Claude
The support matrix is the main reference for what Farhelm does with Claude. This page only adds what is specific to Claude.
Gotchas and quirks
Section titled “Gotchas and quirks”- Guarded selects Claude’s auto mode with
--permission-mode auto: its classifier can approve or block actions without asking about everything. Your allow rules, hooks and later mode switches can loosen it. See Permission modes. Farhelm also tries to notice bypass permissions on in Claude’s status line and show a YOLO shield; this is best effort, not a check of your allow rules or hooks. - Claude may ask whether you trust a directory the first time it starts there. Answer it in the session’s terminal, or opt into answering it for future managed checkouts under Agent trust in Settings. That setting is off by default: it edits Claude’s own settings file and comes with caveats. While Claude asks this in a managed checkout and the setting is off, a note above the session’s terminal says Farhelm can answer it, and its Settings and caveats opens Agent trust at that setting. Directories you choose yourself still get Claude’s question.
- If you start Claude through a wrapper prefix or a custom command, Farhelm can follow its conversation through multiple wrappers that stay running. Each wrapper must pass Claude’s arguments and environment through. See Agent wrappers for the requirements and how Farhelm switches back at the end of a turn if another Claude started inside the session reports first.
- On a Linux host, Claude’s own way of opening links does not reach your browser through Farhelm. Ask it to run
farhelm agent open-url -- <url>instead (Open links from a remote session).