What Already Works With Any AI on Salesforce vs. What's Actually Claude-Exclusive

# What Already Works With Any AI on Salesforce vs. What's Actually Claude-Exclusive

Our last post covered "Claudeforce" — the Aug 26, 2026 Salesforce + Anthropic announcement. The most common follow-up question is worth its own answer: how much of that is actually new capability, versus Salesforce-and-Claude packaging around access that already existed for any AI?

The honest answer, backed directly by Salesforce's own developer documentation: most of the raw "AI talks to my Salesforce data" capability already works today with basically any AI assistant that speaks MCP — Claude, ChatGPT, Gemini, Cursor, Perplexity, a custom agent, doesn't matter. What Claudeforce actually adds is narrower, but real: default-model status inside Salesforce's own products, a curated skill library, and a depth of commercial backing no other AI vendor currently has with Salesforce.

Here's the layer-by-layer breakdown.

Tier 1: Already open to any AI today

Salesforce's official Hosted MCP Server reached general availability in April 2026 — months before Claudeforce existed. Salesforce's own setup documentation names the compatible clients explicitly: "Claude, ChatGPT, Cursor, or a custom agent." It exposes:

• Full SObject CRUD — create, read, update, delete records, respecting field-level security and sharing rules
• Custom Apex Invocable Actions, `@AuraEnabled` methods, Apex REST methods, Flows, and Named Queries
• Data 360 SQL queries and Tableau analytics
• Prompt Builder templates, callable directly from any MCP client

Authentication is per-user OAuth 2.0 with PKCE, so an agent — whichever AI is driving it — only ever acts within the requesting user's own permissions. Setup takes roughly 30 minutes on Enterprise Edition or above: enable it in Setup, create an External Client App, connect any MCP client, done.

Salesforce's official DX MCP Server is a separate, older tool — developer preview since May 30, 2025 — aimed at building and maintaining Salesforce orgs rather than working sales data. It ships 60+ tools across metadata deploy/retrieve, Lightning Web Component and Aura tooling, Apex static analysis (ApexGuru), DevOps workflows, and SOQL. It's documented to work with Cursor, Claude Desktop, VS Code with Cline, and Agentforce for Developers — again, client-agnostic by design.

Community-built MCP servers fill the gap further. Open-source projects like `advancedcommunities/salesforce-mcp-server` (MIT-licensed, 41 tools, built on the Salesforce CLI) explicitly target "any AI app which supports the MCP protocol" — Claude Desktop, Claude Code, Cursor, VS Code with Copilot, Codex CLI. Similar adapters exist for Gemini (via third-party platforms like AnythingMCP, offering roughly 14 tools for SOQL queries, record CRUD, and pipeline stage changes through Gemini CLI or Vertex AI).

One security research piece put the client-agnostic reality plainly: the MCP server "exposes Salesforce to *external* agents (Claude, Cursor, ChatGPT, custom) over the open Model Context Protocol," with "no vendor differentiation" — every client gets the same unfiltered access to whatever the authorizing user can see, including PII in Accounts, Contacts, and Cases, and sensitive data hiding in custom fields. That's a genuine risk worth flagging on its own, independent of which AI is on the other end.

Tier 2: ChatGPT gets its own bespoke deal — separately, and earlier

Before Claudeforce existed, Salesforce had already built a dedicated integration for a different AI vendor. In December 2025, Salesforce launched Agentforce Sales as a native ChatGPT app — sellers can query leads, update opportunities, and delegate prospecting tasks from inside a ChatGPT conversation.

Salesforce was explicit about why. Kris Billmaier, GM of Agentforce Sales, put it this way: "The thing that I worry about... was homegrown MCP servers from customers just spitting out data to OpenAI around the trust boundary." Rather than let customers wire up their own unofficial connectors, Salesforce built an official, governed channel instead.

Architecturally, this isn't the same shape as the raw MCP tools in Tier 1. Queries route through Agentforce's own Trust Layer, where Salesforce's AI scores leads and applies business logic *before* handing a formatted result back to ChatGPT as an interactive "canvas." ChatGPT users get a curated, Salesforce-built experience — not direct access to the four raw MCP verbs a generic client would use.

Tier 3: What's actually Claude-exclusive

This is the part that doesn't exist anywhere else, as of this writing:

Default reasoning model status inside Agentforce itself. Claude isn't just a client calling Salesforce's MCP server from the outside — it's now the default engine behind Agentforce's Atlas Reasoning Engine, and it powers Agentforce Vibes and Agentforce Coworker out of the box. ChatGPT and Gemini can call the same Hosted MCP Server as external agents; neither is running as the reasoning layer inside Salesforce's own agent products.

A curated skill library nobody else gets. The "Salesforce in Claude" plugin ships 37 prebuilt sales skills — meeting prep, deal health scoring, pipeline reviews, stakeholder mapping, call summarization — specifically engineered for Claude's plugin interface. A ChatGPT or Gemini user hitting the same generic Hosted MCP Server gets the raw Discover/Describe/Dispatch tools and has to write their own prompts from scratch. There's no equivalent prebuilt skill pack shipped for those platforms.

A different UX paradigm than ChatGPT's canvas. ChatGPT's Agentforce Sales app is a Salesforce-built, scored, canvas-rendered experience, entirely controlled by Salesforce's own Trust Layer. Claude's integration leans into "the UI is the AI" — a conversational layer that can generate custom, styled dashboards on the fly rather than rendering a fixed canvas Salesforce designed in advance.

