Models & Providers
Configure AI models — run locally, bring your own key, or pick from 130+ providers via models.dev.
OpenBrowse is model-agnostic. Run entirely on-device with no API key, plug in your own key for a frontier model, or pick from 130+ providers and 4,000+ models via the bundled models.dev catalog.
Browse all providers
Head to the Providers directory for the full live list of every provider OpenBrowse can talk to.
Local — no API key
Run fully on-device. Page content and conversations never leave your machine.
| Provider | Setup |
|---|---|
| Chrome Built-in AI (Gemini Nano) | None — works out of the box on Chrome 128+ |
| WebLLM | Pick a model in the model picker; weights download to your browser cache |
| Ollama | Run an Ollama server on your machine |
Cloud — bring your own key
Curated setup guides for the most common providers:
| Provider | Setup |
|---|---|
| OpenAI | API key |
| Anthropic | API key |
| Google Gemini | API key |
| OpenAI-Compatible | API key + base URL (Groq, Together, OpenRouter, DeepSeek, LM Studio, …) |
Beyond these curated guides, OpenBrowse can use any provider in the models.dev catalog — currently 130+ vendors covering hundreds of OSS and proprietary models. The model picker auto-discovers them; for providers that don't ship with first-class SDK support, the OpenAI-Compatible adapter handles the long tail.
OpenBrowse ships with first-class adapters for Anthropic, OpenAI, Google, xAI, Mistral, OpenRouter, and the OpenAI-Compatible umbrella. Other providers in the catalog are routed through one of those adapters using the npm field models.dev publishes for each provider.
How the catalog works
OpenBrowse mirrors the pattern Zed and OpenCode use: a vendored snapshot of https://models.dev/api.json ships with the extension so the model picker works offline and on first run, and a background refresh keeps the cache up to date.
Load order:
- In-memory cache (per service-worker session)
chrome.storage.localcache (refreshed every ~60 minutes)- Bundled snapshot —
apps/extension/src/registry/models-dev/snapshot.json - Live fetch from
https://models.dev/api.json
The snapshot is regenerated via pnpm refresh:models (Zod-validated, sorted-key JSON for clean diffs). Day-to-day, you don't have to think about any of this — the picker stays current automatically.
Configuration
To switch models or add an API key:
- Open the side panel (
⌥I) and click the model picker at the bottom — or go to Settings → Models. - Pick a provider (use the search bar to filter the 130+ available).
- For cloud providers, paste your API key (and base URL if applicable).
- Choose a default model from the list.
API keys are stored in your browser's extension storage and never sent to OpenBrowse servers.
Picking the right model
Different tasks benefit from different models:
- Quick tab tidying / summaries — a fast local model (Gemini Nano via Chrome AI, or a small Llama via WebLLM) is plenty.
- Routine browsing automation — mid-tier cloud models (GPT-4-class, Claude Sonnet, Gemini Flash) hit the sweet spot for tool-calling reliability.
- Complex multi-step tasks — frontier reasoning models when you need the agent to plan across many tools and pages.
Because OpenBrowse uses the Model Context Protocol (MCP) and relies on tool-calling, models that score well on tool-use benchmarks (Anthropic's Claude family, OpenAI's GPT family, Google's Gemini family) are the most reliable. Smaller models work well for narrow tasks but may struggle with long autonomous tool chains.