Ollama

The Ollama provider points the claude CLI at a local Ollama daemon. It’s the simplest non-Anthropic setup: no API key, no gateway translation, and everything stays on your machine. Useful for air-gapped work, privacy-sensitive code, or running offline.

Ollama is anthropic-compatible — it talks Claude’s wire format directly, so the claude CLI doesn’t even know it’s not talking to api.anthropic.com.

Prerequisites

Make sure the Alternative Claude Code backends experimental flag is on.
Install Ollama:
- macOS: brew install ollama or download from ollama.com/download.
- Linux: curl -fsSL https://ollama.com/install.sh | sh.
- Windows: download the installer from ollama.com/download.
Start the daemon. The Ollama installer typically registers a system service that starts on boot. To run it manually: ollama serve.

By default Ollama listens on http://localhost:11434. Confirm it’s reachable: curl http://localhost:11434/api/tags should return JSON.

Pull at least one model

Ollama doesn’t ship models — you pull them on demand. Pick something tool-tuned for agent workflows:

ollama pull qwen3-coder         # solid coding model
ollama pull llama3.1:8b         # general-purpose, smaller
ollama pull qwen2.5-coder:7b    # smaller coding model

Pulled models are cached under ~/.ollama/models/. Run ollama list to see what you have.

Configure in Claudette

Open Settings > Models. The ollama backend appears in the list.
The base URL is pre-filled to http://localhost:11434. Change it if your daemon listens elsewhere (e.g. on a LAN host).
(Optional) Add a bearer token in the API key field if you’ve put Ollama behind an authenticating proxy. Most local installs don’t need this.
Click Test — Claudette pings the daemon to confirm it’s reachable.
Click Refresh models — Claudette queries Ollama’s /api/tags endpoint and discovers everything you’ve pulled. The chips appear under Discovered models.
(Optional) Set a default model.
Toggle Enabled on. The provider is now selectable in the chat header.

Picking a session model

Once enabled, the model picker in the chat header includes Ollama models alongside Claude. Pick one and the next turn runs through Ollama.

The chat header hides the Effort, Fast mode, and 1M-context toggles when an Ollama model is active — AgentBackendConfig::builtin_ollama() declares those capabilities as false. Extended thinking stays available because Ollama supports it on models that have a thinking mode (like qwen3-coder).

Tool use and vision

Tool use and vision are declared true for Ollama in the capability struct, but actual support depends on the model:

Tool use — works on tool-tuned models. qwen3-coder and llama3.1 handle Claude-style tool calls; smaller / non-tuned models may ignore tool definitions.
Vision — works on multimodal models like llama3.2-vision. Plain text models drop image attachments silently.

If a model misbehaves, try switching to a different one before assuming Claudette is broken.

Performance notes

Ollama runs on CPU by default; GPU acceleration depends on your platform (Metal on macOS, CUDA on Linux with the right drivers).
Large models (70B+) need substantial RAM. Watch Activity Monitor / htop while a turn runs.
Concurrent agents amplify resource pressure. Keep parallelism low (1–2 workspaces at a time) when running locally.

Troubleshooting

Test fails with “connection refused” — the daemon isn’t running, or it’s bound to a different host/port. Run ollama serve and confirm curl localhost:11434/api/tags works first.

Refresh models returns nothing — you haven’t pulled any models yet. Run ollama pull <model> and click Refresh again.

Agent hangs on first turn — the model is loading from disk into RAM/VRAM. First-turn latency on a cold model can be 10–60 seconds depending on size. Subsequent turns reuse the loaded weights and are much faster.

Tool calls get ignored — the model doesn’t support tool use, or the prompt is being interpreted as plain chat. Try a tool-tuned model (qwen3-coder, llama3.1).