Ask Danbyte (in-app chat)¶

A chat panel in the top bar, next to the docs button, that answers questions about your own network: "which devices are at Aarhus and are any of them down?", "what changed on this switch yesterday?", "what goes end of support next year?".

It reads through the same tools the agent access endpoint uses, as you, so it sees exactly what you see on the pages - no more. Every lookup it makes is shown in the answer, so you can check it rather than take it on trust.

It is off until you turn it on, and it needs a model connection.

What leaves your network¶

This is the honest part, and the reason for the local option.

Unlike agent access, where your own editor talks to a model you chose, Danbyte itself calls the model here. Your question, the conversation so far, and whatever the tools read on your behalf are sent to the provider a deployment admin configured. With Anthropic or an OpenAI-compatible service, that means your inventory data goes to that company.

Choose the local provider - Ollama or LM Studio on your own network - and nothing leaves the building. That option exists precisely so an airgapped install can use the feature.

Turning it on¶

Two steps, deliberately at different levels.

1. Connect a model (deployment admin, Settings → Security → Assistant model):

Test connection asks the model for one word, so a wrong key or a stopped Ollama is obvious here rather than in the chat.

The key is stored encrypted and never returned by the API. Only a deployment admin can set any of this: it decides where your data goes, so a tenant admin must not be able to change it.

2. Turn the chat on (tenant admin, Settings → Integrations → Assistant chat). The top-bar button appears for everyone in that tenant.

Using it¶

Open it from the top bar or press ⌘J (Ctrl+J). Type a question, Enter to send, Shift+Enter for a new line.

• Lookups are shown. Each answer lists what it read - list · 12 rows - and expanding one shows exactly what was asked.
• New conversation starts a fresh thread. Past conversations lists the ones you have had.
• Conversations are private to you. Nobody else sees them, not other members and not an administrator. Delete one from the list, or all of them at once.

Asking it for a script¶

Some questions want the same answer next month, not once. Ask for a script - "write a script that lists devices with no serial number", "check every month which circuits have no terminations" - and it writes one against the same SDK you would use by hand, saves it and links to it.

It cannot run a script. You open the script, read the code, fill in its parameters and run it yourself; the run log and any files it wrote are on the script's page. Saving a script needs writes on and the script permission; without either, it shows you the code to paste in.

What it can and cannot do¶

• It runs with your permissions. A site-scoped account's chat only sees that site.
• Results are capped at the row limit on the Agent access page, and it is told when a result was cut short.
• "Where do I change X?" is answered from the settings catalog rather than from memory, so the link it gives you is a page that exists.
• Cabling and circuits have their own tools: it cables two ports by name, and lands a circuit's A and Z ends at their sites, rather than assembling the payload by hand.
• It cannot change anything unless Agent access: allow writes is on. With writes off it says so plainly. With writes on, a change follows your own permissions and lands in the change log with Via: Assistant and your own name in the User column.
• Every lookup is recorded on Integrations → Agent access → Recent calls, with the client shown as chat.
• It stops after several rounds of looking things up, so a confused model cannot loop.

Troubleshooting¶

The chat streams over the WebSocket service rather than the API, because a long answer would outlast an ordinary HTTP request. A deployment that proxies /api/ but not /ws/ gets the whole product except this panel.