AI assistant

The assistant answers your questions by looking up AquasManager’s data. It’s optional and off by default: AquasManager ships no AI model, you plug in the one you choose. Once turned on, it opens from the chat icon.

What it can do

The assistant has tools that query the app:

• list your aquariums;
• read a water parameter’s latest value;
• read a sensor’s current value;
• give a device’s state;
• calculate a dose.

“What was the reef’s last KH?” “Is the heater on?” “My product raises 100 L by 1 °dKH per 10 mL: how much should I add to bring the reef’s KH up to 8?”

Reliable dosing

Dosing is never left to the model: it’s a deterministic server tool that the assistant calls.

• The parameter’s current value and the tank volume are read from the app.
• The product’s dosing rate (on the label) is asked of you if unknown: the assistant never makes it up.

Always check the suggestion before dosing.

Connect a model

Everything is set in Settings › AI assistant (Configuration › Assistant IA, administrators):

1. choose a service, pre-filled in one click;
2. enter the API key if the service needs one;
3. click List to see the available models, or type a model’s name;
4. Test the connection, then turn on the assistant and save.

Service	Type	Where your questions go
Ollama	local, free	stay at home
LM Studio, llama.cpp, vLLM…	local, free	stay at home
Mistral AI, OpenAI, OpenRouter…	external, pay per use	sent to the service

Two API types are supported: the Ollama API and any OpenAI-compatible API (address usually ending in /v1).

External service

With an external service, your questions and the data read to answer them (tank names, readings, equipment states) are sent to it and fall under its terms of use. The interface tells you so. To keep everything at home, use a local model.

Settings

• Data access: turn it off for an assistant that only answers in general terms, without reading anything in the app.
• Response timeout: without a GPU, a local model can take a minute or more.
• API key: encrypted on your server with the install’s secret key, never shown again. If the secret key changes, enter the API key again.

From Docker

Inside the API container, localhost means the container itself. For Ollama installed on the same machine, use the machine’s IP address, for example http://192.168.1.20:11434.

Choose a model that supports tool calling, otherwise the assistant won’t be able to read your data. Also check that the model’s licence allows your use.

Connect your own assistant (MCP)

AquasManager is also an MCP server (Model Context Protocol): Claude, Home Assistant or any compatible assistant can look at your aquariums, without going through the built-in assistant.

1. In your account menu, open AI assistant access (Accès pour assistants IA) and create a token. It is only shown once: copy it and keep it like a password.
2. In your assistant, add an HTTP MCP server at the address shown on the page (the API address followed by /mcp), with the Authorization: Bearer <token> header.

Example with Claude Code:

claude mcp add --transport http aquasmanager https://api.example.org/mcp \
  --header "Authorization: Bearer aqm_..."

With Home Assistant: add the Model Context Protocol integration. It only takes an address, with no header: put the token in the URL, https://api.example.org/mcp?token=aqm_.... Then, in your conversation agent (Ollama, Google…), tick AquasManager under “Control Home Assistant”.

Access is read-only: aquariums and livestock, water parameters with their trend and history, sensors, equipment, alerts, tasks, stock, logbook, water changes and dosing calculation. The assistant can’t change or control anything. A token can be revoked at any time from the same page.

Caution

The data you look at goes through the service of the assistant you connect (Anthropic, Google, OpenAI…). To stay 100% local, use a local model, with the built-in assistant or a local MCP client.