MCP Stub problems and fixes

Note whether the tool was absent, visible but failing, or returning an unexpected result. Refresh the tool list, then try a permitted read. Do not send environment dumps, credentials or tokens to support.

Missing startup arguments, a missing executable or a readiness timeout are deployment problems. Give the workspace owner the redacted message. Do not delete sockets, replace binaries or change shared settings: other conversations can depend on them.

The list can remain from before an outage. After recovery, refresh it and try a small read. Before repeating a write, check the destination: the first request may have arrived before its response was lost.

Confirm the authorization finished, then ask the assistant to refresh its tools. If it is still absent, ask the owner whether it is published for your workspace. An empty list can mean the upstream is not reached yet; refresh once, then report whether this happened on a new connection.

Check the workspace and account, and finish the approved authorization flow. Reconnecting does not grant permission, and labels are not access controls. Ask the service owner for access you need.

A service that changes its tool list often causes repeated refreshes; report it. A resource link can expire, so ask for a fresh result. A missing notification does not mean nothing changed: read the current state. A missing display, audio device or GPU belongs with the workspace owner.

Check whether a known permitted item can still be read. An empty list during an outage does not prove that content was deleted. After the connection recovers, refresh that list and read the known item; if it remains absent, ask the service owner to check publication and access.

Inspect the task through the service's own controls. Closing the assistant requests only best-effort cleanup, and abrupt termination can skip it. End remote work through those controls and revoke account access separately when needed.

Ask the owner to confirm that a known permitted skill is published and that the assistant, MCP Stub and upstream service support it; MCP Stub must be at least version 0.5.0. After an upgrade or service recovery, reopen the assistant connection, list skills and read the known skill. A slow or unavailable service at startup can leave skills missing even after tools recover. If discovery works but reading fails, keep the redacted error and have the owner check that skill's availability and access.