One security fix in lo mcp. No change to the piped output of the parity cases.

Security

• lo mcp no longer offers a command whose output is a credential, in any tier (#240). A tool result lands in the model's transcript, and with a hosted model it also reaches the provider. The Go server used to offer four of them in the default readonly tier.

  • Withheld: kubeconfig (the admin kubeconfig with its client key), secrets print, secrets env, kubehz token and kubehz claim-code.
  • The --print-only flag of kubehz node join is never exposed, because it prints the bootstrap token.
  • The lo chat defaults deny the same set.

  The tiers now hold 25, 49 and 88 tools.

Upgrade

Nothing to do. An editor or agent that called one of these tools through lo mcp runs the command itself instead, outside the transcript.