Overview
Nanobrowser is an AI agent that works inside your own, signed-in browser rather than in a separate one. Per the README and website, you describe a task in the side panel (opened from the toolbar icon or with Alt+Shift+N), and the agent reads pages, clicks, types and moves between tabs, with several tasks able to run in separate tab groups.
The README lists Chrome and Edge as supported and Firefox, Safari and other Chromium browsers (Opera, Arc, Brave) as untested. The code is on GitHub under the Apache License 2.0; the repository was created on 31 December 2024, was last pushed to on 17 November 2026, and the latest release, 0.3.0, is from 15 October 2026, so development is active.
The changelog says version 0.3.0 rebuilt the extension around one agent in the side panel, and version 0.2.0 moved model calls from LangChain to the Vercel AI SDK and removed the Groq, Cerebras and Llama API providers.
How to Get It
Nanobrowser installs like any browser extension: add it from the Chrome Web Store (an Edge Add-ons listing is also linked), then connect a model. No terminal is needed unless you use a local model or build from source.
Route | How to get it |
|---|---|
| Chrome Web Store | Nanobrowser listing (free) |
| Microsoft Edge | Edge Add-ons listing |
| Website | nanobrowser.ai |
| Source code | GitHub (Apache-2.0) |
| Privacy policy | nanobrowser.ai/docs/privacy |
| Build from source | Node.js 24 LTS and pnpm 10 |
Version reviewed: 0.3.0, released on 15 October 2026 per GitHub Releases; the Chrome Web Store listing shows 0.3.0, updated on 16 October 2026. Last upstream update (push) on 17 November 2026. The GitHub release only ships nanobrowser-0.3.0.zip, an unpacked build for developer mode.
# Optional: build from source (Node.js 24 LTS, pnpm 10)
corepack enable
pnpm install
pnpm build
# Local Ollama started by hand: allow the extension
OLLAMA_ORIGINS="chrome-extension://*" ollama serveFeatures
Everything below is a project statement from the README or documentation; none of it was tested by PromptQuorum.
- Side-panel chat with parallel tasks. Several tasks can run in separate tab groups, with image, file and screenshot attachments and results saved to Downloads as CSV or Markdown.
- Ask and Auto modes. Ask mode (the default) shows an approval card before submitting, sending, deleting or uploading; Auto mode skips it. Payments, passwords, codes and access grants always go to you, per the docs.
- Per-site rules. Each site can be set to Allow, Ask or Block; a site rule overrides the mode and matches the exact host.
- Model choice per chat. You pick the model and reasoning effort for each chat, and a Test models button in Settings marks models that lack tool calling.
- Vision when the model supports it. With a vision model the agent can take screenshots and see attached images; without one it reads pages as text.
- Shortcuts, voice and memory. Saved prompt shortcuts triggered with a slash, voice input, persistent memory and automatic summarising of long chats.
- Chat history with export. Chats can be exported to Markdown, and the README lists English, Simplified Chinese and Traditional Chinese interfaces.
- Element IDs. The README says a control layer with no AI model in it gives each page element a stable ID, performs real browser input and checks state before acting, using Chrome's debugger API and Puppeteer.
Privacy and Data Flow
Per its privacy policy, Nanobrowser has no account and no Nanobrowser server that receives chats or pages: your messages, attached files, the content of the tabs the agent works in and, for vision models, screenshots go from your browser to the model provider you choose. With a local Ollama or LM Studio model, the policy says that data stays on your computer. It adds that password, card and one-time-code field values are withheld when recognised, and that logins and cookies stay in the browser.
Two more flows come from the privacy policy. Voice input, if enabled, sends audio to the chosen OpenAI or Google voice model, and the policy says recordings are not kept. Two telemetry switches, both on by default, send an anonymous daily count and daily usage totals (task counts, tools used, error codes) to PostHog in the United States, kept for one year; the policy says this never includes pages, sites visited, messages, files or keys, and both can be turned off in Settings, Data. Uninstalling opens an optional Google Form survey.
The permissions docs add three cautions: opening pages never needs approval, so a page could trick the agent into opening an address that carries your data elsewhere; typing is not gated, so a page can observe what the agent types; and prompt injection is a risk, so watch the agent and use Stop. The policy texts read are dated 13 October 2026 (GitHub) and 18 October 2026 (website); PromptQuorum has not inspected network traffic.
📌Note: These are the project's own statements, not audit results. For work or customer accounts, read the policy first, use a separate browser profile and keep Ask mode on. Automating a website can breach its terms of service, so check them before pointing an agent at a site. This is general information, not security or legal advice.
Pricing and License
The extension is free under the Apache License 2.0; you pay your model provider for what you use. The docs say long tasks and text-heavy pages cost more, and that a local model costs nothing to run.
The website and the Chrome Web Store listing show no sign-up, subscription or paid tier. API keys are stored in the browser and sent only to their own provider, per the docs.
Item | Cost | Notes |
|---|---|---|
| Extension | Free | Apache-2.0, no account |
| Cloud model | Provider rate | Billed by your provider |
| Local model | Free to run | Your hardware, 16k context+ |
Costs depend on the provider and the task. PromptQuorum did not measure cost per task.
Who Should Use It
- People who want an AI agent inside their signed-in Chrome or Edge, with a visible approval step before sensitive actions.
- Ollama or LM Studio users who want to try a browser agent with a local model and accept that the docs advise short, clear tasks.
- Developers who want an Apache-2.0 codebase they can build from source with Node.js 24 LTS and pnpm 10.
