English | 简体中文
ninja
Reverse engineered ChatGPT
proxy (bypass Cloudflare 403 Access Denied)
Features
- API key acquisition
- Email/password account authentication (Google/Microsoft third-party login not supported)
- Supports obtaining RefreshToken
ChatGPT-API
/OpenAI-API
/ChatGPT-to-API
Http API proxy (for third-party client access)- Support IP proxy pool (support using Ipv6 subnet as proxy pool)
- ChatGPT WebUI
- Very small memory footprint
Limitations: This cannot bypass OpenAI's outright IP ban
ArkoseLabs
Sending GPT-4/GPT-3.5/Creating API-Key
dialog requires sending Arkose Token
as a parameter. There are only two supported solutions for the time being.
- Use HAR
- Supports HAR feature pooling, can upload multiple HARs at the same time, and use rotation training strategy
The
ChatGPT
official website sends aGPT-4
session message, and the browserF12
downloads thehttps://tcr9i.chat.openai.com/fc/gt2/public_key/35536E1E-65B4-4D96-9D97-6ADB7EFF8147
interface. HAR log file, use the startup parameter--arkose-gpt4-har-dir
to specify the HAR directory path to use (if you do not specify a path, use the default path~/.gpt4
, you can directly upload and update HAR ), the same method applies toGPT-3.5
and other types. Supports WebUI to upload and update HAR, request path:/har/upload
, optional upload authentication parameter:--arkose-har-upload-key
- Use YesCaptcha / CapSolver
The platform performs verification code parsing, start the parameter
--arkose-solver
to select the platform (useYesCaptcha
by default),--arkose-solver-key
fill inClient Key
- Both solutions are used, the priority is:
HAR
>YesCaptcha
/CapSolver
YesCaptcha
/CapSolver
is recommended to be used with HAR. When the verification code is generated, the parser is called for processing. After verification, HAR is more durable.
Currently OpenAI has updated
Login
which requires verification ofArkose Token
. The solution is the same asGPT-4
. Fill in the startup parameters and specify the HAR file--arkose-auth-har-dir
. If you don't want to upload, you can log in through the browser code, which is not required. To create an API-Key, you need to upload the HAR feature file related to the Platform. The acquisition method is the same as above.
Recently,
OpenAI
has canceled theArkose
verification forGPT-3.5
. It can be used without uploading HAR feature files (uploaded ones will not be affected). After compatibility,Arkose
verification may be turned on again, and startup parameters need to be added.--arkose-gpt3-experiment
enables theGPT-3.5
modelArkose
verification processing, and the WebUI is not affected.
Http Server
*
represents any URL
suffix
Public interface, -
ChatGPT-API
/public-api/*
/backend-api/*
-
OpenAI-API
/v1/*
-
Platform-API
/dashboard/*
-
ChatGPT-To-API
/to/v1/chat/completions
About using
ChatGPT
toAPI
, useAceessToken
directly asAPI Key
, interface path:/to/v1/chat/completions
-
Authorization
- Login:
/auth/token
, formoption
optional parameter, default isweb
login, returnsAccessToken
andSession
; parameter isapple
/platform
, returnsAccessToken
andRefreshToken
- Refresh
RefreshToken
:/auth/refresh_token
- Revoke
RefreshToken
:/auth/revoke_token
- Refresh
Session
:/api/auth/session
, send a cookie named__Secure-next-auth.session-token
to call refreshSession
, and return a newAccessToken
About the method of obtaining
RefreshToken
, use theChatGPT App
login method of theApple
platform. The principle is to use the built-in MITM agent. When theApple device
is connected to the agent, you can log in to theApple platform
to obtainRefreshToken
. It is only suitable for small quantities or personal use(large quantities will seal the device, use with caution)
. For detailed usage, please see the startup parameter description. - Login:
API documentation
Basic services
- ChatGPT WebUI
- Expose
ChatGPT-API
/OpenAI-API
proxies API
prefix is consistent with the official oneChatGPT
toAPI
- Can access third-party clients
- Can access IP proxy pool to improve concurrency
