Skip to main content

Ollama in Sibyl

Choose Settings → AI models → Local → Ollama. Sibyl checks the endpoint shown in the form and lists installed models and models currently loaded in memory. Choose an installed model, then Save as workspace default. Presets and custom IDs remain available; a preset is not proof that a model has been downloaded.

Check status probes the native version, model catalog and loaded-model endpoints without running inference. A reachable server with an installed model may still have insufficient resources or unsuitable tool-calling behavior; these checks do not evaluate model answers.

If Ollama is installed locally but stopped, Start Ollama launches its installed CLI using ollama serve. Only a plain loopback HTTP endpoint on the machine hosting Sibyl's Python backend can be started. The browser cannot supply commands, executable paths or extra process arguments. Startup may take a moment; the panel checks a starting process automatically. Stop local server appears only for a process started by this Sibyl instance. Existing desktop/system Ollama processes are reused and must be stopped in Ollama itself. Sibyl stops its own child when the backend shuts down.

Context window​

Expand Advanced model settings to set the context window in tokens. It holds instructions, the specialist directory, tool results, conversation and reply. The default is 32,768; choose a value supported by the model and available memory. Larger windows use more memory. Sibyl's full specialist directory needs more room than a basic chat, so small windows can prevent coordinator requests. An input that cannot fit produces setup guidance. Save the profile to apply the value to new requests; running turns keep their previous configuration.

Sibyl uses ProtoLink's capability-based native tool selection for Ollama. Model tool-following and answer quality still vary. See acceptance observations for the actual local-model checks.

Installation and models​

Install Ollama on the Python backend machine, which may differ from your browser's computer. Use the official download page and quickstart. Open Ollama after installing, then check status again. Remote servers can be checked, but cannot be started from Sibyl.

Model installation is separate. The panel displays a command such as ollama pull qwen3:8b for the selected model; run it yourself in a terminal on the Ollama machine. Sibyl does not install software, download models or silently change the selected AI profile when starting a server. See the Ollama CLI.

Implementation lives in src/sibyl/local_models/ollama.py. POST /api/v1/settings/ollama/check, /start and /stop accept only base_url and model, under the normal workspace access controls. Saved credentials are used only at their saved endpoint. Native contracts: version, model catalog, loaded models. Tests use fake HTTP and process boundaries; they never start the developer's Ollama installation.

Optional Agent models settings let Sibyl and each specialist use a different saved provider/model without changing the workspace default. See per-agent setup.