Salesforce’s Headless 360 MCP server is live in beta, and it changes the threat model for AI agents in enterprise orgs. The server allows an AI agent to act exactly as the user who authorized the connection, inheriting every permission that user holds. This isn’t a sandboxed agent with a fixed scope; it’s a runtime explorer that discovers, describes, and dispatches operations based on the connected user’s actual access rights.
The Four-Tool Discovery Loop
The integration relies on four core tools: discover, describe, dispatch, and dispatch_readonly. The 'discover' tool takes a natural-language query and returns ranked candidate operations using a hybrid semantic and keyword search. For example, querying for 'query data in a grid worksheet' returns Setup Operation Recipes (SORs) with relevance scores and step counts. The 'describe' tool then unpacks these SORs into detailed multi-step workflows, including preconditions and agent guidance, before 'dispatch' executes the specific API call.
Permission Sprawl Is Now Agent Reach
The critical insight is that most Salesforce orgs are riddled with legacy permissions. Users often retain 'View All Data' or 'Modify All Data' from past projects or migrations because revoking them is risky. In the UI, this sprawl is hidden by list views and obscure setup pages. Headless 360 removes this friction. An agent doesn’t care about UI complexity; it simply asks what it can do. If a support lead has 'View All Data', the agent can now query every Case in the org, a capability that sat idle for years because no human ever bothered to do it.
The Ambiguity of Write Operations
Reads are straightforward, but writes introduce policy risks. An agent asked to share a record might choose to widen the org-wide default sharing setting rather than creating a specific share. Both actions satisfy the request and are permitted by the user’s profile, but one changes the org for everyone. Salesforce’s documentation notes that annotations are 'hints, not enforcement.' Client-side gates often collapse into 'require approval for every write,' which is too blunt to distinguish between a minor record update and a global configuration change.
Key Takeaways
- Headless 360 agents inherit the exact permissions of the authorizing user, with no separate identity.
- 'Discover' and 'describe' allow agents to find operations at runtime, meaning the available tool surface grows with every Salesforce release without client updates.
- Legacy permissions like 'View All Data' become active attack surfaces when agents bypass UI constraints.
- The server is off by default; admins must check Setup > MCP Servers to see if it is active.
- Refresh tokens for the External Client App remain valid indefinitely by default, creating long-lived access vectors.
The Bottom Line
Headless 360 is not just a convenience layer; it is a permission amplifier that exposes the neglected hygiene of enterprise Salesforce orgs. Until admins actively restrict External Client App access and rotate refresh tokens, the agent's reach is effectively the org's weakest link.