Agent Comments works with local command-line agents. Pick a preset in settings, run the health checks, then adjust the commands if your local CLI differs.
Background reply agent:
codex execTerminal agent:
codexUse this when you want Codex to reply in comment threads or write CriticMarkup back into the Markdown file.
Background reply agent:
claude -p --permission-mode acceptEditsTerminal agent:
claude--permission-mode acceptEdits lets headless claude -p write CriticMarkup back to the file. Without it, Claude runs with normal permissions, can't apply edits non-interactively, and will report that it couldn't write instead of replying. If your Claude CLI uses different flags for stdin/non-interactive mode, update Reply extra args and run the background health check.
Background reply agent:
opencode runTerminal agent:
opencodeOpenCode CLI behavior may vary by version. Use the preset as a starting point and run the health checks.
Any command can work if it can:
- read the prompt from stdin
- write a response to stdout, or edit the Markdown file directly
- run from the vault root
For comment replies, the command should either append {>>AgentName: reply<<} to the target thread or print a concise reply that Agent Comments can append.
Agents differ in whether their non-interactive mode can edit files:
- Codex (
codex exec) writes to the file directly. - Claude (
claude -p) needs--permission-mode acceptEditsto write; otherwise it falls back to printing a reply. - OpenCode and others vary — check your CLI's headless/permission flags.
If an agent can't write, Agent Comments still appends its stdout as the reply. Granting write permission lets the agent apply CriticMarkup edits in place.
Obsidian launched from the GUI inherits a minimal PATH. Agent Comments augments it with common locations (~/.local/bin, /opt/homebrew/bin, /usr/local/bin, …). If your agent binary lives elsewhere, set the Reply command / Terminal launch command to an absolute path.