Localhost belongs to the Harness host.

Prove the host before the form.

Run both commands in the same machine, VM, container, or WSL environment that runs DeepSeek Harness:

curl http://127.0.0.1:11434/v1/models
ollama list

The first command proves OpenAI-compatible discovery is reachable. The second gives the exact model ID, including its tag.

Enter the complete route.

1. Open Settings → Models → Add a custom provider.
2. Use Provider ID ollama. The ID must be lowercase and becomes permanent.
3. Use base URL http://127.0.0.1:11434/v1.
4. Select openai-completions.
5. Leave API key blank. Local Ollama does not require one by default.
6. Add a model row. Use the exact installed ID, such as qwen2.5-coder:7b.

Browser localhost is not the route.

The browser only edits settings. Model discovery and inference run from the Harness host. If Harness runs in Docker, a VM, WSL, or another machine, 127.0.0.1 points inside that environment—not necessarily to the Ollama process on your laptop.

Use an address reachable from the Harness host. Keep Ollama bound away from untrusted networks; changing reachability should not silently widen exposure.

Route the symptom.

• Create remains disabled: add at least one model row and check the Provider ID format.
• Fetch available models fails: run the direct /v1/models request from the Harness host.
• Discovery works but the first turn fails: verify the exact model tag and inspect the provider response.
• Laptop curl works but container Harness fails: test from inside the container; loopback is different.

Primary evidence.

• Custom-provider creation gates in the Web client
• Official model-provider guide
• Original Ollama configuration report
• Ollama OpenAI compatibility documentation