The model behind the panel

Choose the setup that matches your patience, privacy, and budget.

WebBrain works with its built-in cloud option, direct cloud APIs, model routers, and local OpenAI-compatible servers. You only need one active provider.

The quick choice

Fewest steps

WebBrain Cloud

Built in, active by default, and includes a free daily allowance. No model server or API key setup.

Private & offline-capable

Local model

Use llama.cpp, Ollama, LM Studio, Jan, vLLM, SGLang, or LocalAI. Your model requests stay with the server you run.

Direct relationship

Cloud API

Bring a key for OpenAI, Anthropic, Gemini, Mistral, DeepSeek, xAI, Groq, and other supported providers.

Broad catalog

Router

Use one account to choose among many models. OpenRouter and Hugging Face are useful when you experiment often.

Configure a provider

Provider list with Local, Cloud and Router filters plus the active WebBrain Cloud card — Use filters to shorten the list. The active provider remains visible so you always know what the panel is using.

Find the provider. Use Local, Cloud, Router, or the search field.
Open its card. Non-active cards are collapsed to keep the list readable.
Enter the values. Local servers usually need a URL and model name. Cloud providers also need an API key.
Save. This stores the card’s current fields in the browser.
Test Connection. Fix any reported URL, key, model, or server problem before continuing.
Set Active. Only after the test succeeds. A saved provider is not necessarily the provider the side panel is using.

What the fields mean

Field	Meaning	New-user advice
Server / Base URL	The OpenAI-compatible root endpoint or provider API root.	Use the exact value shown by your local app or provider. Do not paste a chat webpage URL.
API key	A credential sent to that provider.	Local servers normally do not need one unless you enabled authentication.
Model	The model identifier the server expects.	Spelling and prefixes matter. “Model not found” usually means this field does not match the provider catalog.
Context window	How much prompt, history, page content, and tool schema fits in one request.	16k tokens is the usable target for reliable local runs. 8k can work with Compact; 4k is too small.
Supports vision	Whether the selected local or open catalog model accepts images.	Enable only when the exact served model can process image input.
Prompt tier	How large a prompt and normal browser-tool set WebBrain exposes.	Start with Mid for local models. Use Compact for small contexts and Full for capable models.
Input / output cost	Optional price estimates per million tokens.	Fill these when you want WebBrain’s cost allowances to estimate providers that do not report usage cost.

Local models: the extra moving parts

Server	Typical base URL	API key
llama.cpp	`http://localhost:8080`	Usually none
Ollama	`http://localhost:11434/v1`	None
LM Studio	`http://localhost:1234/v1`	None
Jan	`http://localhost:1337/v1`	Optional
vLLM	`http://localhost:8000/v1`	Optional
SGLang	`http://localhost:30000/v1`	Optional
LocalAI	`http://localhost:8080/v1`	Optional

Load models asks supported local servers which models are currently available. For llama.cpp, Ollama, and LM Studio, Test Connection or Load models can also detect the live context window when the backend reports it.

Browser extensions still need permission to reach local servers

If the server works in a terminal but not in WebBrain, check the server’s CORS/origin setting. Ollama commonly needs OLLAMA_ORIGINS="chrome-extension://*,moz-extension://*" before ollama serve.

Prompt tiers in plain language

CompactSmallest prompt and action-tool surface. Designed for constrained local models and roughly 8k contexts.

MidCommon browser tasks, iframes, downloads, scheduling, and form verification. The best local-model starting point.

FullAdds advanced UI and DOM fallbacks such as hover, drag/drop, frames, and shadow DOM tools.

Cloud providers are forced to Full. Local providers default to Mid unless you choose another tier. OpenRouter exposes the tier because its catalog includes models with very different capabilities. Dev mode is blocked for Compact-tier providers.

When Test Connection fails

Connection refused: start the local server and confirm the port.
CORS or origin error: allow the WebBrain extension origin in the local server.
Unauthorized: paste a current key for the same provider as the base URL.
Model not found: copy the exact model ID from the provider or use Load models.
Requests time out: increase the General request timeout for large local models on CPU.
Runs fail after a few turns: confirm the real context window; lower the prompt tier or load the model with a larger context.

Test before you switch

Keep the working provider active while you configure a new one. Save and test the new card first; only then click Set Active.