- Supports obtaining RefreshToken
- Support file feature pooling in HAR format
Parameter Description
--level
, environment variableLOG
, log level: default info--bind
, environment variableBIND
, service listening address: default 0.0.0.0:7999,--tls-cert
, environment variableTLS_CERT
', TLS certificate public key. Supported format: EC/PKCS8/RSA--tls-key
, environment variableTLS_KEY
, TLS certificate private key--proxies
, Proxy, supports proxy pool, multiple proxies are separated by,
, format: protocol://user:pass@ip:port, if the local IP is banned, you need to turn off the use of direct IP when using the proxy pool,--disable-direct
turns off direct connection, otherwise your banned local IP will be used according to load balancing--workers
, worker threads: default 1--disable-webui
, if you don’t want to use the default built-in WebUI, use this parameter to turn it off
# Generate certificate
ninja genca
ninja run --pbind 0.0.0.0:8888
# Set the network on your mobile phone to set your proxy listening address, for example: http://192.168.1.1:8888
# Then open the browser http://192.168.1.1:8888/preauth/cert, download the certificate, install it and trust it, then open iOS ChatGPT and you can play happily
For
Web login
, a cookie named:__Secure-next-auth.session-token
is returned by default. The client only needs to save this cookie. Calling/api/auth/session
can also refreshAccessToken
Install
-
Ubuntu(Other Linux)
Making Releases has a precompiled deb package, binaries, in Ubuntu, for example:
wget https://github.com/gngpp/ninja/releases/download/v0.7.8/ninja-0.7.8-x86_64-unknown-linux-musl.tar.gz
tar -xf ninja-0.7.8-x86_64-unknown-linux-musl.tar.gz
./ninja run
-
OpenWrt
There are pre-compiled ipk files in GitHub Releases, which currently provide versions of aarch64/x86_64 and other architectures. After downloading, use opkg to install, and use nanopi r4s as example:
wget https://github.com/gngpp/ninja/releases/download/v0.7.8/ninja_0.7.8_aarch64_generic.ipk
wget https://github.com/gngpp/ninja/releases/download/v0.7.8/luci-app-ninja_1.1.6-1_all.ipk
wget https://github.com/gngpp/ninja/releases/download/v0.7.8/luci-i18n-ninja-zh-cn_1.1.6-1_all.ipk
opkg install ninja_0.7.8_aarch64_generic.ipk
opkg install luci-app-ninja_1.1.6-1_all.ipk
opkg install luci-i18n-ninja-zh-cn_1.1.6-1_all.ipk
-
Docker
docker run --rm -it -p 7999:7999 --name=ninja \
-e WORKERS=1 \
-e LOG=info \
gngpp/ninja:latest run
- Docker Compose
CloudFlare Warp
is not supported in your region (China), please delete it, or if yourVPS
IP can be directly connected toOpenAI
, you can also delete it
version: '3'
services:
ninja:
image: ghcr.io/gngpp/ninja:latest
container_name: ninja
restart: unless-stopped
environment:
- TZ=Asia/Shanghai
- PROXIES=socks5://warp:10000
command: run --disable-direct
ports:
- "8080:7999"
depends_on:
- warp
warp:
container_name: warp
image: ghcr.io/gngpp/warp:latest
restart: unless-stopped
watchtower:
container_name: watchtower
image: containrrr/watchtower
volumes:
- /var/run/docker.sock:/var/run/docker.sock
command: --interval 3600 --cleanup
restart: unless-stopped
Command Manual
$ ninja --help
Reverse engineered ChatGPT proxy
Usage: ninja [COMMAND]
Commands:
run Run the HTTP server
stop Stop the HTTP server daemon
start Start the HTTP server daemon
restart Restart the HTTP server daemon
status Status of the Http server daemon process
log Show the Http server daemon log
genca Generate MITM CA certificate
gt Generate config template file (toml format file)
update Update the application
help Print this message or the help of the given subcommand(s)
Options:
-h, --help Print help
-V, --version Print version
$ ninja run --help
Run the HTTP server
Usage: ninja run [OPTIONS]
Options:
-L, --level <LEVEL>
Log level (info/debug/warn/trace/error) [env: LOG=] [default: info]
-C, --config <CONFIG>
Configuration file path (toml format file) [env: CONFIG=]
-b, --bind <BIND>
Server bind address [env: BIND=] [default: 0.0.0.0:7999]
-W, --workers <WORKERS>
