Local model, visible approval

Choose a model in Ollama. Use it inside WebBrain.

WebBrain can use Ollama’s local OpenAI-compatible endpoint today. Configure it in Settings with the official Ollama release, or try the preview launch handoff that sends the model, endpoint, and context window to WebBrain for one-click approval.

The launch command is still a preview

ollama launch webbrain is not part of upstream Ollama builds yet. The normal Settings path works with the official release; the one-command handoff currently requires the linked preview branch.

Two setup paths

Use the path that matches your Ollama build

Path	Use it when	What you do
Manual provider setup	You installed an official Ollama release.	Start Ollama with extension origins allowed, then enter its local endpoint and model in WebBrain Settings.
Launch handoff	You built the current preview branch and want the shortest setup.	Run `ollama launch webbrain --model <model>`, review the browser prompt, and approve the provider change.

WebBrain and Ollama logos joined by a heart — Ollama serves the model locally; WebBrain gives that model the browser-agent tools.

Works with upstream Ollama

Configure the official Ollama release manually

Install Ollama and pull a capable model. Choose a model that follows tool calls reliably. Agent runs work best with at least a 16k-token context window.
Allow browser-extension origins. Quit another Ollama desktop process if it already owns port 11434, then start the server from a terminal with the origin allowlist below.
Open WebBrain Settings. Go to Providers & Models, find Ollama (Local), and open the card.
Enter the endpoint and model. Use http://localhost:11434/v1, then select the exact model you pulled.
Save and Test Connection. Use Load models when you want WebBrain to read the available model IDs and detect the live context window.
Set Active. Start in the Mid prompt tier. Use Compact only when the model’s context is constrained.

ollama pull <model>
OLLAMA_ORIGINS="chrome-extension://*,moz-extension://*" ollama serve

Terminal requests can succeed while WebBrain gets 403

curl does not send a browser-extension origin. If Ollama works in a terminal but fails in WebBrain, restart the Ollama server with OLLAMA_ORIGINS rather than changing the provider URL.

Preview · fewer settings

Try the Ollama launch handoff

The preview command opens a WebBrain launch URL containing the selected model, loopback /v1 endpoint, and detected context window. The extension shows those values in a confirmation dialog before it changes anything.

git clone https://github.com/esokullu/ollama.git
cd ollama
git switch codex/ollama-webbrain-launch-handoff
cmake -S . -B build -G Ninja -DOLLAMA_MLX_BACKENDS=
cmake --build build --parallel 8

OLLAMA_ORIGINS="chrome-extension://*,moz-extension://*" ./ollama serve

Keep the server running. In another terminal, replace the example model with one available in your local Ollama installation:

./ollama launch webbrain --model qwen3.5:9b

Let the command open webbrain.one. The extension only accepts this handoff on the WebBrain Ollama launch page.
Read the confirmation. Check the model, loopback provider URL, and context window.
Approve the change. WebBrain updates the local Ollama provider and makes it active.
Open the side panel. The active provider should now read Ollama (Local).

WebBrain confirmation dialog showing the Ollama model, provider URL, and context window — The handoff remains a user decision: nothing is saved until you approve the browser prompt.

The handoff stays local

WebBrain only accepts a loopback Ollama host, an HTTP(S) URL without embedded credentials, and the OpenAI-compatible /v1 endpoint. The model name and context window are normalized before the provider is saved.

Pick a model and context that can drive an agent

Start at 16k context or higher. That is the practical minimum for the system prompt, browser tools, page content, and a useful conversation history.
Tool-use training matters. A model that answers normal chat well may still produce malformed or missing tool calls.
Use Mid first. It exposes the common browser toolset without the full prompt cost. Compact can help around 8k; 4k is too small for reliable runs.
Vision is model-specific. Only enable vision when the exact Ollama model you serve accepts images.
Test the live model. Test Connection and Load models can read Ollama’s runtime information and refresh the detected context window.

Troubleshooting by symptom

Symptom	Likely cause	Fix
WebBrain reports 403	Ollama rejected the extension origin.	Restart the server with `OLLAMA_ORIGINS="chrome-extension://,moz-extension://"`.
Connection refused	Ollama is stopped, or WebBrain is using the wrong port.	Run `ollama serve` and keep the base URL at `http://localhost:11434/v1`.
Model not found	The saved model ID does not exactly match Ollama’s catalog.	Run `ollama list` or use Load models, then copy the complete model name including its tag.
The model talks instead of using browser tools	The model is weak at tool use, or the context/prompt tier is too constrained.	Try a tool-trained model, increase context, or move from Compact to Mid.
`ollama launch` is unknown	You are running an upstream Ollama build.	Use manual provider setup, or build the linked preview branch.
The launch page opens but nothing prompts	The WebBrain extension is missing, disabled, or outdated.	Enable the current extension on `webbrain.one`, reload the launch page, and run the command again.

For the announcement, screenshots, and current preview rationale, read WebBrain now has an Ollama launch handoff.