# OpenBrowse
**The open-source Browser Use Cloud alternative.** Self-host AI browser agents on a Raspberry Pi or any VPS, drive them through the same v3 REST API the `browser-use-sdk` already speaks, and watch every run live in a real browser. Built on top of the [Browser Use](https://github.com/browser-use/browser-use) SDK. It's cheaper, faster, and more reliable than BU Cloud.
[openbrowse.co](https://openbrowse.co)
---
## Benchmarks
Given the same real-world extraction task (a careers page with 14 records behind an embedded, cross-origin board, requiring full schema output):
| Runtime | Model | Reasoning | Time | Tokens | LLM cost | Records |
| --- | --- | --- | ---: | ---: | ---: | --- |
| BU Cloud | claude-sonnet-5 | high | 2m 36s | 859k | $0.78 | 14/141 |
| **OpenBrowse** | **gpt-5.6-terra** | **none** | **1m 47s** | **202k** | **$0.24** | **14/14** |
| **OpenBrowse** | **claude-sonnet-5** | **high** | 4m 02s | **242k** | **$0.40** | **14/14** |
1 Extracted successfully, though some returned fields (e.g. job seniority) were hallucinated when not shown on the page.[All runs, every model tried, and the task specification](https://openbrowse.co/benchmarks).
## Why OpenBrowse over BU Cloud?
| | BU Cloud | OpenBrowse |
| --- | --- | --- |
| Hosting | Managed, per-task pricing | Your hardware, you pay only LLM tokens |
| How it works | Code-first: the agent scripts its way through pages | Visual-first: the agent opens real tabs you can watch live, like a human working |
| Bulk page reads | One page at a time | `read_pages` opens whole listings in parallel tab waves, one step |
| Structured output | Schema-validated | Schema-validated, plus a live answer store with a completeness gate the agent must pass before finishing |
| Anti-hallucination | Often fills fields the page never shows | On-screen data first, enriched from the page's own structured data (JSON-LD, APIs), never guessed: values without evidence are refused at the store boundary |
| Profiles | Cloud profiles | Import your BU Cloud profiles (cookies and localStorage) with one command |
| Live view | Replay | Real-time VNC of the actual browser, a step feed with the model's reasoning, and an IDE-style code tab that streams the agent's sandbox scripts live as they're written |
| API | v3 REST | The same v3 REST surface: point `browser-use-sdk` at your box and change nothing but `baseUrl` and `apiKey` |
[The full dimension-by-dimension comparison](https://openbrowse.co/vs/browser-use-cloud).
### โก See it in action
Here's a snippet of our benchmark run, with `claude-opus-5` taking agentic actions across parallel tabs while its code streams into a live IDE.
## Model providers
### Recommended models
1. **For most use cases**, `gpt-5.6-terra { "reasoningEffort": "none" }`, `gpt-5.6-sol { "reasoningEffort": "none" }` and `claude-sonnet-5 { "reasoningEffort": "high" }` all strike a great balance of reliability, accuracy, and cost.
2. **For intense workflows**, use `claude-opus-5 { "reasoningEffort": "medium" }` or `gpt-5.6-sol { "reasoningEffort": "none" }` โ both are great options, but watch out for token burn.
3. **On a budget?** Use `gpt-5.6-luna { "reasoningEffort": "max" }` with a tightly focused prompt. It might take a while, and it's more prone to hallucinations (especially with broad prompts), but the actual extractions are still great quality.
### Comparisons
The same real-world extraction task (a careers page with 14 records behind an embedded, cross-origin board, full schema output) run against BU Cloud and against OpenBrowse on a Raspberry Pi 5 (16GB) without concurrency, ordered cheapest to most expensive:
| Runtime | Model | Reasoning | Steps | Time | Tokens | LLM cost | Records |
| --- | --- | --- | --- | ---: | ---: | ---: | --- |
| OpenBrowse | **gpt-5.6-luna** | **max** | 36 | 17m 03s | 1.08M | **$0.22** | 14/14 |
| OpenBrowse | **gpt-5.6-terra** | **none** | 11 | **1m 47s** | **202k** | **$0.24** | 14/14 |
| OpenBrowse | **gpt-5.6-sol** | **none** | **8** | **2m 03s** | **136k** | **$0.33** | 14/14 |
| OpenBrowse | **claude-sonnet-5** | **high** | 10 | 4m 02s | **242k** | **$0.40** | 14/14 |
| OpenBrowse | **claude-sonnet-5** | **none** | **9** | 5m 18s | **237k** | **$0.51** | 14/14 |
| OpenBrowse | **gpt-5.6-terra** | **high** | 17 | 5m 05s | **434k** | **$0.66** | 14/14 |
