简体中文 | English

Gpt-Agreement-Payment

End-to-end replay tool for ChatGPT Plus / Team subscription agreements: reverse-engineered from packet capture the entire chain Stripe Checkout → PayPal / GoPay / QRIS → ChatGPT manual-approval → Codex OAuth + PKCE and implemented as a runnable client. Includes a from-scratch hCaptcha visual solver, PayPal fraud detection early-exit branch, and a set of empirical anti-fraud mechanism data collected from real execution.

Author-run API Relay Station

_{👆 Click the image to visit lukyface · Invite code 9ipF · Exchange group 1104565621}

[!CAUTION] Using this project constitutes acceptance of all terms in NOTICE. The project is provided AS IS, with no warranties, and the maintainers assume no responsibility. Authorized use only on systems you own / legitimate CTF / authorized bug bounty in-scope assets / security research. Strictly prohibited for fraud, payment evasion, bulk account creation for resale, violation of third-party ToS, unauthorized targets. All legal responsibility rests with the user. If you do not accept these terms, do not use.

What This Is

Supports four subscription activation paths:

Given a clean proxy + payment credential, run the command to completion and get the OAuth refresh_token.

Four points worth examining:

• N-worker concurrency + phone-lock OTP critical section mutual exclusion (webui/backend/parallel_runner.py). Same phone across multiple workers serializes OTP phase using advisory lock, pre/post-OTP fully parallelized; DB atomic claim + placeholder INSERT prevents multiple workers contending for same promo_link / same inventory email. Frontend configures phone pool (M rows) + concurrency N (can exceed M), worker maps to phone pool via i % M round-robin.
• hCaptcha visual solver (CTF-pay/hcaptcha_auto_solver.py, ~4000 lines, independently usable). VLM primary path + CLIP/OpenCV heuristic fallback + Playwright human action synthesis, covers 12 known hCaptcha question types.
• Anti-fraud mechanism empirical data. IP string-level exact fingerprinting, batch correlation delayed banning, probe layer vs ban layer separation. Measured sample of ~2% 24-hour survival rate across 45 accounts, including corrected models. See docs/anti-fraud-research.md.
• Twelve-path self-healing daemon loops (pipeline.py::daemon()). Webshare API auto IP rotation (webshare jitter fallback to /tmp/gost_last.json cache), CF DNS quota cleanup, tmpfs orphan recovery, gost relay watchdog, DataDome slider auto-drag. Design goal is unattended operation for weeks.

Architecture

flowchart LR
    A[pipeline.py / webui] --> B[CTF-reg/<br/>browser_register.py<br/>Camoufox + Turnstile]
    B --> C[CTF-pay/card.py<br/>Stripe Checkout replay]
    C --> D{Payment path}
    D -->|paypal| E1[Camoufox / Node RPA<br/>PayPal billing agreement]
    D -->|gopay| E2[Midtrans linking<br/>+ WhatsApp OTP]
    D -->|qris| E3[QRIS QR<br/>+ remote preview]
    E1 --> F[Stripe poll<br/>state=succeeded]
    E2 --> F
    E3 --> F
    F --> G[Camoufox second login<br/>Codex OAuth + PKCE]
    G --> H[refresh_token<br/>output/webui.db &#40;SQLite&#41;]

For detailed subsystem decomposition, file organization, and protocol chain details, see docs/architecture.md.

Current Status and Entry Barriers

To be frank, this is not an out-of-the-box tool. Running through the entire pipeline requires at least:

• A real, login-capable PayPal account (or go through PayPal guest checkout authorization flow)
• A proxy with exit in EU / US / ID (region locked based on selected payment path)
• A Cloudflare zone (optional, for enabling catch-all subdomain email registration; also supports Outlook code pool)
• A Linux machine capable of running Camoufox + Playwright (approximately 5 GB disk + 2 GB memory)
• (Required for PayPal guest checkout) An SMS code gateway API key for PayPal signup
• (Required for GoPay) A WhatsApp online number + WhatsApp code verification service
• (Optional) An OpenAI-compatible VLM API key for hCaptcha solving; residential / pseudo-residential exits typically don't trigger hCaptcha, and CLIP degradation is available without VLM
• (Optional) A captcha solving platform API key compatible with createTask/getTaskResult protocol as a fallback for browser passive captcha

Full end-to-end setup typically takes 1–3 hours for initial configuration tuning. Once daemon mode runs stably, a single pipeline takes ~5 minutes; concurrent mode with 2 workers on the same phone generates 2 accounts in ~3 minutes.

Code is research-oriented, organized by protocol stage, and doesn't prioritize maximum readability.

