Skip to content
Nautelle

Using Nautelle

Assistant and AI providers

A chat that reads your board, and the provider keys behind it.

The assistant is a chat on every page of the web app. It reads your Board and fleet, drafts cards and queues work for you. It calls model providers directly from the API using keys an admin sets, separate from the Claude Code sessions on your runners.

Using it

Open the launcher at the bottom of any page. The window can be dragged and resized, and it opens as a sheet on a phone. Conversations are listed on the left, grouped by day, and each one is personal to you: nobody else can open it, an admin included.

In the composer you can attach a card, an agent session (only in projects that ship session content) or a file or screenshot, and pick a model and a project. The project defaults to Auto: the assistant matches the project you name against the ones you can access, and asks when several match. Replies stream, and you get them in the language you wrote in.

What it can do

  • Read projects, cards, fleet status, the queue, budgets, activity, sessions and approvals, with the same role checks as the web.
  • Draft and queue cards, as New work does.
  • Request a stop, a handoff or a merge. These only create a pending action request for you to approve. The assistant never executes them, and writes need operator or higher.

In Telegram, free text that is not a command goes to the same assistant as your linked user. See Telegram.

Admin: AI providers

An admin sets this up under Admin → AI providers.

  • Providers are Anthropic, OpenAI, xAI Grok and OpenRouter.
  • A key is write-only: you can set, test, replace and delete it, and Nautelle shows only its last four characters. Keys are sealed with AES-256-GCM under LLM_SECRETS_KEY, a base64 32-byte master key separate from APP_ENCRYPTION_KEY. Without it the gateway does not start.
  • Enable models per provider and set the default.
  • Each provider has a monthly cap. Anthropic defaults to $200 and the others to 0, which means disabled. A provider at its cap refuses calls until the first of the next month (UTC), and the chat says which provider and when it resumes.
  • Every call is recorded with user, model, tokens, cost and feature, priced from the price table.

Next steps

ConfigurationSet LLM_SECRETS_KEY.Usage and budgetsSee what the assistant costs.