MCP is a first-class Lumic interface and the highest-leverage way to operate multiple nodes.
The target workflow is:
- install Lumic independently on each VPS;
- configure each Lumic MCP server in the coding agent;
- let the agent inspect node status and capabilities;
- describe the infrastructure outcome instead of manually debugging SSH sessions.
Example instruction:
Create production on node-01 and staging on node-02. Inspect this repository’s requirements, provision both environments, configure Git deployment and HTTPS, and verify health.
Typed tools, not shell
MCP tools should look like:
inspect_server
install_package
install_runtime
install_component
install_service
create_application
plan_deployment
deploy_application
rollback_deployment
search_logs
diagnose_server
Lumic must not expose unrestricted root shell execution as its normal MCP model.
Policy
MCP access is capability-based. A production identity might be allowed to inspect, deploy and restart a service while being denied database deletion, firewall changes, OS upgrades or raw execution.
Mutating operations should carry actor/interface/correlation metadata into Lumic’s audit trail.
Current implementation
The v2 repository already contains an MCP crate boundary. Phase 0 introduces the first read-only server status capability before broader mutations are exposed.