<div align="center">
<h1>🛡️ DeepCloak</h1>
### Local-first research with explicit fetch decisions and evidence.
A CLI and MCP wrapper around local-deep-research and CloakBrowser. It retries selected failed fetches in a browser and records the result. Success depends on the source and its access controls.
[](https://pypi.org/project/deepcloak/)
[](https://github.com/Mrbaeksang/deepcloak/actions/workflows/ci.yml)
[](LICENSE)
[](pyproject.toml)
[](#-use-it-from-an-ai-agent)
[](CONTRIBUTING.md)
[](https://github.com/Mrbaeksang/deepcloak/stargazers)
[](https://youtu.be/p5ompjDLzaI)
**English** · [한국어](README.ko.md) · [简体中文](README.zh-CN.md)
[Quickstart](#-quickstart) · [How it works](#-how-it-works) · [Use from an agent (MCP)](#-connect-it-to-your-agent-mcp) · [Why we built it](#-why-we-built-this) · [Changelog](CHANGELOG.md)
### [▶ Live demo — deepcloak.vercel.app](https://deepcloak.vercel.app) · [▶ Watch on YouTube](https://youtu.be/p5ompjDLzaI)
[](https://youtu.be/p5ompjDLzaI)
</div>
---
## The problem
You ask a research tool a question. Half the best sources sit behind a **Bot Wall** — Cloudflare, Datadome, a Turnstile, a reCAPTCHA. Every other tool gets a `403`, silently drops those pages, and hands you a thinner report. **You never even learn what it missed.**
## What DeepCloak does
When a plain fetch hits a Bot Wall, DeepCloak **Escalates** that one URL to a **Stealth Fetch** and **Bypasses** the wall — recovering the content other agents abandon. Then it tells you, at the bottom of every report, exactly how many walls it broke through.
It's a thin, local-first orchestrator over two great projects: [`local-deep-research`](https://github.com/LearningCircuit/local-deep-research) (the research loop) and [`CloakBrowser`](https://github.com/CloakHQ/CloakBrowser) (the stealth browser). Use it as a **CLI**, an **MCP server**, or a **Claude skill**. MIT.
## 🌑 Why we built this
The open web is quietly closing. More of the best writing now sits behind a bot check, and AI research agents — the tools we increasingly trust to read the web *for us* — go blind at exactly those doors, without ever saying so. A report that silently skips every walled source isn't neutral; it's wrong in a way you can't see.
DeepCloak's stance is simple: **your agent should be able to read what a person with a browser can read** — and it should be honest about how it got there. It attempts a Stealth Fetch when needed and prints an Evidence Record of the result. The process runs locally; search engines receive queries, source websites receive requests, and a configured cloud LLM receives research context. Choose a local LLM when that context must stay on the machine. Capability *and* transparency, MIT-licensed, no lock-in.
## ✨ Why it's different
| | Plain deep research | **DeepCloak** |
| --- | :---: | :---: |
| Reads the open web | ✅ | ✅ |
| Reads Cloudflare / Datadome / Turnstile / reCAPTCHA pages | ❌ *dropped silently* | ✅ **Bypassed** |
| Tells you which sources were walled | ❌ | ✅ Evidence Record |
| Local-first (no API key required) | ✅ | ✅ |
| Fast on open pages | — | ✅ *plain-first, stealth only when needed* |
> **Verified live — not mocked.** The clip above is an **unedited screen recording** (captured with `ffmpeg`, no compositing) of a real `deepcloak` run against a **local LLM (Qwen) + SearXNG — no API key**. It Escalates on each Bot Wall and **Bypasses 8 Cloudflare/Turnstile walls** in one pass, then writes a cited report. Full clip: [`docs/media/demo-real.mp4`](docs/media/demo-real.mp4); a raw asciinema session is also kept at [`docs/media/demo.cast`](docs/media/demo.cast). Wall counts vary per run (8–20) because the open web does.
## 🚀 Quickstart
```bash
pip install deepcloak
deepcloak setup # one-time: downloads the stealth browser
export OPENAI_API_KEY=... # or ANTHROPIC_API_KEY / GEMINI_API_KEY — or --provider ollama
deepcloak "How does Cloudflare Turnstile detect bots?" --depth detailed --out report.md
```
You get a cited `report.md` ending with a `🛡️ Bypassed N bot-walled sources` section, plus a `report.md.evidence.json` sidecar.