| BU Cloud | claude-sonnet-5 | high | 10 | 2m 36s | 859k | $0.78 | 14/141 |
| OpenBrowse | gpt-5.6-sol | medium | 16 | 5m 24s | **339k** | $0.90 | 14/14 |
| OpenBrowse | claude-opus-5 | medium | 15 | 3m 56s | **398k** | $1.32 | 14/14 |
| OpenBrowse | claude-opus-5 | none | 17 | 4m 53s | **480k** | $1.62 | 14/14 |
1 Extracted successfully, though some returned fields (e.g. job seniority) were hallucinated when not shown on the page.
OpenAI and Anthropic models are generally at their best at the opposite ends of the reasoning dial. For example, OpenAI's GPT-5.6-Terra performs better with less reasoning, spending less time planning ahead and more time reacting to the page in front of it, while Anthropic's 5-series Claude models lean towards rabbit holes and need reasoning time to refocus on the goal.
### Common questions
What's the difference between "thinking" and "reasoning"?
> OpenBrowse separates two kinds of reasoning. **Browser thinking** is our way of describing how the platform works in "steps" (the ๐๏ธ see / ๐ plan / โก๏ธ next / ๐ญ thinking cards in the live feed), so it can't be disabled.
>
>**Model reasoning** is different: it's the Chain-of-Thought reasoning provided by LLM providers (e.g. Anthropic's extended thinking, OpenAI's reasoning effort), and can be controlled per session by changing `reasoningEffort` in the API. Values are validated per model at runtime.
See all supported models
> - OpenAI: `gpt-5.6-sol`, `gpt-5.6-terra`, `gpt-5.6-luna`
> - Anthropic: `claude-mythos-5`, `claude-fable-5`, `claude-opus-5`, `claude-sonnet-5`, `claude-opus-4.8`, `claude-opus-4.8[1m]`, `claude-opus-4.7`, `claude-opus-4.7[1m]`, `claude-opus-4.6`, `claude-opus-4.6[1m]`, `claude-sonnet-4.6`, `claude-sonnet-4.6[1m]`
> - Google: โ ๏ธ Coming soon
## Quick start
Install OpenBrowse with your favorite package manager:
```bash
# uv
uv tool install openbrowse
openbrowse start
# pipx
pipx install openbrowse
openbrowse start
# venv
source /bin/activate
pip install openbrowse
openbrowse start
```
To install from source, clone the repo and use uv directly:
```bash
git clone git@github.com:lujstn/openbrowse.git
cd openbrowse
uv sync
uv run openbrowse serve
```
Then, open `http://localhost:8420` in a browser. You'll be guided through everything you need to run OpenBrowse. [View the installation guide](https://openbrowse.co/docs/installation) for more details.
## Exposing it to the web
OpenBrowse runs on plain Python + Chromium and can be easily port forwarded. To reach it from outside that box without opening ports, put it behind [Tailscale](https://tailscale.com/):
```bash
# private access from your own devices
tailscale up
# or expose the API publicly over TLS with Tailscale Funnel
sudo tailscale funnel --bg 8420
```
## Features
- **v3-compatible REST API**: sessions, structured output schemas, cost caps, live URLs.
- **Visual, tab-based browsing**: parallel foreground tab waves for bulk reads; a code tab shows when the agent runs a script; everything visible over VNC.
- **Schema answer store**: every write validated live against your JSON Schema, coverage tracked per field, a completeness gate before `done`, and mark-absent semantics for data a site genuinely does not publish.
- **Grounding guards**: shell-read detection with automatic in-frame retry, evidence-checked enum writes, URL fields validated as absolute http(s) links at the store boundary, honest failure over invented data.
- **Browser-like profiles**: sessions share a profile's logins the way a browser does. Each site's data goes into the browser's own storage before the first page loads, so a page only sees its own site's data; each session merges back only what it changed, even when several share one profile; and the least recently used sites are evicted within fixed limits, so profiles never grow without bound. A site-data page in the dashboard lists and deletes them per site.
- **Profile import**: bring BU Cloud profiles (cookies + localStorage) via CLI or the dashboard.
- **Dashboard**: live session feed with model reasoning, per-step costs, JSON export (full / steps / output-only), profile management.
- **CAPTCHA solving**: optional CapSolver integration.
- **Multi-provider**: Anthropic and OpenAI models behind one API, with per-provider repair layers for each family's failure modes.
## Citation
This project is licensed under the MIT License. If you use OpenBrowse as part of your research or project, please cite:
```bibtex
@software{openbrowse2026,
author = {Johnston Kurilov, Lucas},
title = {{OpenBrowse}: Self-hosted AI browser agents},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.21986248},
url = {https://github.com/lujstn/openbrowse}
}
```
[](https://doi.org/10.5281/zenodo.21986248)
