Skip to main content

Managed Judgments through MCP

The hosted SchemaBounce MCP exposes judgment information through the product domain that owns it. There is no general-purpose judgment tool that bypasses product policy.

Reference resource​

Read schemabounce://reference/judgments to discover the current provider-neutral capability contract, data classes, fallbacks, payer, documentation path, and owning MCP domain.

Available reads​

  • sb_workspace with action=judgments_inspect returns the workspace's settings and effective capability state. The caller needs the normal settings read permission.
  • sb_schema with action=sensitive_findings returns workspace-scoped sensitive-column findings. The caller needs the normal schema read permission. When source_ref is supplied, the response also includes the reduced automatic-scan lifecycle and bounded next actions.

The findings action returns bands and provider-neutral provenance. It removes exact probability, engine model, and deciding-user identifiers from the standard MCP projection.

Lifecycle next actions may tell an authorized agent to wait for an in-flight scan, ask a person to test a source connection again, review findings in the console, or inspect judgment settings. This read cannot start a scan, change a finding, restore a route suggestion, create a transform, or bypass the console confirmation path.

Planned capabilities appear in the reference manifest but do not get callable actions until their product implementation exists.

Agent grants​

Agents receive no new default grants. A workspace owner must explicitly grant the relevant MCP domain to an agent. The authenticated user or service account must also have the underlying API permission.

Both checks matter:

  1. The MCP connection authorizes a workspace identity.
  2. The agent's explicit tool grant authorizes the domain action.

Draft boundary​

Future agent-authored configuration actions may offer preview and submit operations.

  • Preview is side-effect free.
  • Submit revalidates the target and creates a pending Inbox proposal.
  • Approval is performed by a person in the console.
  • MCP cannot approve, enable, or execute the proposal.

If the target changed after preview, submission fails as stale and requires a new preview.

What MCP cannot access​

The judgment reference and read actions do not expose prompts, raw submitted state, question text, provider credentials, exact probability distributions, conversation content, source event values, tool arguments, tool results, database row values, or file contents. Planned domain actions may accept the bounded feature inputs listed in the reference manifest, but they do not make that submitted state readable afterward.