Host support
Every agent reads the same content, from the folders that agent looks in:
| Agent | Skills | Subagents | Notes |
|---|---|---|---|
| Claude Code | .claude/skills/ |
.claude/agents/*.md |
|
| Codex | .agents/skills/ |
.codex/agents/*.toml |
Start a skill with $skill-name or the /skills picker. Codex never starts a familiar skill on its own. Each skill’s agents/openai.yaml sets this. |
| Gemini CLI | .agents/skills/ |
.gemini/agents/*.md |
No command per skill. Ask for a skill by name (“run the project-status skill”) and confirm when asked. Skill names show without a prefix. This is a limit of the skills format, not a bug. |
Other agents. An agent that reads AGENTS.md and the skill folders is expected to load familiar.
Cursor, GitHub Copilot in VS Code and OpenCode are examples. Skills can appear twice, because the
agent reads both skill folders. familiar has no record of what such an agent supports. It treats
every capability as unknown, does subagent steps itself, and asks before it starts an extension
skill.
In this guide, python means Python 3.10 or newer under the command your computer uses: python3,
python or py. The agent picks the command the same way. On macOS, the built-in
/usr/bin/python3 is an older copy from Apple. Install a current Python from python.org.
.claude/skills/ and .agents/skills/ hold the same skill files, because agents look in different
folders. The one difference is agents/openai.yaml in each skill under .agents/skills/. It holds
the Codex settings that Claude Code reads from SKILL.md instead. If you edit a skill, edit both
copies.
A capability is active only after its released version and a live health check both pass.
Otherwise familiar reports it as available, unsupported, unavailable or unknown. It never
claims a fallback. .familiar/installed.json records the current state for each agent.