Salesforce 通过 MCP 服务器和可复用代理技能扩展 Headless 360

Salesforce is pushing its platform further toward agent-first development. In an August 25 announcement, the company described an expanded Headless 360: hosted Model Context Protocol (MCP) servers, more Data 360 and Slack integrations, reusable Agent Skills and Plugins, a headless experience layer, and Salesforce Multi-framework for modern React applications. The pitch is simple: organizations should be able to reuse their existing identity, permissions, metadata, workflows, and business logic wherever an authorized agent works, instead of rebuilding an integration for every new AI surface. Salesforce's announcement

For coding teams, the most concrete piece is the Headless 360 MCP Server. Salesforce's developer reference says the beta exposes four stable tools backed by a growing library of operations, rather than presenting thousands of individual tools to an agent. The server can cover tasks such as querying or updating records, reading and deploying Apex triggers, managing users, and building event-driven integrations. The important design choice is progressive discovery: the agent first searches for a capability with Discover, reads its contract with Describe, and only then executes it through Dispatch or the read-only Dispatch variant. The reference documentation

That shape matters beyond Salesforce. A smaller tool surface is easier to explain and review, while runtime discovery lets the operation library grow without forcing every client to hard-code a new tool list. Salesforce also says the server requires an External Client App with the mcp_api scope and an OAuth-configured MCP client; the beta is not an anonymous plug-in.

The broader developer story is the combination of MCP servers with portable skills. Salesforce's developer blog describes skills as reusable, validated instructions for narrow tasks, and plugins as packages that bundle skills, MCP servers, and utilities for a particular agent platform. That creates a practical division of labor: the protocol connects the agent to live systems, while a skill explains how to perform a repeatable development task. The developer walkthrough

The takeaway for builders is not “give an agent access to everything.” It is to expose one governed capability at a time, make discovery and authorization explicit, and keep a human review boundary around writes, deployments, and identity changes. Headless tooling makes the agent's reach broader; the trust layer still determines whether that reach is safe.