Limits and What We Could Not Verify
Nanobrowser is a poor fit if page content must not reach a third-party model, or if a missed action could cost money or data. Several points remain unverified.
- Not a fit: tasks where an unnoticed click does harm; the docs say icon-only buttons, auto-saving switches and drag-and-drop actions can be missed and then run without approval, even in Ask mode.
- Not a fit: Firefox, Safari or other Chromium browsers, which the README lists as untested.
- Unverified: task success rates, speed and cost per task; the sources give no benchmark and PromptQuorum did not run one.
- Unverified: how well local models drive the agent; the docs recommend a model for local use but give no quality figures.
- Unverified: the legal entity behind the project; the privacy policy names only "the Nanobrowser team" with a contact address, and gives no company name, address or country.
- Unverified: whether the privacy statements hold in practice; no independent audit was seen.
Alternatives
Nanobrowser is a browser extension; the alternatives in the same directory segment are a Python library, whole AI browsers and a desktop agent. PromptQuorum has not compared them hands-on.
Nanobrowser
- Noted in sources:
- Chrome and Edge extension; Apache-2.0; cloud or local models
Articles about Nanobrowser (1)
- Local Coding Assistants, Agents & Workflow Tools Compared (2026): IDE Plugins, Agent Frameworks and Autonomous AgentsUpdated November 20, 2026
Also mentioned in:
- Page Assist Review 2026: Browser Sidebar for Local AI ModelsUpdated November 20, 2026
Browser Use
- Noted in sources:
- Python library and CLI for scripting a browser agent; MIT; optional paid cloud
Articles about Browser Use (2)
- Browser Use Review: Open-Source AI Browser Agent for PythonUpdated November 19, 2026
- Local Coding Assistants, Agents & Workflow Tools Compared (2026): IDE Plugins, Agent Frameworks and Autonomous AgentsUpdated November 20, 2026
Also mentioned in:
- Cline Review: The Open-Source Agent That Grew From Claude DevUpdated September 5, 2026
- OpenHands Review: The Open-Source Answer to DevinUpdated September 5, 2026
BrowserOS
- Noted in sources:
- AI browser (Chromium fork) for Mac, Windows and Linux; AGPL-3.0
Articles about BrowserOS (1)
- BrowserOS Review: Open-Source AI Browser With Local Ollama and LM Studio SupportUpdated November 14, 2026
LastBrowser
- Noted in sources:
- Desktop browser for Windows with a built-in Sidekick AI sidebar
Articles about LastBrowser (1)
- LastBrowser Review: Free AI Browser With a Local Ollama SidekickUpdated November 6, 2026
Local Operator
- Noted in sources:
- Desktop agent app that writes and runs code on your computer
Articles about Local Operator (2)
- Local Operator Review: Free Desktop AI Agent for Your ComputerUpdated October 24, 2026
- Local Coding Assistants, Agents & Workflow Tools Compared (2026): IDE Plugins, Agent Frameworks and Autonomous AgentsUpdated November 20, 2026
FAQ
Does Nanobrowser work in Firefox or Brave?
The README lists only Chrome and Edge as supported and says Firefox, Safari and other Chromium browsers such as Opera, Arc and Brave are untested. PromptQuorum has not tried them.
How do I connect Ollama or LM Studio?
Per the docs, set OLLAMA_ORIGINS to chrome-extension://* and restart Ollama, then enter http://localhost:11434 as the server address and click Test models. LM Studio needs Enable CORS in its Developer settings, with the server on port 1234.
Can the agent act without asking me?
In Auto mode it submits, sends, deletes and uploads without asking, but payments, passwords, card numbers, one-time codes and access grants still go to you. A site rule set to Ask forces a prompt even in Auto mode, and Block stops the agent from touching the site.
What changed in version 0.3.0?
Per the changelog, a single agent replaced the planner and navigator, Ask mode became the default, and every page element got an ID tied to that exact element, so clicks target the intended one.
Can several tasks run at once with a local model?
They can start in separate tab groups, but the docs say each Ollama or LM Studio server handles one request at a time, so local tasks queue; servers added through the OpenAI-compatible provider, such as llama.cpp or vLLM, do not queue.
Verdict
Nanobrowser is worth trying if you want an open-source AI agent inside Chrome or Edge, with approval steps and the option of a local model, and you accept that nothing was tested here.
It is Apache-2.0, actively developed (last push on 17 November 2026, release 0.3.0 on 15 October 2026) and free apart from your model provider's charges. The trade-offs: page content goes to the model provider you pick, the docs say some actions can be missed and run unapproved, local models work best on simple tasks, and success rates and speed rest on no data we could check. PromptQuorum has not tested it; start with a separate browser profile and low-risk tasks.
Sources
- Nanobrowser website — product description and install links, checked 20 November 2026.
- Models documentation — providers, tool calling and pricing, checked 20 November 2026.
- Local models documentation — Ollama and LM Studio setup, checked 20 November 2026.
- Permissions documentation — Ask and Auto modes, site rules and warnings, checked 20 November 2026.
- Privacy page — data flows and telemetry, dated 18 October 2026, checked 20 November 2026.
- PRIVACY.md — privacy policy on GitHub, dated 13 October 2026, checked 20 November 2026.
- Changelog — versions 0.1.9 to 0.3.0, checked 20 November 2026.
- GitHub repository and README — license, install steps and last push, checked 20 November 2026.
- Release 0.3.0 — latest release, published 15 October 2026.
- Chrome Web Store listing — version 0.3.0, updated 16 October 2026, free, checked 20 November 2026.