## 🧠 How it works
```
search (DuckDuckGo, no setup) ─▶ candidate URLs
│
▼ for each page:
plain fetch ─▶ Bot Wall detected? ──no──▶ use it (fast)
│ yes
▼
Escalate ─▶ Stealth Fetch (CloakBrowser) ─▶ Bypass
│
▼
research loop (local-deep-research) ─▶ cited report + Evidence Records
```
Stealth is heavy, so DeepCloak tries a cheap plain fetch first and only launches the stealth browser when it actually detects a Bot Wall (`--stealth auto`, the default). Use `--depth detailed`/`report` to fetch full pages where Bypasses happen.
## 🤖 Connect it to your agent (MCP)
DeepCloak runs as a stdio **MCP server** exposing `deep_research(query, depth, provider, model)`, `quick_summary(query, provider, model)`, `get_evidence(run_id)`, and `list_models(provider)`.
**Claude Code** — add to your project's `.mcp.json` (an example ships in this repo):
```json
{ "mcpServers": { "deepcloak": { "command": "deepcloak", "args": ["mcp"] } } }
```
**Codex** — add to `~/.codex/config.toml`:
```toml
[mcp_servers.deepcloak]
command = "deepcloak"
args = ["mcp"]
```
Then your agent can call `deep_research` and read bot-walled sources directly. Prefer a slash-style skill? Drop [`skill/SKILL.md`](skill/SKILL.md) into `~/.claude/skills/deepcloak/`.
## ⚙️ Configuration
| Flag | Default | Notes |
| --- | --- | --- |
| `--depth` | `detailed` | `quick` / `detailed` / `report` |
| `--engine` | `duckduckgo` | `searxng` / `auto` |
| `--stealth` | `auto` | `always` / `off` |
| `--provider` / `--model` | auto-detected | `OPENAI_API_KEY` → `ANTHROPIC_API_KEY` → `GEMINI_API_KEY` → `OPENROUTER_API_KEY`; or `ollama` / `lmstudio` / `llamacpp` / `openai-endpoint` (`--base-url`) |
| `--respect-robots` | off | honor robots.txt |
| `--proxy` | — | SOCKS5 for the Stealth Fetch |
| `--json` | off | print the full result (report + Evidence Records + settings) as JSON |
**Picking a model:** list what a provider offers with `deepcloak models --provider openrouter` (works for every provider; local servers via `--base-url`), then pass `--model <id>`. With several credentials in your environment and a real terminal, DeepCloak simply asks. Settings persist in `~/.config/deepcloak/config.toml` (override with `$DEEPCLOAK_CONFIG`); precedence: CLI flags > environment > config file.
**Run history:** finished runs are saved to `~/.local/share/deepcloak/runs/` — browse with `deepcloak runs`, replay a report with `deepcloak show <id>` (`--json` for the full record).
## ⚠️ Responsible use
DeepCloak Bypasses bot-detection. **You are responsible for having the right to access whatever you fetch.** robots.txt is **ignored by default**; pass `--respect-robots` to honor it ([ADR-0002](docs/adr/0002-ignore-robots-by-default.md)). Don't use it to violate sites' terms or the law.
## 🗺️ Roadmap
- More Bot Wall signatures + smarter Escalation heuristics
- More search backends beyond DuckDuckGo / SearXNG
- Cache Bypassed pages across runs
- Richer Evidence Record export (HTML / JSON schema)
Ideas welcome — start a [Discussion](https://github.com/Mrbaeksang/deepcloak/discussions) or open a [feature request](https://github.com/Mrbaeksang/deepcloak/issues/new/choose).
## 🛠️ Built on
[`local-deep-research`](https://github.com/LearningCircuit/local-deep-research) (MIT) + [`CloakBrowser`](https://github.com/CloakHQ/CloakBrowser) (MIT), via pip — no vendored code. Domain glossary in [CONTEXT.md](CONTEXT.md); design decisions in [docs/adr/](docs/adr/); contributing guide in [CONTRIBUTING.md](CONTRIBUTING.md).
## 📄 License
MIT — see [LICENSE](LICENSE) and [NOTICE](NOTICE).
<div align="center">
**If DeepCloak read a page your last tool gave up on, drop a ⭐ — it helps others find it.**
Built and maintained by **[Sanghyeon Baek (백상현 · @Mrbaeksang)](https://baeksang.dev)** · [baeksang.dev](https://baeksang.dev) · [contact@baeksang.dev](mailto:contact@baeksang.dev)
[](https://star-history.com/#Mrbaeksang/deepcloak&Date)
</div>