ai-enabled-hr-talent-automation

MCP Tool Governance

Title: MCP Tool Governance Version: 1.0 Owner: [TENANT_CONFIGURATION_REQUIRED — Integration Architecture / Security] Status: Draft Last reviewed: 2026-09-07 Next review: [TENANT_CONFIGURATION_REQUIRED] Reviewers: Architecture, Security, AI Governance

Purpose and scope

Defines the lifecycle for adding, versioning, and retiring MCP tools/servers, and the onboarding/offboarding/security-review process referenced by mcp-architecture.md.

Tool manifest and versioning

Every MCP server publishes a versioned tool manifest (schema: hr-tools.schema.json) declaring: tool name, tier (read-only/propose-only/approval-required write), input/output schema, required scopes, and rate limits. Manifest changes are versioned; the Tool Gateway pins to a specific manifest version per environment and only upgrades after review.

Onboarding process for a new MCP server

flowchart LR
    A[Proposal: new integration domain] --> B[Security review: data flows, credential scope, egress needs]
    B --> C[Define tool manifest - tiers, schemas, scopes]
    C --> D[Implement + deploy in isolated namespace]
    D --> E[Red-team / abuse-case testing per red-team-plan.md]
    E --> F{Pass?}
    F -- No --> C
    F -- Yes --> G[Approve for staging, then prod rollout]
    G --> H[Register in mcp/servers/README.md and integration-adapter-catalog.md]

Offboarding process

  1. Confirm no active skill references the server’s tools (dependency check against agent-skill-catalog.md).
  2. Revoke credentials and remove network egress allowlist entries.
  3. Retain audit logs per data-classification-and-retention.md retention rules (audit logs outlive the integration itself).
  4. Remove from the tool manifest registry and update documentation.

Change governance for existing tools

Change type Approval required
Add a new read-only tool Architecture review
Add a new propose-only tool Architecture + AI Governance review
Add or modify an approval-required write tool Architecture + Security + AI Governance review (this is the highest-risk category)
Change a tool’s scope/permissions Security review, re-run security-test-plan.md tenant-isolation and authorization tests
Deprecate a tool Follow versioning-and-deprecation-policy.md pattern

Ongoing security review

Each MCP server undergoes a security review on a configurable cadence (default: annually — [TENANT_CONFIGURATION_REQUIRED]) or upon any credential/scope change, covering: credential scope minimality, egress allowlist accuracy, and manifest-to-implementation consistency.

Change control

Version Date Author Change
1.0 2026-09-07 Documentation package generation Initial creation
This site is open source. Improve this page.