AI Agent Skill

Version-matched UniFi guidance for coding agents

The installable unifly skill is a discovery stub. It tells an agent to run unifly skill, which prints the guide embedded in the installed binary. Updating the CLI updates its guide at the same time.

🔗Install

Install the unifly binary on the machine where the agent runs, then add the discovery skill for your user:

unifly skill --install

This writes ~/.agents/skills/unifly/SKILL.md, the shared user skills directory. Claude Code reads its own directory, so target it with --claude ($CLAUDE_CONFIG_DIR/skills, or ~/.claude/skills by default), or pick any skills directory with --dir:

unifly skill --install --claude
unifly skill --install --dir ~/work/agent-skills

The command works offline without configuration. It prints the installed path on stdout and its status on stderr (--quiet keeps only the path). Running it again is safe: an identical skill is left alone, and one installed by npx skills add or an earlier release is upgraded in place. Any other SKILL.md at that path, such as a locally edited copy, stays untouched and the command exits with code 6 unless you pass --force. A symlinked skill directory or SKILL.md means something else manages the skill (a checkout, or npx skills add in link mode), so the command reports where the link points and writes nothing, even with --force.

The quick install scripts run this step for you (set UNIFLY_NO_SKILL=1 to skip it). install.sh skips it when run as root, since that home directory may belong to someone else. Homebrew and AUR install unifly for every user and cannot write into your home directory, so they print a reminder instead. After cargo install, run it yourself.

The skills CLI installs the same stub for many agents at once:

npx skills add hyperb1iss/unifly

For Claude Code, the repository also provides a plugin with a network manager agent and status/audit shortcuts:

/plugin marketplace add hyperb1iss/unifly

Older binaries that do not recognize unifly skill need an update before they can serve the guide. Until then, use the installed binary’s --help for syntax rather than assuming current online instructions match its behavior.

🔗Load Guidance as Needed

unifly skill
unifly skill commands
unifly skill concepts
unifly skill workflows
unifly skill examples
unifly skill examples switch-ports.jsonc

The core guide covers targeting, authentication, inspection, authorized changes, and verification. Command notes explain operational exceptions; the appended inventory is generated from the binary’s Clap command tree. Concepts covers configuration, identifiers, output, and secrets. Workflows provides focused recipes for diagnostics, settings, switch ports, and firewall ordering. Named examples print the raw template for saving and editing.

Every topic works offline without a configured controller. Commands used to inspect or modify the network still need the appropriate credentials. Configuration setup uses unifly config init locally or unifly config cloud-setup for Site Manager.

Runtime guidance removes installed skill/binary version mismatch. Operational prose still needs maintenance when behavior changes, and a payload that passes local schema checks may require different values on a particular controller.

🔗Agent and Slash Commands

The plugin’s network manager agent and both shortcuts load the same runtime guide instead of carrying separate command recipes:

Entry pointPurpose
Network manager agentInspect, provision, and verify authorized infrastructure changes
/unifly-statusRead-only health, device, client, and alarm summary
/unifly-auditRead-only WiFi, firewall, firmware, and isolation review

The shortcuts preserve the requested profile and report checks that the available authentication cannot perform. Audit findings include evidence and impact; a failed query is not proof that a resource is unused.

🔗Example Requests

  • “Create an IoT VLAN with a dedicated SSID and isolation from my trusted network.”
  • “Audit firewall policy ordering without changing anything.”
  • “Find the client named ring-doorbell and inspect its WiFi experience.”
  • “Export the switch port configuration so I can review the VLAN assignments.”
  • “List the consoles visible through my Site Manager profile.”

The agent reads the relevant guide topic, inspects state, carries the user’s authorization through the task, and verifies changes afterward.

🔗Next Steps