unifi-agent

Agentic command and control for UniFi networks. A safety-first hybrid API client, MCP server, and CLI that lets an LLM (or a cron job, or you) manage a UniFi network end to end over the LAN: read full performance telemetry, tune AP radios, run speedtests, block clients, toggle SSIDs, restart devices, take backups, and run unattended diagnostic loops that surface optimization recommendations.

Built local-first and security-first: no new inbound attack surface, TLS certificate pinning for the console's self-signed cert, secrets in the OS keyring (never in the repo or the agent's context), and every mutation gated by a blast-radius-aware safety layer with dry-run previews and an audit log.

Not affiliated with Ubiquiti Inc. "UniFi" and "Ubiquiti" are trademarks of their owner.

Why

UniFi has three partial APIs and no official MCP. The clean official Integration API can't tune radios, run speedtests, or read events; the powerful classic API is unofficial and needs a local login; and an API key doesn't reliably authenticate the classic endpoints. Most tools pick one side and lose capability. unifi-agent uses both — the official API for what it covers, the classic API for the gaps — and labels the provenance and risk of every operation. See docs/ARCHITECTURE.md.

Related MCP server: go-unifi-mcp

Features

• Full visibility: device/client inventory, per-subsystem health, per-AP radio status, events, historical reports, last speedtest — normalized into compact, LLM-friendly shapes.

• Full control (guarded): set radio channel/width/TX-power, enable/disable WLANs, block/unblock/kick clients, authorize guests, restart devices, power-cycle PoE ports, run speedtests, create backups.

• Two front ends, one core: a stdio MCP server (17 tools) and a CLI that share the same safety-guarded facade.

• Diagnostic loops: diagnose / watch run read-only passes that flag offline devices, weak clients, subsystem warnings, and co-channel interference, and emit JSONL for cron/launchd or an LLM loop.

• Safety by construction: read-only mode, blast-radius ceiling, mandatory confirm, dry-run previews, pre-change backups, race-mitigating read-modify-write, redacted audit log. See SECURITY.md.

Quickstart

uv venv && source .venv/bin/activate && uv pip install -e ".[dev]"
unifi-agent setup      # stores API key / local admin in the OS keyring, pins the TLS cert
unifi-agent doctor     # verifies connectivity + each API surface
unifi-agent overview

(Not activating the venv? Prefix commands with uv run, e.g. uv run unifi-agent doctor.)

Full provisioning (creating the API key and a dedicated local admin) is in docs/SETUP.md.

CLI examples

unifi-agent devices                     # inventory
unifi-agent radios                      # per-AP channel/width/power
unifi-agent diagnose                    # findings + optimization recommendations
unifi-agent speedtest-last              # most recent ISP test

# Every write previews first; add --confirm to apply.
unifi-agent set-radio "Main Floor U6+" 5GHz --channel 44 --width 80 --dry-run
unifi-agent set-radio "Main Floor U6+" 5GHz --channel 44 --width 80 --confirm
unifi-agent speedtest --confirm
unifi-agent block aa:bb:cc:dd:ee:ff --confirm

unifi-agent watch --interval 300        # unattended diagnostic loop (JSONL)

MCP

{
  "mcpServers": {
    "unifi": {
      "command": "/absolute/path/to/repo/.venv/bin/unifi-mcp",
      "env": { "UNIFI_HOST": "192.168.1.1" }
    }
  }
}

Use the absolute path to the venv's unifi-mcp (or uv run --directory /path/to/repo unifi-mcp); unifi-mcp is not on the MCP client's PATH. Read tools are always safe. Write tools return a preview (applied: false) with the blast radius and predicted change unless called with confirm=true; the model must preview, show you, and confirm. The server is stdio only — it opens no port.

Safety model (short version)

Blast radii: none < client < device < wlan_group < site < gateway. WLAN edits and radio changes reprovision APs and drop their wireless clients for ~30-60s; the tool warns before these and encourages one-AP-at-a-time changes.

Remote access

Don't expose the tool or the console to the internet. To manage on the go, VPN back into the LAN (the gateway already runs a WireGuard server) and use the same commands. Same security properties, zero new surface.

Development

uv venv && source .venv/bin/activate && uv pip install -e ".[dev]"
python -m pytest -q          # 47 tests: safety, radio logic, redaction, auth/CSRF, TLS pinning, facade, loops
ruff check src tests

License

MIT — see LICENSE.