

What to try next

If your org is considering a pilot, the admin checklist guide covers what to verify before connecting anything to Salesforce data, which is the part that actually determines whether a pilot goes well.