Hover a node · click to focus · ←/→ steps
Spec → allowlisted tools → model loop. Writes pause for human approval; MCP goes through an allowlist gateway.
When you need this
- Developers install MCP plugins ad hoc
- Security asks for MCP inventory
- Agents need prod read access
- Shadow MCP bypasses org controls
Prerequisites
- List of IDE/agent tools in use
- Security policy owner
- Preferred gateway (e.g. Runlayer-managed) selected
Steps
- 1
Inventory shadow MCP
Survey teams for MCP configs in Cursor, Claude Desktop, and internal agents. Document data access per server.
- 2
Define allowlist policy
Approved servers only; deny by default. Separate dev (synthetic data) from prod (read-only, audited).
- 3
Deploy gateway
Route MCP through a managed gateway with RBAC and audit logs. Block direct prod connections from unapproved clients.
- 4
Catalog skills and tools
Publish an approved skills catalog so teams reuse vetted tools instead of inventing new MCP servers.
- 5
Monitor and review quarterly
Alert on new MCP registrations. Review tool invocation logs for anomalous patterns.
Adoption pitfalls
- Policy memo without technical enforcement
- Prod credentials in local MCP configs
- No distinction between read and write tools
Adoption checklist
- Complete MCP inventory documented
- Allowlist enforced technically (not policy-only)
- Audit logs retained 90+ days
- Quarterly access review scheduled
SEER REAL assessment / sprint
Assessment inventories shadow MCP and data exposure. Sprint stands up allowlist gateway + catalog for the highest-risk agent surface.