Salesforce made a pitch at Dreamforce last week that it has never quite made before: you do not need to learn Salesforce to build on Salesforce. The AIforce Headless Toolkit ships with 60+ MCP tools, 4,000+ APIs, 220+ CLI commands, and 30+ prebuilt coding skills for Claude Code, Cursor, and Windsurf. A backend developer can now wire Salesforce data and workflows into an application without opening Setup, without memorizing SOQL object hierarchies, and without a Salesforce consultant on retainer. The interesting question is how much of that is available today and how much ships next quarter.
What the Headless Toolkit Actually Contains #
The marketing layer calls it AIforce. The developer layer is the Headless Toolkit, and it has four production-facing components:
- Salesforce Hosted MCP Servers — GA since April 2026. Expose org data, flows, Apex actions, and SOQL queries to any MCP client. Requires Enterprise Edition or higher.
- Marketing Cloud Engagement MCP Server — GA since June 2026. Manages journeys, data extensions, and automations.
- Salesforce DX MCP Server (
@salesforce/mcp) — In beta. Local, CLI-based server distributed as an npm package. Useful for development workflows without a hosted org dependency. - Agentforce Native MCP Client — In beta since January 2026. Lets Agentforce agents call external MCP servers as tools.
The Claudeforce plug-in — 37 prebuilt sales skills in Claude and 40+ additional skills for Claude Code — entered open beta on AgentExchange on September 14, one day before the Dreamforce keynote. Deloitte, GitLab, and Legora piloted it. Agentforce Coworker (the Lightning UI assistant) is generally available with no migration required; 100,000 users activated it in the first 35 days.
Connecting Claude Code to Salesforce #
The Hosted MCP Server path is the most practical starting point. Here is the full connection flow:
Salesforce configuration:
- Go to Setup → MCP Servers. Enable
sobject-reads. Start read-only. - Go to Setup → External Client App Manager. Create a new application.
- Set the OAuth callback URL to
https://claude.ai/api/mcp/auth_callback. Add scopes:mcp_api,refresh_token,offline_access. - Enable PKCE. Enable “Require Secret for Refresh Token Flow.”
- Copy the Consumer Key and Consumer Secret.
Claude configuration:
- In Claude.ai Settings (or Claude Desktop’s Customize menu), select “Add custom connector.”
- Enter the MCP server URL:
https://api.salesforce.com/platform/mcp/v1/platform/sobject-reads. Paste the Consumer Key as the OAuth Client ID. - Click Connect. Authenticate through Salesforce OAuth. Done.
After connecting, Claude can run natural-language SOQL queries, describe sObject schemas, retrieve related records, and surface aggregated pipeline data. The @salesforce/mcp npm package does the same locally with the DX MCP Server if you prefer to avoid cloud OAuth setup during development.
The Permission Model Is Not Optional Reading #
Every AIforce request runs with the connecting user’s existing permissions. That sentence sounds like a security feature, and it is — until it is not.
If your Salesforce org has accumulated permission debt over the years (and most enterprise orgs have), AIforce inherits all of it. Sharing rules that were supposed to be tightened two releases ago, field-level security left open for convenience, over-broad profiles from a migration that nobody revisited — an agent asking “show me all enterprise accounts with overdue invoices above $500K” will return exactly what the connecting user can technically access, not what they should be able to see.
This is not a bug in Salesforce’s design. It is an acceleration of existing exposure. The practical response: before you connect any agent, run a permission audit. Create a dedicated integration user scoped to the minimum necessary objects and fields. Test in a sandbox. The Hosted MCP Server’s read-only sobject-reads starting point exists precisely because Salesforce knows how this goes.
Koa: The Purpose-Built CRM Reasoning Model #
The component that got the least Dreamforce coverage may matter most for developers building Agentforce agents. Koa is Salesforce’s first purpose-built CRM reasoning model, post-trained on NVIDIA Nemotron 3 Super using synthetic data drawn from 27 years of CRM deployment scenarios across 14 industries. Salesforce claims it matches or beats leading general-purpose models on CRM tasks with three times fewer errors.
Those errors matter in enterprise context. A general-purpose LLM guessing at opportunity stage logic or misinterpreting a Salesforce junction object relationship makes mistakes that are hard to detect and expensive to fix. Koa is currently in pilot with Formula 1, UChicago Medicine, and Xero, among others. General availability is targeted for winter 2026 in U.S. regions.
What to Do This Week #
The Hosted MCP Servers are GA and free in Developer Edition orgs. That makes the entry cost zero.
- Existing org on Enterprise Edition or higher: Enable
sobject-reads, set up the External Client App, connect Claude. Start with read-only queries before touching write operations. - Evaluating whether to build on Salesforce: Spin up a free Developer Edition org. You get Hosted MCP Servers, Agentforce Vibes IDE, and Claude Sonnet 4.5 as the default coding model at no cost.
- Building Agentforce agents now: Bookmark Koa’s GA timeline (Winter 2026). General-purpose LLMs work for prototyping; for production agents doing complex CRM reasoning, the purpose-built model will matter.
The pitch Salesforce made at Dreamforce is more grounded than most enterprise AI announcements. The toolkit is real, the MCP Servers are in production, and the connection path is documented. The catch — Enterprise Edition gating, permission debt risk, Koa still in pilot — is also real. Both things are true at the same time, which puts this in the “worth evaluating seriously” category rather than the “wait and see” one.