lanowl

Setting it up

The model

The owl is a local language model, run by Ollama on a machine of yours. It explains incidents, answers your questions, reviews security every morning and proposes fixes. It never decides whether something is down: that is lanowl's own code, and it works the same with the model switched off.

Setting it up

Until config.yaml has a model: section (or LANOWL_MODEL_URL gives one), there is no model: the owl is off. lanowl asks nothing, checks no model server and warns about none; the alerts, the digests and the dashboard work as always, and the owl's card on Now says where to set one. A new install starts this way.

  1. Install Ollama on the machine with the most memory or the best GPU (it can be another machine than lanowl's).
  2. Pull a model with tool calling, for example ollama pull qwen3.8:27b, the one lanowl is tested with.
  3. Make Ollama listen on the network if it is on another machine (OLLAMA_HOST=0.0.0.0), and point lanowl at it:
model:
  provider: ollama
  url: "http://192.168.88.6:11434"    # or LANOWL_MODEL_URL in docker/.env
  name: "qwen3:30b"
  think: true
  num_ctx: 32768
  keep_alive: "10m"
  request_timeout_s: 300
  max_tool_iters: 10
  temperature: 0.2
  persona: "owl"
Key What it does
url, name Where Ollama is, and which model. LANOWL_MODEL_URL wins over url.
think How much a thinking model reasons before it answers, one setting for every call: false, true (the model's own default level) or one of the levels the model lists (for example low, medium, xhigh). Settings โ†’ The model offers the levels your model has; a level it does not list means its default. false answers fastest; the level can change what the owl concludes.
num_ctx The context window. Every tool result stays in the prompt, so give it room: past it, Ollama silently drops the start of the prompt, and lanowl's log warns when an audit came within 10% of it.
keep_alive How long the model stays loaded after a call. 0 frees the memory at once.
request_timeout_s The longest one call may take.
max_tool_iters The most tool calls in one audit.
persona The voice of its prose: owl (the owl's voice, opening with its call, "Hoo.", on Telegram), or "" for plain prose.

Which model. Any Ollama model with tool calling works. A larger model reasons better; a 27โ€“35B model on a machine with 32โ€“64 GB of memory is comfortable.

Tested with qwen3.8:27b, and it works well: lanowl is developed against a real home network with it (the Apple-silicon build qwen3.8:27b-mlx, about 18 GB, with think: true and num_ctx: 131072).

Only Ollama is supported for now; cloud models are on the roadmap.

The owl's voice

With persona: owl, the model speaks as the owl. On Telegram its words open with its call, "Hoo.", once in each message (the dashboard shows them without it), and are marked ๐Ÿฆ‰ wherever they appear, so you can always tell the model's words from the monitor's measurements, which keep their ๐Ÿ”ด๐ŸŸก๐ŸŸข. It says what happened and whether it needs you, in a sentence or two:

๐Ÿฆ‰ Hoo. NAS (192.168.88.20) is down since 03:12 and needs you. Likely power: its switch port dropped the same second as the UPS beside it.

What it keeps to:

  • Only what you need. What is fine, and what the message already prints under its words, is left out; what it told you before comes back only as "still" or "again".
  • Every device with its address. A device never seen before also with what it is: its maker, and its hostname or MAC.
  • How sure it is, in one word: likely or a guess. One error in a log is one error, never a dying disk.
  • Where it read the why: the router's log, the ARP table, a device's own log, quoted as written.
  • "Needs you" only for a critical or high problem that is still open. It suggests and never orders, and never offers something lanowl cannot do.
  • It never changes a severity and never invents a device, a number or a cause.

With persona: "" the model writes plain prose, without the call.

What makes its answers good is network.description in config.yaml: a few sentences about your network that every prompt includes. Say what no inventory says.

What it can do

Look, with read-only tools: the history of every device, the internet's history and evidence, the router's log and tables, a machine's log, MQTT values, ping, TCP and HTTP to an inventory device, and everything lanowl has kept: what it sent you, the update check, the security review, the backups, earlier conversations. An address outside your inventory is refused.

Read a device itself, in your questions: the speed a port really linked at, its own log, its Wi-Fi clients. lanowl logs in with the device's own login, the one it uses for updates and backups, and runs one fixed read-only command per read; the owl picks the read. What each kind offers: Device kinds.

Run a command on a device, in your questions, for what the fixed reads do not show: the routes, a setting, a counter. The owl writes the command; one on the device kind's read-only list runs at once, anything else waits for your approval with a button and runs exactly as you read it (Device kinds).

Run diagnostics, from a fixed catalog: mtr, traceroute, DNS and TLS checks, a port scan, the router's own ping, a packet capture, each internet link on its own, a check from another LAN host or from your VPS. When you ask a question, the ones it needs run at once and are listed under its answer. The hourly audit, which nobody asked, may run a few passive ones by itself (actions.session.audit_checks); for anything more, or anything active like a scan, it asks you to open an investigation session, with a button.

Propose changes, from another fixed catalog, that run only when you approve: Actions and approvals.

Remember what you tell it ("remember that the NAS sleeps at night"), in notes you can read, edit and delete on the dashboard or with /memory.

Its shell, in a sandbox

For questions a fixed tool cannot answer, the owl can get a shell: one bash command at a time, in a container of its own beside lanowl. It walls off the machine lanowl runs on, so it needs that machine's address (Settings โ†’ The model's shell gives the line, filled in), on the command line or in docker/.env:

LANOWL_HOST_IP=192.168.88.5 docker compose -f docker/compose.yaml --profile shell up -d
shell:
  enabled: true
  canary_lan: "192.168.88.10:22"     # a LAN service that always answers a handshake
  audit: {enabled: true}             # the hourly audit's own, offline sandbox

The sandbox has no key, no config and no state of lanowl's. It reaches the internet, and the LAN only for handshakes and scans. It cannot reach lanowl's dashboard. Its walls are proven before the tool is offered: from inside, a handshake to canary_lan must succeed (the network is there), the HTTP request after it must get nothing, and lanowl's dashboard must not answer at all. Anything else switches the shell off, and you are told once. The proof runs at start, every morning with the security review, and whenever the sandbox restarts.

The hourly audit gets a second sandbox with no internet and no DNS at all, and at most a few commands per audit. Each command is listed in the next digest that goes out, under "The audit looked for itself", and with what came back on the dashboard (Ask โ†’ The model's shell).

Switching it off

/model off on Telegram, or the switch on the dashboard's Ask tab, puts the owl to sleep: no diagnoses, answers or reviews, and Ollama is told to free its memory. Alerts, digests and the dashboard go on. /model on wakes it.

Edit this page on GitHub.