Getting Started

Beginner Path: WebUI Configuration Wizard (Recommended)

Compress 1–3 hours of manual configuration into ~15 minutes. 14-step wizard + real-time preflight self-checks + built-in runtime controller (SSE log stream + concurrency panel) generate CTF-pay/config.auto.json + CTF-reg/config.paypal-proxy.json configurations.

Docker Deployment (Fastest Path, One-Click Start)

The repo comes with Dockerfile (multi-stage build: Node frontend + Ubuntu 24.04 runtime) + docker-compose.yml, packing all system dependencies / Playwright Chromium+Firefox / Camoufox / gost SOCKS5 relay / Node QuickJS (for OpenAI Sentinel) into the image. The git working tree on the host serves as the single source of truth, bind-mounted into the container; restart with docker compose restart for instant effect after modifying Python code, no rebuild needed.

git clone https://github.com/DanOps-1/Gpt-Agreement-Payment
cd Gpt-Agreement-Payment
docker compose up -d --build
# Default listening on 127.0.0.1:8765 (host port), open http://127.0.0.1:8765/ in browser
# First visit redirects to /setup to create admin

Common Maintenance Commands:

# Real-time logs
docker compose logs -f webui

# Debug inside container (pip list / run tests / inspect SQLite)
docker compose exec webui bash

# Reload uvicorn after Python code changes (bind mount already syncs source code, just restart process)
docker compose restart webui

# Rebuild dist in container after frontend code changes
docker compose exec webui sh -c "cd /app/webui/frontend && npm run build"

# Completely stop + clean containers
docker compose down

# Image upgrade (pull new base / upgrade Python packages)
docker compose build --no-cache && docker compose up -d

Data Persistence

The output/ directory is a host directory bind mount containing webui.db (SQLite) + run results / logs visible and backupable to the host. webui/frontend/dist and node_modules use anonymous volumes, which are overridden by the baked image version (not overridden by an empty host directory).

Public Access (nginx Reverse Proxy + HTTPS)

See webui/README.md. By default, docker-compose.yml binds the port to 127.0.0.1:8765; to expose directly to 0.0.0.0, modify the ports section (not recommended without an authentication layer in front).

Manual Installation (without Docker)

# 1. Backend dependencies
pip install -r webui/requirements.txt

# 2. Frontend build (one-time)
cd webui/frontend && pnpm i && pnpm build && cd ../..

# 3. Start
python -m webui.server
# Open http://127.0.0.1:8765 in browser, first visit redirects to /setup to create admin

Support single run (PayPal billing agreement / GoPay / QRIS / promo long links) + concurrent mode (frontend configured with phone pool + concurrency N), public network access via nginx reverse proxy see webui/README.md.

Installation

git clone https://github.com/DanOps-1/Gpt-Agreement-Payment
cd Gpt-Agreement-Payment
pip install requests curl_cffi playwright camoufox browserforge mitmproxy pybase64
playwright install firefox
camoufox fetch

hCaptcha solver ML dependencies (torch / transformers / opencv) are recommended to be installed separately in a venv, approximately 4 GB:

python -m venv ~/.venvs/ctfml
~/.venvs/ctfml/bin/pip install torch transformers opencv-python pillow numpy

Complete dependency list and system packages, see docs/installation.md.

Configuration

Copy the template and fill in values:

cp CTF-pay/config.paypal.example.json     CTF-pay/config.paypal.json
cp CTF-reg/config.paypal-proxy.example.json   CTF-reg/config.paypal-proxy.json

For field meanings and schema, see docs/configuration.md. The Docker deployment entrypoint will automatically bootstrap these two config files on first startup. After editing on the host, run docker compose restart webui to take effect.

Running

Docker users only need to use webui to run. The CLI commands below are for users running natively locally; to run the same commands in Docker, first docker compose exec webui bash to enter the container.

# 1) Single complete workflow (PayPal billing agreement)
xvfb-run -a python pipeline.py --config CTF-pay/config.paypal.json --paypal

# 2) Plus protocol authorization + promo long link (PayPal guest checkout)
xvfb-run -a python scripts/no_card_paypal_plus.py \
    --config CTF-pay/config.paypal.json --paypal-node-rpa \
    --phone <your_phone> --otp-timeout 240

# 3) N worker concurrency (same phone is fine, OTP phase auto-queues via advisory lock)
# Recommended to start via webui: select concurrency mode, frontend configures phone pool + concurrency N

# 4) Continuous account pool maintenance
xvfb-run -a python pipeline.py --config CTF-pay/config.paypal.json --paypal --daemon

Four Operating Modes (Single / Batch / Self-Dealer / Daemon) + Concurrency Mode Differences and Parameters

See docs/operating-modes.md.

