Salesforce pipeline lands inside Claude, in beta

Ask Claude to send your forecast and the action still runs through Salesforce, so whatever validation rules sit on your records keep applying. Salesforce in Claude is now in beta, bringing accounts, opportunities and pipeline into the chat with 37 pre-built sales skills, announced by Anthropic on X.

At a glance

• Anthropic puts the plugin's scope at chat and Claude Cowork, on web and desktop; an admin connects it, and the announcement names four of the 37 skills and leaves the rest unnamed.
• Connecting the two uses an OAuth 2.0 client credentials flow, according to Anthropic's documentation: the credential belongs to the agent rather than a person, and read-only scoping is the recommended starting point.
• When Claudeforce was announced in late August, CNBC reported the plugin was going to a select group of pilot customers, with more prebuilt skills promised later in the year; no price is mentioned.

If you did not follow the earlier episodes: in January Anthropic launched interactive Claude apps for Slack, Canva, Figma, Box and Clay, and TechCrunch noted then that a Salesforce implementation was expected soon. That system is built on MCP, the open standard Anthropic introduced in 2024, which according to TechCrunch added app support in November. The Salesforce piece was announced in late August as Claudeforce; CNBC reported it came alongside Salesforce's quarterly results.

Anthropic names four of the 37 skills

The four in the announcement are the ones a rep does every week: prep a call, review a deal, create a pipeline dashboard, send your forecast. All of it happens without leaving the conversation, and the data behind it is your accounts, opportunities and pipeline rather than a snapshot pasted into a prompt. The remaining 33 are not listed.

Scope is narrower than "Claude everywhere": Anthropic's product page limits it to chat and Claude Cowork, web and desktop. An admin has to connect the plugin, so this is not something an individual seat turns on by itself. Timing held, too. At the August announcement, CNBC reported Claudeforce was going out to a select group of pilot customers with a preview capacity launch promised for the following month, and the open beta landed in September as penciled in.

The credential belongs to the agent, not to a person

The interesting part is the wiring. According to Anthropic's documentation, the connection uses the OAuth 2.0 client credentials flow, lives inside an Access bundle, and runs as a dedicated integration user set as the connected app's run-as user. Nothing is tied to whoever happens to be typing. The docs recommend starting with read-only permission scoping, which means the first thing your org grants is the ability to look, not to write.

The key itself never reaches the model. Anthropic's documentation describes an Agent Proxy that injects the Salesforce credential at the network boundary, so neither the model nor the sandbox holds it. Picture a courier who carries the envelope to the door but never carries the key: the door is opened from the outside, at the last moment, by something the courier cannot reach or read.

Salesforce calls the plumbing AIforce

On the Salesforce side, the enabling layer has its own name. Salesforce describes AIforce as its trusted enterprise harness that brings business data and workflows to any agent through MCP servers, APIs and CLI tools, "without complicated and costly integrations." The same harness is what any other agent would plug into.

Permissions got explicit attention at the announcement. CNBC reported that Dario Amodei said Anthropic put a huge amount of effort into managing permissions for Claudeforce, and that the two companies are building what they call Enterprise Frontier Safeguards, meant to keep Salesforce customer data private and keep Anthropic's models from running out of control.

Amodei also said, per CNBC, that Claude had long been accelerating Anthropic's research teams, and that Claudeforce is the first time the company has been able to accelerate its own go-to-market work inside Claude. CNBC also reported Anthropic told investors its annualized revenue run rate reached 65 billion dollars at the end of July, seven times higher than a year earlier.

What the announcement does not give you is the rest of the list: 33 of the 37 skills go unnamed, and there is no word on pricing, seat requirements or whether coverage extends past chat and Cowork. In our view the read-only-first guidance in the docs is the honest part of this launch, because a chat box that can push a forecast is exactly where a scoping mistake stops being theoretical.

When the other 33 skills land

No date has been given for the rest. CNBC reported in August that additional prebuilt skills were promised later in the year, which is the next checkpoint worth watching, along with whether the surface widens beyond chat and Claude Cowork on web and desktop. The other open question is how many orgs move past read-only scoping once beta feedback comes back, and what Enterprise Frontier Safeguards turns out to cover in practice.