Main
- Get permissions right first: README calls out Screen Recording + Accessibility; without them, agents will fail mid-run.
- Use screenshots as evidence: capture before/after images so agent runs can be audited and reproduced.
- Prefer structured element IDs + actions (
click,set-value,perform-action) over coordinate clicks to reduce UI flakiness.
Source-backed notes
- README says it’s a macOS CLI and optional MCP server, exposing the same tools through both interfaces.
- README Quick start shows commands like
peekaboo image ...and an MCP server run vianpx -y @steipete/peekaboo. - README lists a requirement for Screen Recording + Accessibility permissions (linked docs/permissions.md).
FAQ
- Do I have to use Homebrew?: No—README also supports running the MCP server via
npx -y @steipete/peekaboo. - Why are actions failing?: Most issues are missing permissions; confirm Screen Recording + Accessibility are enabled.
- When should I use MCP mode?: Use MCP mode when you want agents to call tools directly from Codex/Claude Code/Cursor.