Doxloop supports three agent CLIs: Codex, Claude Code, and the Gemini CLI. It discovers an executable on PATH in that order, or you can force one with --agent.
Invocation modes
| Workflow | Codex | Claude Code | Gemini CLI |
|---|---|---|---|
| Create / update | Interactive prompt | Interactive prompt | Interactive -i prompt |
| Review | Read-only ephemeral execution | Print with plan permissions | Plan approval mode |
| External IDE | --print prepared prompt | --print prepared prompt | --print prepared prompt |
review is deliberately non-interactive so it stays read-only. create and update may pause once for consolidated reader or scope decisions.
Screenshot capture
Application guide screenshot capture requires the Codex desktop app or the Claude desktop app because the workflow needs desktop browser access. It is not supported through Codex CLI, Claude Code CLI, or Gemini CLI.
Prepare the request without launching an agent CLI:
doxloop create --screenshots --print "Create an onboarding guide with screenshots."Then submit the printed prompt in a supported desktop app opened at the project root. See Add guide screenshots for application configuration and capture requirements.
Model and reasoning options
Doxloop forwards --model using each CLI's own native model option, so any model name your installed agent CLI accepts works:
doxloop create \
--agent claude \
--model <model-name> \
"Create API documentation for developers integrating this product."--reasoning is Codex-only and accepts minimal, low, medium, high, or xhigh:
doxloop create \
--agent codex \
--reasoning high \
"Create troubleshooting guidance for operators running this product."Where skills are installed
doxloop init (and doxloop agent setup/update) install the shared doxloop-authoring skill and your selected generator's format skill (doxloop-doxbrix and so on) under:
.agents/skills/— read by Codex and Gemini CLI.claude/skills/— read by Claude Code
Manage them explicitly at any time:
doxloop agent status
doxloop agent setup
doxloop agent updateagent status reports each skill path as missing, current, or modified. setup installs anything missing, but refuses to overwrite a skill with local changes — use agent update to replace it explicitly once you've reviewed the diff. review never installs or updates skills, because doing so would mutate the project during a read-only workflow.
When an agent CLI changes behavior
Agent CLIs evolve independently of Doxloop. If an invocation starts failing, confirm the installed CLI version by running it directly, then fall back to a version-independent path:
doxloop create --print "Create a quickstart for developers adopting this product."Paste the printed prompt into your agent session manually.