Server worker-pool size (Recommended number of CPU cores) [default: 1]
--concurrent-limit <CONCURRENT_LIMIT>
Enforces a limit on the concurrent number of requests the underlying [default: 1024]
-x, --proxies <PROXIES>
Server proxies pool, Only support http/https/socks5 protocol [env: PROXIES=]
-i, --interface <INTERFACE>
Bind address for outgoing connections (or IPv6 subnet fallback to Ipv4) [env: INTERFACE=]
-I, --ipv6-subnet <IPV6_SUBNET>
IPv6 subnet, Example: 2001:19f0:6001:48e4::/64 [env: IPV6_SUBNET=]
--disable-direct
Disable direct connection [env: DISABLE_DIRECT=]
--cookie-store
Enabled Cookie Store [env: COOKIE_STORE=]
--timeout <TIMEOUT>
Client timeout (seconds) [default: 360]
--connect-timeout <CONNECT_TIMEOUT>
Client connect timeout (seconds) [default: 20]
--tcp-keepalive <TCP_KEEPALIVE>
TCP keepalive (seconds) [default: 60]
--pool-idle-timeout <POOL_IDLE_TIMEOUT>
Set an optional timeout for idle sockets being kept-alive [default: 90]
--tls-cert <TLS_CERT>
TLS certificate file path [env: TLS_CERT=]
--tls-key <TLS_KEY>
TLS private key file path (EC/PKCS8/RSA) [env: TLS_KEY=]
-A, --auth-key <AUTH_KEY>
Login Authentication Key [env: AUTH_KEY=]
-D, --disable-webui
Disable WebUI [env: DISABLE_WEBUI=]
--cf-site-key <CF_SITE_KEY>
Cloudflare turnstile captcha site key [env: CF_SECRET_KEY=]
--cf-secret-key <CF_SECRET_KEY>
Cloudflare turnstile captcha secret key [env: CF_SITE_KEY=]
--arkose-endpoint <ARKOSE_ENDPOINT>
Arkose endpoint, Example: https://client-api.arkoselabs.com
-E, --arkose-gpt3-experiment
Enable Arkose GPT-3.5 experiment
--arkose-gpt3-har-dir <ARKOSE_GPT3_HAR_DIR>
About the browser HAR directory path requested by ChatGPT GPT-3.5 ArkoseLabs
--arkose-gpt4-har-dir <ARKOSE_GPT4_HAR_DIR>
About the browser HAR directory path requested by ChatGPT GPT-4 ArkoseLabs
--arkose-auth-har-dir <ARKOSE_AUTH_HAR_DIR>
About the browser HAR directory path requested by Auth ArkoseLabs
--arkose-platform-har-dir <ARKOSE_PLATFORM_HAR_DIR>
About the browser HAR directory path requested by Platform ArkoseLabs
-K, --arkose-har-upload-key <ARKOSE_HAR_UPLOAD_KEY>
HAR file upload authenticate key
-s, --arkose-solver <ARKOSE_SOLVER>
About ArkoseLabs solver platform [default: yescaptcha]
-k, --arkose-solver-key <ARKOSE_SOLVER_KEY>
About the solver client key by ArkoseLabs
-T, --tb-enable
Enable token bucket flow limitation
--tb-store-strategy <TB_STORE_STRATEGY>
Token bucket store strategy (mem/redis) [default: mem]
--tb-redis-url <TB_REDIS_URL>
Token bucket redis connection url [default: redis://127.0.0.1:6379]
--tb-capacity <TB_CAPACITY>
Token bucket capacity [default: 60]
--tb-fill-rate <TB_FILL_RATE>
Token bucket fill rate [default: 1]
--tb-expired <TB_EXPIRED>
Token bucket expired (seconds) [default: 86400]
-B, --pbind <PBIND>
Preauth MITM server bind address [env: PREAUTH_BIND=]
-X, --pupstream <PUPSTREAM>
Preauth MITM server upstream proxy, Only support http/https/socks5 protocol [env: PREAUTH_UPSTREAM=]
--pcert <PCERT>
Preauth MITM server CA certificate file path [default: ca/cert.crt]
--pkey <PKEY>
Preauth MITM server CA private key file path [default: ca/key.pem]
-h, --help
Print help
Platform Support
- Linux
x86_64-unknown-linux-musl
aarch64-unknown-linux-musl
armv7-unknown-linux-musleabi
armv7-unknown-linux-musleabihf
arm-unknown-linux-musleabi
arm-unknown-linux-musleabihf
armv5te-unknown-linux-musleabi
- Windows
x86_64-pc-windows-msvc
- MacOS
x86_64-apple-darwin
aarch64-apple-darwin
Compile
- Linux compile, Ubuntu machine for example:
git clone https://github.com/gngpp/ninja.git && cd ninja
cargo build --release
- OpenWrt Compile
cd package
svn co https://github.com/gngpp/ninja/trunk/openwrt
cd -
make menuconfig # choose LUCI->Applications->luci-app-ninja
make V=s
Instructions
- Open source projects can be modified, but please keep the original author information to avoid losing technical support.
- Project is standing on the shoulders of other giants, thanks!
- Submit an issue if there are errors, bugs, etc., and I will fix them.