Skip to main content

Ollama smart home integration, free and open source

Ollama integration for Gladys Assistant

Ask your local Ollama models from a scene, and follow the state of the server.

View on GitHubAll integrations

Ollama runs large language models (Llama, Mistral, Qwen, Gemma…) on your own machine. This integration lets your Gladys scenes ask them questions, with the scene information in the question, and use the answer in the next actions: send it as a message, have it read aloud, test it in a condition. Everything stays on your network: no account, no cloud.

It also follows the server: online or not, models installed, models in memory and the memory they hold, with a dashboard widget and two scene triggers.

Gladys already has an AI assistant through Gladys Plus (OpenAI). This integration is its local counterpart: free, private, and as quick as your hardware.

What you need​

  • Gladys 5.1 or later.
  • An Ollama server on your network, with at least one model (ollama pull llama3.2, for example). A small model (1 to 3 billion parameters) answers in a few seconds on a recent processor; a graphics card makes larger models usable.
  • If Ollama runs on another machine than Gladys, it must accept the network: start it with OLLAMA_HOST=0.0.0.0 (see the Ollama FAQ). Ollama has no password: keep port 11434 inside your network.

Set up​

  1. In Gladys, open Integrations, find Ollama and install it.
  2. In Configuration, fill in:
    • Address: the IP address or name of the server, e.g. 192.168.1.20 (port 11434 is added), 192.168.1.20:8080, or the full URL of a reverse proxy (https://ai.example.org);
    • API key (optional): only for a reverse proxy that asks for a Bearer token;
    • Default model: the model used when a scene action names none, e.g. llama3.2;
    • Longest wait for an answer: 30 to 110 seconds (60 by default);
    • Keep the model in memory: how long Ollama keeps the model after a question or a load. Longer means quicker answers, but the memory stays taken;
    • Refresh interval: how often the state of the server is read (one minute by default);
    • Language of the feature names: the names given to the features and the texts of the scene events and errors.
  3. Save, then press Test the connection: it shows the version of Ollama and the models installed, with their exact names.
  4. In the Discovery tab, add the Ollama device to Gladys.

The Ollama device​

FeatureMeaning
Online1 when the server answers, 0 after two failed readings in a row
Installed modelsnumber of models on the server
Models in memorynumber of models loaded, ready to answer
Memory usedmemory held by the loaded models, in megabytes

Dashboard​

The Ollama widget shows the server (online, version), the number of installed models, the models in memory with their size and the minutes before Ollama unloads them, and the memory used. Two buttons:

  • Load model: puts the default model in memory, so that the next question is quick (a check mark when it already is);
  • Free memory: unloads every model in memory.

Scenes​

Triggers​

  • Ollama is offline: the server stops answering (two failed readings in a row, 15 seconds apart). Variable: the reason.
  • Ollama is back online: the server answers again. Variable: the length of the outage, in minutes.

Actions​

  • Ask Ollama: asks a model a question.

    • Question: type {{ to insert scene information, e.g. the temperature of a sensor or a value of the trigger;
    • Model: empty for the default model;
    • Instructions: how the model must answer, e.g. "Answer in one short sentence, in English.";
    • Creativity: the model's own setting, or from precise (0) to very creative (1.2);
    • Longest answer: in characters (1000 by default, 10,000 at most). A longer answer is cut at the end of a sentence (or of a word), marked with ….

    Its results are available to the following actions: Answer, Model used, Answer time (seconds) and Answer cut (yes/no). In a Send Message action, type {{ and pick Answer. When the question fails (server offline, model not installed, too long), the answer is empty and the scene goes on.

  • Load an Ollama model: puts a model in memory ahead of time (empty: the default model). Result: the model loaded.

  • Unload an Ollama model: frees the memory of a model (empty: every model in memory). Result: the number of models unloaded.

Scene ideas​

  • Morning briefing: at 7:00, Ask Ollama "In two friendly sentences, sum up the day: outside temperature {{…}}, weather {{…}}.", then send the Answer to your phone.
  • Readable alert: when the washing machine plug drops below 5 W, ask "Write a short and funny message saying the laundry is done." and use Talk on a speaker with the Answer.
  • Free the memory at night: at 1:00, Unload an Ollama model with an empty model, so that the server is free for its other tasks.
  • Watch the server: on Ollama is offline, send yourself a message with the reason.

Troubleshooting​

  • "Ollama does not answer": check the address and the port (11434), that the server is started, and that it listens on the network (OLLAMA_HOST=0.0.0.0). With Docker, the port must be published.
  • "model not installed": the name must be the exact one listed by Test the connection (ollama list on the server). Install it with ollama pull <name>.
  • "Ollama took too long to answer": the first question after a while loads the model from disk. Use Load an Ollama model ahead of time, keep the model in memory longer, raise the longest wait, or choose a smaller model.
  • "access refused": the reverse proxy refused the API key.
  • Problems and ideas: GitHub issues.

Limits​

  • One question, one answer: the action keeps no conversation history between two runs.
  • Text only: no image in the question, no tool calls.
  • Gladys waits two minutes at most for an action: very large models on a processor alone may not fit. The answer is capped at 10,000 characters.
  • Reasoning models (DeepSeek-R1, Qwen3…) are asked to answer without their reasoning; any reasoning left is removed from the answer.
  • The questions and the answers never go to the integration logs.
  • Tested against a real Ollama 0.40 server with small models (qwen2.5 0.5B, SmolLM2 135M).
  • Not affiliated with Ollama.

Configuration settings​

These are the settings Ollama asks for in its configuration screen in Gladys.

SettingTypeRequiredDescription
Your Ollama serversectionNoOllama listens on port 11434. If it runs on another machine than Gladys, start it with OLLAMA_HOST=0.0.0.0 so that it accepts the network.
AddressstringYesIP address or name of the server (port 11434 by default), or the full URL of a reverse proxy.
API key (optional)secretNoOnly for a reverse proxy that asks for a Bearer token: Ollama itself has no password.
Default modelstringNoUsed when a scene action names no model. "Test the connection" lists the installed models.
Longest wait for an answerselectNoA large model on a processor without graphics card can take a minute or more.
Keep the model in memoryselectNoAfter a question or a load. Longer: quicker answers, but the memory stays taken.
Refresh intervalselectNoHow often the state of the server is read.
Language of the feature namesselectNoNames given to the features when the device is added, and texts of the scene events and errors.

How to install Ollama in Gladys​

  1. In Gladys, open Integrations: Ollama appears in the catalog, next to the native integrations, with a community badge.
  2. Click Install. Gladys pulls the Docker image (ghcr.io/guim31/gladys-ollama:1.0.1), starts it in a sandbox isolated from the core, and generates the integration's interface (devices, discovery and configuration).
  3. Open the Configuration screen of the integration, fill in the settings, and save.
  4. You can also install it directly from its repository URL: https://github.com/guim31/gladys-ollama.

Ollama requires Gladys >=5.1.0. The catalog inside Gladys refreshes every hour, so a new version becomes available at most one hour after its release.

Not running Gladys yet? It is free and open source: follow the installation guide to get started.

About external integrations​

Ollama is an external integration: a community integration packaged as a Docker container and published on GitHub, that Gladys installs in one click and runs in a sandbox isolated from its core. It is published and maintained by guim31, not by the Gladys core team.

Subscribe to the Gladys Assistant newsletter

A few emails per month about new releases and project news. Sent by Pierre-Gilles Leymarie, founder of the project. Unsubscribe anytime 🙂
Last updated on