A multi-client pattern needs one policy layer
When shared workflows move through MCP, the client list grows quickly. One team may standardize on Codex, another may prefer Claude Code, and another may run opencode in terminal sessions.
The policy layer should remain stable. Invokora keeps organization Skill access, delivery mode, and audit evidence in one place.
- Name supported clients as examples of MCP compatibility.
- Keep the canonical Skill handle and policy server side.
- Avoid maintaining separate source-delivery rules per client.
Review local tools separately
MCP Skill access is only one part of local AI client risk. Filesystem tools, shell execution, network access, and secrets still need client-side review.
A clear rollout separates Invokora-managed Skill policy from the broader local tool surface that each MCP host exposes.
- Keep local tools narrow and named.
- Block write-capable actions in plan-only or review modes.
- Document which local capabilities are outside Invokora Skill policy.
Operate through evidence, not preference
Security teams do not need to approve a favorite client. They need evidence that organization controls are applied consistently regardless of the client used.
The review packet should show identity checks, source delivery boundaries, delivery approvals, blocked outputs, and audit export paths.
- Track the execution surface for each run.
- Make delivery mode visible before source sync.
- Use resources and security pages as shared review material.