MCP Docker problems and fixes

Check which container service is selected and whether your account can reach it. Point the tool at the service you mean, then ask for its system information. Local connections and plain network connections work; secure web and remote shell connections are refused, and Windows pipes need Windows. With nothing set, the local service is used.

A successful connection does not prove the versions fit. Use a service or a tool build that your team has approved as compatible; the tool does not negotiate versions by itself and has no setting for it.

Compare the image or destination with the list of approved registries. Pick an approved one, or ask whoever runs the tool to add yours and restart it. Sign-in for a private registry is supplied with each download or upload; it never overrides the list.

Look at what already runs in the session before raising the limit. Remove only what is disposable, or ask for the approved higher limit before the tool starts. Cleaning up a session can delete the volumes it tracks, so use it only when you mean to tear down.

Look at Docker before you repeat a change: running out of time can leave a container, image or finished build behind. Requests have a fixed one-minute limit. A time-out while waiting does not stop the container; stop it separately only if you mean to.

Read the whole error and compare the request with the tool’s current description; remove fields it does not list. A missing path does not mean the container is gone: fix the source path or create the destination folder first.

Only mounts of a folder from the service’s computer are supported today: use one with an approved source and a destination in the container. Creating a volume does not attach it.

Disconnecting does not remove containers, and releasing a session only stops tracking. Remove what is disposable explicitly, read clean-up errors and keep volumes that hold valuable data. Images are removed separately.

Use the original session if you can, or have the container adopted. Adopting shares control with the original owner; it is not an exclusive transfer. Networks and volumes cannot be adopted.

Records can be lost if the tool keeps them only in memory, cannot read its saved state or has two writers. Keep the container IDs, recover one step at a time and check the container before adopting again. Missing tracking does not mean the data was deleted.

Use the tool’s built-in help for the command, and use your agent’s tools for anything the command line does not support. Watching events and statistics can fail before showing results; watch for a short, fixed time instead. Give --since and --until as Unix timestamps in seconds; a duration such as 10m is rejected.

Read the whole answer, not only the status: it can say degraded and still answer. Fix access to the service, then try a harmless action to prove it.