Ghost Browser by Gemma

Free, local browser automation agent powered by Gemma4 27B + Ollama —
ranked 3rd in our 79-task Stealth Bench V1 run, competitive with several paid cloud browser providers.

Experimental local browser automation runner evaluated on Browser-Use Stealth Bench V1.
Results are environment-dependent and not a guarantee of bypassing any specific protection.

---

Stealth Bench V1 Results

Rank	Agent	Tasks	Passed	Accuracy	Type
1	browser-use-cloud	80	58	73%	☁️ Paid cloud
2	anchor	80	55	69%	☁️ Paid cloud
3	Ghost Browser (Gemma4)	79	53	67.1%	💻 Local / Free
4	onkernel	80	53	66%	☁️ Paid cloud
5	browserless	80	42	52%	☁️ Paid cloud
6	local_headful (bu-2-0)	80	40	50%	💻 Local
7	steel	80	39	49%	☁️ Paid cloud
8	browserbase	80	32	40%	☁️ Paid cloud
9	hyperbrowser	80	28	35%	☁️ Paid cloud
10	local_headless	80	2	3%	💻 Local

67.1% with a free local model — no API key, no subscription, no cloud dependency.

Note: other agents ran 80 tasks; our run completed 79. Results are from the same benchmark suite under the same judge LLM, but are not perfectly matched conditions.

Reproducibility

Field	Value
Model	gemma4:latest (27B) via Ollama
Browser	local_headful (Chromium, persistent profile)
OS	Windows 11
Date	2026-05-10
Tasks run	79 / 80
Judge LLM	gemini-2.5-flash (Google AI Studio)
Per-task timeout	1800s
Ollama context	262144 tokens
Parallel tasks	3

---

Failure Analysis (26 failed tasks)

Being transparent about what failed and why:

Category	Count	Sites	Fixable?
Press & Hold antibot	13	walmart, wayfair, yeti, fiverr, samsclub…	Custom JS mouse-hold action needed
LLM timeout	5	x.com, reddit, zillow, bloomberg, lianjia	Retry logic / faster hardware
Cloudflare / Access Denied	3	gamestop, homedepot, crocs	Profile warmup may help
Geo-block (Korean IP)	2	sephora, williams-sonoma	US proxy/VPN
Site down during test	2	belk, davidjones	Not our fault
Login popup blocking content	1	douyin	Task-specific handling

Rough upper bound (if each failure category were independently resolved): timeouts → ~73%, Press & Hold → ~90%+. These are estimates, not guarantees.
Full per-task breakdown in benchmark_results.json.

---

What is Ghost Browser?

Ghost Browser is a customized browser-use agent harness that runs entirely on your local machine:

• Gemma4 27B via Ollama — no API key, no cost
• Reflexion-style retry loop — learns from failed attempts, avoids repeating mistakes
• Guard layer — intercepts bad actions before execution, blocks repeated failures
• Auto-recovery — detects internal browser pages (edge://, chrome://) and navigates back
• Persistent browser profile — reduces bot detection fingerprint over time
• Partial result saving — saves progress even if the agent doesn't fully complete a task

This is a research and benchmarking tool, not a commercial scraping or bypass service.

---

Requirements

• Python 3.11+
• uv package manager
• Ollama with gemma4 model pulled

---

Setup

# 1. Install Ollama and pull Gemma4
ollama pull gemma4

# 2. Clone and install
git clone https://github.com/ddangkong/ghost-browser
cd ghost-browser
uv venv --python 3.11
uv sync

# 3. Run a task
uv run python run.py

---

Architecture

Task Input
    ↓
generate_plan()           # LLM drafts step-by-step plan
    ↓
[Attempt Loop]
    ↓
generate_recovery_plan()  # Reflexion: learns from previous failures
    ↓
Agent.run()               # browser-use agent executes
    ↓
guard_model_actions()     # intercepts each action before execution
  ├─ forbidden_actions    # blocks repeated failed signatures
  ├─ internal URL check   # recovers from edge:// / chrome:// pages
  └─ recovery injection   # forces alternative action on loop detection
    ↓
save_partial_results()    # saves progress if agent doesn't complete
    ↓
[Next Attempt or Done]

---

Join the Development

We're sharing this as an open experiment in local LLM browser automation.
Gemma4 is getting better fast — and we want to grow this together.

If you're interested in:

• Improving challenge interaction handling (Press & Hold, browser compatibility)
• Trying with other local models (Mistral, Llama, Qwen…)
• Building a UI layer (chat interface + live browser view)
• Running benchmarks on your own hardware and comparing results
• Packaging as an MCP server for Claude / Cursor integration

Feel free to open an issue, submit a PR, or just share your benchmark results.
Every run on different hardware/models adds to the picture.

---

Known Limitations

1. SEC EDGAR / DART XBRL viewer → empty DOM (iframe structure, no visible content)
2. Press & Hold challenges → browser-use doesn't support mouse-hold natively
3. Repeated extraction loops — harness-injected extract not always blocked by forbidden_actions
4. Performance is hardware-dependent — Gemma4 27B needs a capable GPU or fast CPU

---

Based on

• browser-use — MIT License, Copyright (c) 2024 Gregor Zunic
• Ollama
• Gemma4 by Google