What shipped

Agent

• A single agent replaces the planner and navigator.
• Several tasks can run side by side, each with its own chat and tab group.
• Long chats are summed up automatically, and your own messages are kept in full.
• The agent can upload your files to a page and save results, such as CSV or Markdown, to Downloads.

Models

• Any model with tool calling works: Anthropic, OpenAI, Gemini, OpenRouter, Grok, DeepSeek, Mistral, Ollama, LM Studio and any OpenAI-compatible endpoint.
• Pick the model and reasoning effort for each chat.
• The Test models button in Settings checks which of your models work.

Interface

• Attach images, text files or a screenshot of the tab.
• Start a saved prompt by typing /.
• Speak your task with OpenAI’s or Gemini’s voice model.
• Custom instructions tell the agent what to keep in mind in every chat.
• Edit and rerun any message.
• The interface is available in English, Simplified Chinese and Traditional Chinese, in light or dark.

History

• Export a chat as Markdown.

Safety

• Ask mode, the default, checks with you before the agent submits, sends, deletes or uploads. Auto mode goes ahead without asking.
• Payments, passwords, codes and access grants are always handed to you.
• Set any site to Allow, Ask or Block.

Privacy

• Nanobrowser sends a daily user count and anonymous usage counts, with no ID attached. They never include your pages, sites, messages or keys.
• Turn both off in Settings → Data.

Reliability

• Each element has an ID tied to that exact element, so the agent clicks the one it means.
• The agent clicks and types like a real user, so pages respond as they would to you.
• Every action is sent once and reports whether the page responded.

Upgrading

• Your providers and API keys, default model, saved prompts and blocked sites move to 0.3.
• Earlier conversations stay in your browser until you export or delete them in Settings → Data.

Models

• Structured responses are more reliable with Anthropic, Gemini 3, DeepSeek and GPT-6 models.
• Models that support reasoning now use a reasoning effort setting instead of temperature and top P.
• Default models are updated for OpenAI, Anthropic, Gemini, Grok, DeepSeek and Ollama.
• Ollama now connects through its OpenAI-compatible API. Existing Ollama settings keep working.

Upgrading

• Groq, Cerebras and Llama API are no longer built in, and their saved settings and keys are deleted. To keep using one, add it as a custom OpenAI-compatible provider and choose a new model for any agent that used it.
• The Minimal reasoning effort option is gone. A saved Minimal setting now uses Low.

Developers

• The extension is now a single package built with WXT, with instant page reloads and automatic extension reloads during development.
• LLM calls run on the Vercel AI SDK instead of LangChain.
• Building requires Node.js 24 LTS and pnpm 10.

Models

• Gemini responses are read correctly, including through OpenAI-compatible endpoints.
• Nanobrowser recovers more often when a model returns a badly formatted response.
• Default Gemini and Grok models are updated.

Interface

• Portuguese (Brazil) is now a supported language.
• Attach text files to a chat.
• Settings use clean icons instead of emoji.

Safety

• The side panel only accepts connections from the extension itself.

Reliability

• Fixed an infinite loop on pages with live elements and sandboxed iframes, such as amazon.com.
• Finding clickable elements no longer fails on deeply nested pages.

Models

• gpt-5-chat-latest, a non-reasoning model, is now available.
• Model parameters are now saved correctly in settings.

Safety

• Instructions hidden in web pages are filtered out and treated as untrusted, to guard against prompt injection.

Privacy

• Anonymous usage analytics are on by default to help guide improvements. Turn them off in Settings.

Agent

• The planner now confirms when a task is done, so tasks no longer need a separate validation step.

Models

• GPT-5 and GPT-5 mini are supported, with a new Minimal reasoning effort option.
• Llama API is now a built-in provider.
• Default models are updated, including GPT-5 and Claude Opus 4.1.

Interface

• The interface follows your browser’s language, starting with Traditional Chinese.

Reliability

• Frames that can’t be accessed, such as cross-origin frames, no longer interrupt a task.

Models

• Models with the same name from different providers no longer get mixed up.

History

• Replay a past task from your history. It’s experimental and off by default; turn it on in Settings.

Reliability

• Scrolling inside pages and scrollable elements is faster and more accurate.

Models

• Groq and Cerebras are now built-in providers.

Interface

• Speak your task with the new mic button.
• New users get a guided setup instead of an “API key not found” error.

Interface

• Messages show clear separators between sections.
• Element highlighting turns off when Use Vision is off.

History

• Bookmark favorite tasks in the history list to start them quickly.

Safety

• A basic firewall lets you allow or deny specific URLs.
• Text from web pages is treated as data, not instructions, to guard against prompt injection.

Reliability

• Stopping a task cancels the model request right away.
• The agent no longer keeps scrolling past the top or bottom of a page.

Agent

• The agent can close tabs and pause for a few seconds.

Models

• Azure OpenAI and OpenRouter are now built-in providers.
• Default models are updated for every provider.

Agent

• The agent can read and choose options in dropdown menus.
• On a Mac, the agent uses Command where it would press Ctrl.

Interface

• When typing with a Chinese or Japanese input method, pressing Enter to confirm a word no longer sends the message.
• Long messages wrap instead of overflowing.
• The chat input clearly shows when it is disabled.

Reliability

• Fixed a bug that affected Google search.
• A task now stops after three failed checks instead of occasionally looping forever.
• “Access denied” (403) errors explain more about what went wrong.

Models

• DeepSeek’s official API is available as a built-in provider.
• Fixed model selection being disabled when only one model is available.

Interface

• Nanobrowser is now on the Chrome Web Store.

Models

• Claude 3.7 and Grok are now supported.
• Built-in Ollama support lets you run local models such as deepseek-r1, qwen2.5 and mistral-small.
• Add custom OpenAI-compatible providers, such as OpenRouter.
• Set temperature and top P for each agent.

Interface

• Model settings are redesigned, so adding and configuring providers is easier.
• Dark mode follows your system setting.

Models

• Gemini 2.0 is now supported.
• Fixed errors when using GPT-4o mini.

Agent

• A multi-agent system plans and carries out complex web tasks from plain-language instructions.
• The agent runs in your browser, not on a cloud server.

Models

• OpenAI and Anthropic models are supported.

Interface

• A side panel shows what the agent is doing in real time.