Case study

DB MCP Safety Boundary

MCP safetyPublic-safeSafety boundary case

Safe multi-database MCP workflow pattern for agent-assisted schema and query analysis.

Problem

Agent database access must be constrained by read-only policy, schema visibility, query validation, and human review.

Approach

Use synthetic connection registry and query validation screens to show safety boundaries without exposing real schemas or logs.

Public Evidence

  • Connection registry concept with redacted metadata and explicit safety modes.
  • Query validation path that blocks unsafe operations before execution.
  • Human-review boundary for database access in agent-assisted workflows.

Redaction Boundary

  • No database URLs, credentials, real table names, customer schemas, or query logs.
  • Connection names and schemas remain abstracted into public-safe examples.

Verification

  • read-only boundary described at public-safe level
  • secret values excluded
  • synthetic examples only

Evidence focus: Read-only query validation and human-review boundaries for database-assisted workflows.

Screens

S-MCP-001Planned
connection listpermission modehealth stateredacted metadata
defaultdisconnectedpermission denied

secret-free connection metadata

Connection Registry

List redacted DB connections and safety mode.

Target
Reviewer evaluating infrastructure safety boundaries.
Entry
DB MCP safety page.
Signal
secret-free connection metadata
Boundary
No URLs, credentials, real schemas, customer tables, or production hostnames.
S-MCP-002Planned
query previewpolicy resultblocked reasonreview handoff
validblockedneeds review

agent safety gate

Query Validation

Validate query safety before execution.

Target
Reviewer evaluating agent safety gates.
Entry
DB MCP safety page screen evidence section.
Signal
agent safety gate
Boundary
Synthetic query examples only; no raw logs, private schema, or production data.