Documentation

Known Limitations

• PayPal billing agreement only available in EU. Stripe account restrictions; can only place orders as EU identities like IE.
• PayPal guest checkout path subject to multiple risk controls. INSTRUMENT_SHARING_LIMIT_EXCEEDED / CC_LINKED_TO_FULL_ACCOUNT / CREATE_CARD_ACCOUNT_CANDIDATE_VALIDATION_ERROR / DataDome captcha etc. will all trigger; the script has early exit + automatic retry branches for each known error.
• Batch registration next-day survival rate ~2%. This is caused by batch correlation effects in ChatGPT's anti-fraud mechanism, not the tool itself. See docs/anti-fraud-research.md.
• Free account path currently not working. OpenAI changed the free account secondary login flow; redirect to /add-phone cannot be bypassed without a real phone number; ChatGPT-Web client's access_token calling Codex API has audience mismatch.
• Stripe runtime fingerprint drifts. runtime.version / js_checksum / rv_timestamp need realignment approximately every few weeks.
• hCaptcha question type coverage incomplete. Current 12 common types; when uncovered, VLM directly outputs coordinates as fallback, success rate not guaranteed.

Contribution

Most valuable contributions ranked by impact:

1. New hCaptcha question type solvers
2. Protocol adaptation when Stripe / PayPal / OpenAI introduces breaking changes
3. New failure mode daemon self-healing branches observed in practice (with logs)
4. Anti-fraud empirical data supplements (desensitization format per existing examples)
5. Multi-IP pool / proxy rotation implementation to free concurrency from single IP limitation
6. Documentation improvement / translation

⚠️ Maintainers cannot manually reproduce PRs. When submitting PRs, please follow the PR template providing detailed description + running evidence (requirements vary by change type: solver question types need round JSON, protocol adaptation needs packet capture comparison, daemon self-healing needs trigger logs and recovery logs). PRs lacking evidence will be closed without explanation.

See CONTRIBUTING.md for complete workflow, code style, and desensitization checklist for research contributions. See CODE_OF_CONDUCT.md for community guidelines. Do not open issues for security problems; see SECURITY.md.

Acknowledgments

Key dependencies in the tool chain:

• Camoufox — antidetect Firefox build, foundation of the entire browser automation layer
• mitmproxy — protocol packet capture
• Playwright — browser automation
• curl_cffi — TLS fingerprint simulation
• OpenAI CLIP — visual backbone for heuristic solver
• gost — SOCKS5 relay

Code Contributors

Thanks to the following friends for code contributions (by PR time):

• @Lium-7768 — #12 Align GoPay step visibility to hide 06 and 13
• @DragonBaiMo — #15 Algorithmicized persona generator + email name co-sourcing
• @laochendeai — #21 detect blocking challenge pages

Community

Special thanks to the LINUX DO community — the earliest feedback sources, testers, and protocol change reporters all come from here.

Sponsorship

If this project has been helpful to you, you're welcome to buy the author a coffee ☕

Donation Thanks

Thanks to the following friends for their support (no particular order):

The gesture matters more than the amount; every bit of support is motivation to continue maintaining the project 🙏

Star History

Disclaimer

[!IMPORTANT] Using this project constitutes your acceptance of having fully read, completely understood, and explicitly agreed to all terms in NOTICE. If you cannot accept — do not use this project, delete all copies.

License is MIT, but the License itself is not the complete disclaimer. Full disclaimer terms are in NOTICE; below is the key summary:

This project is provided "AS IS" without warranties of any kind. Including but not limited to merchantability, fitness for a particular purpose, non-infringement, safety, stability, and continued compatibility with third-party services. You assume all risks of using this project.

Use only within authorized scope. Permitted: systems you own, legal CTF, authorized bug bounty projects on in-scope assets, security research. Prohibited: fraud, payment evasion, bulk account creation for resale, violating third-party ToS, unauthorized targets.

You bear all legal responsibility. Including but not limited to account suspension, payment loss, criminal liability, civil damages, administrative penalties, third-party claims, reputation loss, business loss. Applicable laws may include US CFAA, EU GDPR, UK CMA, China Criminal Law Articles 285/286/287, etc. See section 4 of NOTICE.

Maintainers have no obligation to reply to issues, review PRs, fix bugs, maintain availability, or adapt to protocol changes. Reserved the right to archive, delete, rename, or stop maintaining this project at any time without prior notice.

Uncertain whether your use is legal — do not run. Ask a lawyer first, or talk to the target platform's security team.