dsk++ — 忘れられたDeepSeekライブラリを完全に非同期に書き直し（いくつかの機能を追加）

A while back, xtekky wrote a small library called deepseek4free for interacting with DeepSeek's chat infrastructure. It worked, but it was synchronous — meaning it blocked the event loop on every API call, making it painful to use in any async Python project. I needed it for a Discord bot I was building, and the blocking behavior was causing real problems. So I rewrote it. The result is dskpp (dsk++) — a drop-in async replacement that adds a few things the original never had. What's new in dsk++ Fully async — built on curl_cffi async sessions, async for streaming, asyncio.to_thread() for WASM PoW Concurrent file uploads — asyncio.gather() uploads multiple files simultaneously Session deletion — await api.delete_chat_session(session_id) Conversation history — await api.get_history(session_id) Proper error hierarchy — AuthenticationError , RateLimitError , NetworkError , CloudflareError , UploadFilesUnavailable , APIError Automatic Cloudflare detection and cookie refresh Docker support via DOCKERMODE=true The architecture The library has three layers: API layer ( api.py ) handles session management, streaming SSE parsing, file uploads, and automatic retry on Cloudflare blocks. Bypass layer ( server.py ) is a FastAPI + Chromium automation server that solves Cloudflare challenges and extracts cookies. PoW layer ( pow.py ) runs the WASM-based proof-of-work solver via asyncio.to_thread() so it doesn't block your event loop during CPU-bound hashing. Quick example import asyncio from dskpp.api import DeepSeekAPI async def main (): api = DeepSeekAPI ( " your_token_here " ) session = await api . create_chat_session () # Upload files concurrently file_ids = await api . upload_files ([ " report.pdf " , " data.xlsx " ]) # Stream response async for chunk in api . chat_completion ( session , " Summarize these files " , ref_file_ids = file_ids , search_enabled = False ): print ( chunk . get ( " content " , "" ), end = "" ) await api . delete_chat_session ( session ) await api . close () asyncio . run ( main ()) Before using the API, you need to generate cookies once: python run_and_get_cookies.py This launches Chromium, solves the Cloudflare challenge, and stores cookies locally. After that, the client handles refresh automatically. Getting your token Log into chat.deepseek.com , open DevTools, go to Application → Local Storage → chat.deepseek.com → USER_TOKEN . A word of caution This is built on reverse-engineered infrastructure. DeepSeek can change their API at any time and break things. It may also violate their ToS — use at your own risk and don't do anything that would get your account banned. Why bother? The original library was solid but unmaintained and sync-only. For anyone building bots, automation tools, or anything async in Python, a blocking DeepSeek client is a dead end. This fixes that. Also not everyone can afford the DeepSeek API, so this helps people automate the web interface endpoints for chat. Credit to xtekky and Doremii109 whose work this builds on. GitHub → Fundiman/dskpp Stars appreciated 🥺 A while back, xtekky wrote a small library called deepseek4free for interacting with DeepSeek's chat infrastructure. It worked, but it was synchronous — meaning it blocked the event loop on every API call, making it painful to use in any async Python project. I needed it for a Discord bot I was building, and the blocking behavior was causing real problems. So I rewrote it. The result is dskpp (dsk++) — a drop-in async replacement that adds a few things the original never had. What's new in dsk++ Fully async — built on curl_cffi async sessions, async for streaming, asyncio.to_thread() for WASM PoW Concurrent file uploads — asyncio.gather() uploads multiple files simultaneously Session deletion — await api.delete_chat_session(session_id) Conversation history — await api.get_history(session_id) Proper error hierarchy — AuthenticationError , RateLimitError , NetworkError , CloudflareError , UploadFilesUnavailable , APIError Automatic Cloudflare detection and cookie refresh Docker support via DOCKERMODE=true The architecture The library has three layers: API layer ( api.py ) handles session management, streaming SSE parsing, file uploads, and automatic retry on Cloudflare blocks. Bypass layer ( server.py ) is a FastAPI + Chromium automation server that solves Cloudflare challenges and extracts cookies. PoW layer ( pow.py ) runs the WASM-based proof-of-work solver via asyncio.to_thread() so it doesn't block your event loop during CPU-bound hashing. Quick example import asyncio from dskpp.api import DeepSeekAPI async def main (): api = DeepSeekAPI ( " your_token_here " ) session = await api . create_chat_session () # Upload files concurrently file_ids = await api . upload_files ([ " report.pdf " , " data.xlsx " ]) # Stream response async for chunk in api . chat_completion ( session , " Summarize these files " , ref_file_ids = file_ids , search_enabled = False ): print ( chunk . get ( " content " , "" ), end = "" ) await api . delete_chat_session ( session ) await api . close () asyncio . run ( main ()) Before using the API, you need to generate cookies once: python run_and_get_cookies.py This launches Chromium, solves the Cloudflare challenge, and stores cookies locally. After that, the client handles refresh automatically. Getting your token Log into chat.deepseek.com , open DevTools, go to Application → Local Storage → chat.deepseek.com → USER_TOKEN . A word of caution This is built on reverse-engineered infrastructure. DeepSeek can change their API at any time and break things. It may also violate their ToS — use at your own risk and don't do anything that would get your account banned. Why bother? The original library was solid but unmaintained and sync-only. For anyone building bots, automation tools, or anything async in Python, a blocking DeepSeek client is a dead end. This fixes that. Also not everyone can afford the DeepSeek API, so this helps people automate the web interface endpoints for chat. Credit to xtekky and Doremii109 whose work this builds on. GitHub → Fundiman/dskpp Stars appreciated 🥺