Five tools and one guidance prompt.
| Tool | Use it to | Main fields |
|---|---|---|
aide_query | Inspect project guidance, cluster or route state, setup or diagnostics | action, optional project context and route namespace |
aide_project_init | Write a new project file without overwriting an existing one | name, domain, a trust-material choice and project context |
aide_cluster_apply | Create, start, stop or select a cluster | action, name for selection, project context |
aide_cluster_delete | Delete a cluster and optionally its configured named volumes | confirm, purge, confirmPurge, project context |
aide_route_apply | Redirect or reset an existing Service | action, service, namespace, optional host and port, project context |
Use the exact field spelling and types in the tools reference. projectDir is a directory in the server's environment; projectFile selects the project configuration. Relative file paths use that working directory. Supply absolute paths when the server starts elsewhere.
Read before the change.
aide_query actions are project, clusters, cluster_status, routes, route_status, setup and doctor. Cluster listing reads registered records; route listing shows managed routes, not every Service. Setup checks prerequisites without applying setup. Diagnostics and route reports need their findings read even when the command exits zero. A report cannot guarantee a later operation will succeed.
Keep project and default selection distinct.
Initialize once with a name, domain and either generateCa:true or the caCert/caKey path pair. Never send key contents. If you will stop and start the cluster, supply your own CA files rather than generateCa. See the cluster tutorial for creating the pair. Existing project files are not overwritten.
Create uses the project file and rejects an existing cluster. Start and stop act on that project's cluster. Selection requires name and changes the default and Kubernetes context for a registered cluster recorded as running; it does not rewrite the project file. Later project operations still require the intended project context.
Deletion without literal boolean confirm:true returns a generic reminder without inspecting or changing the target. Review the target and retention policy separately. To request configured named-volume purge, also send purge:true and confirmPurge:true. These fields express the client's request; obtain approval through the client's workflow. Verify residual state after cleanup.
Verify the actual application.
An existing Service and namespace are required. External routing accepts a host and port, port only for automatic local-address selection, or neither to reuse a saved destination. Internal routing accepts both endpoint fields for a pin, or neither to reuse a saved pin or ordinary routing. Reset requests restoration and clears the override. Use a literal IPv4 address reachable from the target environment and an integer port 1–65535; the endpoint is a backend destination, not a public address to create.
Record a working application baseline before redirecting. Query route_status and repeat the normal application request after redirect and after reset. Read warnings and involve the application owner if restoration fails. Keep the development backend running until the original response is verified.
Separate completion from verification.
Completed commands return MCP text holding JSON with command, exitCode, stdout, stderr and truncated. Read isError, then parse the text. Invalid arguments are protocol errors, including a purge request missing its second confirmation. Start and timeout failures instead return tool-error text without a command envelope.
The wait is 30 minutes and each returned stream retains at most 16 KiB of original output bytes. truncated:true means output was shortened. Timeout or failure can leave earlier changes; inspect before retrying. See Troubleshooting.
The static aide_manual prompt takes no arguments and summarizes task/context safeguards. It does not inspect the project or authorize changes.