Slack, specifically. Claude becomes the default model for Slack itself — Slackbot, Claude Tag — with governed Salesforce actions reachable from a Slack thread. There's no announced equivalent default-model arrangement for ChatGPT or Gemini inside Slack.

Commercial depth that buys engineering priority. Salesforce has reportedly committed roughly $300 million a year in Anthropic token spend on top of an existing $300 million equity stake. That kind of commitment is why Claude gets embedded as a default inside core Agentforce products, while other models remain reachable, but arms-length, through the open protocol.

The comparison, side by side

| Access path | Who can use it today | What you actually get | Salesforce-official? | |---|---|---|---| | Hosted MCP Server (GA April 2026) | Any MCP client — Claude, ChatGPT, Cursor, Perplexity, custom agents | Raw record CRUD, Apex actions, Flows, Named Queries, Data 360, Tableau | Yes — first-party | | DX MCP Server (dev preview, May 2025) | Cursor, Claude Desktop, VS Code + Cline, Agentforce for Developers | 60+ dev tools: metadata, LWC/Aura, Apex analysis, DevOps, SOQL | Yes — first-party, dev-focused | | Community MCP servers | Any MCP client | Similar or richer CRUD/Apex/SOQL access, sometimes 40+ tools | No — unofficial, outside governance and audit | | Agentforce Sales in ChatGPT (Dec 2025) | ChatGPT only | A bespoke, Salesforce-built "canvas" scored by Agentforce's Trust Layer | Yes — first-party, OpenAI-specific | | Gemini via third-party adapters | Gemini CLI / Vertex AI | ~14 tools: SOQL, record CRUD, reporting | No — third-party infrastructure dependency | | "Salesforce in Claude" (Claudeforce) | Claude only | Same Hosted MCP Server, plus 37 prebuilt skills, default-model status inside Agentforce, default Slack model | Yes — first-party, Claude-exclusive packaging |

What this means practically

If your goal is simply "let an AI read and write my Salesforce data through governed APIs," you don't need to wait for Claudeforce's open beta, and you don't need Claude specifically. The official Hosted MCP Server already does that today, with any MCP-compatible AI, and has since April 2026. What Claudeforce buys you on top of that is Salesforce's own curation and default-model treatment for one specific AI vendor — genuinely useful if you want the polished, pre-built experience, but not a new door that was previously locked.

It's also worth being honest about the flip side of "any AI can connect": the same openness that lets ChatGPT, Gemini, or a homegrown script talk to your CRM data is a real governance surface. Every unfiltered SOQL query, every custom field an admin forgot was sensitive, flows to whichever AI client is authorized — audit your permission profiles and field-level security before connecting anything, regardless of which AI you're connecting.

FAQ

Do I need Claude to use MCP with Salesforce? No. Salesforce's own Hosted MCP Server documentation lists Claude, ChatGPT, Cursor, and custom agents as compatible clients. The protocol is open; Claude is one option among several, not a requirement.

Is the Claudeforce announcement mostly marketing, then? Not entirely — the underlying data access genuinely isn't new or exclusive, but the default-model embedding inside Agentforce's own products, the 37 prebuilt skills, and the Slack default-model status are real and don't exist for any other AI vendor today.

Can I connect ChatGPT to my Salesforce org right now? Yes, two ways: Salesforce's own Agentforce Sales app (native ChatGPT integration, launched Dec 2025) for a curated sales experience, or by connecting any MCP client — including ChatGPT — to the general Hosted MCP Server for raw data access.

What about Gemini? Gemini doesn't have a Salesforce-built native app the way ChatGPT does, but it can reach Salesforce data through the same open Hosted MCP Server, or through third-party adapters like AnythingMCP that bridge Gemini CLI or Vertex AI to Salesforce's APIs.

What should I check before connecting any AI to Salesforce? Audit permission profiles and field-level security first — the MCP server passes through whatever the authorizing user can already see, unfiltered, to whichever AI is on the other end. Community/unofficial MCP servers add an extra layer of risk since they sit outside Salesforce's own audit trail entirely.

---

*Sources: [Salesforce Hosted MCP Servers overview](https://developer.salesforce.com/docs/platform/hosted-mcp-servers/guide/hosted-mcp-servers-overview.html), [Salesforce DX MCP Server docs](https://developer.salesforce.com/docs/atlas.en-us.sfdx_dev.meta/sfdx_dev/sfdx_dev_mcp.htm), [Salesforce Developers blog on DX MCP](https://developer.salesforce.com/blogs/2025/06/level-up-your-developer-tools-with-salesforce-dx-mcp), [SalesforceDevops.net on Agentforce in ChatGPT](https://salesforcedevops.net/index.php/2025/12/22/salesforce-launches-agentforce-in-chatgpt-to-head-off-homegrown-mcp-servers/), [usecarly.com on the Hosted MCP Server](https://www.usecarly.com/blog/salesforce-mcp/), [Strac's Salesforce MCP security analysis](https://www.strac.io/blog/salesforce-mcp-server), [advancedcommunities/salesforce-mcp-server on GitHub](https://github.com/advancedcommunities/salesforce-mcp-server), and [AnythingMCP's Gemini connection guide](https://anythingmcp.com/guides/connect-salesforce-to-gemini).*