**Identify the publisher and source.** Find the canonical repository or registry listing and confirm the install command points at that source, not a similarly named copy. Expected: publisher, repository, and package name all match each other.

**Pin an exact version.** Install a specific released version instead of a floating latest tag. Success check: the config records the exact version, so the code cannot change silently on the next launch.

**Read the tool list before connecting.** From the server documentation or a dry listing, note every tool name, what it reads, and what it writes. Expected: no tool surprises you; any tool with broad file write or shell access is flagged for step 5.

**Review the credential scope it asks for.** List each credential or environment value the server requests and whether a read only or narrowly scoped option exists. Success check: the server runs, or can be trialled, without your primary long lived credentials.

**Run an isolated trial.** Start the server in a separate agent profile or sandbox folder with test data only, and call only its read only tools. Expected: calls return documented results and touch nothing outside the test folder.

**Decide and record the outcome.** Promote, restrict, or reject the server, and note the version and the tools you approved. Success check: a future update triggers a fresh check against that record instead of silent acceptance.