The original author of one-api has not been active for a long time, resulting in a backlog of PRs that cannot be updated. Therefore, I forked the code and merged some PRs that I consider important. I also welcome everyone to submit PRs, and I will respond and handle them actively and quickly.
Fully compatible with the upstream version, can be used directly by replacing the container image, docker images:
ppcelery/one-api:latestppcelery/one-api:arm64-latest
Also welcome to register and use my deployed one-api gateway, which supports various mainstream models. For usage instructions, please refer to https://wiki.laisky.com/projects/gpt/pay/cn/#page_gpt_pay_cn.
- One API
- Turtorial
- New Features
- (Merged) Support gpt-vision
- Support update user's remained quota
- (Merged) Support aws claude
- Support openai images edits
- Support gemini-2.0-flash-exp
- Support replicate flux & remix
- Support replicate chat models
- Support OpenAI o1/o1-mini/o1-preview
- Get request's cost
- Support Vertex Imagen3
- Support gpt-4o-audio
- Support deepseek-reasoner & gemini-2.0-flash-thinking-exp-01-21
- Support o3-mini
- Support gemini-2.0-flash
- Support OpenRouter's reasoning content
- Support claude-3-7-sonnet & thinking
- Automatically Enable Thinking and Customize Reasoning Format via URL Parameters
- Support AWS cross-region inferences
- Support OpenAI web search models
- Support gemini multimodal output #2197
- Support coze oauth authentication
- Support gemini-2.5-pro
- Support o3 & o4-mini & gpt-4.1
- Support gpt-image-1's image generation & edits
- Bug fix
Run one-api using docker-compose:
oneapi:
image: ppcelery/one-api:latest
restart: unless-stopped
logging:
driver: "json-file"
options:
max-size: "10m"
environment:
# (optional) SESSION_SECRET set a fixed session secret so that user sessions won't be invalidated after server restart
SESSION_SECRET: xxxxxxx
# (optional) DEBUG enable debug mode
DEBUG: "true"
# (optional) DEBUG_SQL display SQL logs
DEBUG_SQL: "true"
# (optional) If you access one-api using a non-HTTPS address, you need to set DISABLE_COOKIE_SECURE to true
DISABLE_COOKIE_SECURE: "true"
# (optional) ENFORCE_INCLUDE_USAGE require upstream API responses to include usage field
ENFORCE_INCLUDE_USAGE: "true"
# (optional) MAX_ITEMS_PER_PAGE maximum items per page, default is 10
MAX_ITEMS_PER_PAGE: 10
# (optional) GLOBAL_API_RATE_LIMIT maximum API requests per IP within three minutes, default is 1000
GLOBAL_API_RATE_LIMIT: 1000
# (optional) GLOBAL_WEB_RATE_LIMIT maximum web page requests per IP within three minutes, default is 1000
GLOBAL_WEB_RATE_LIMIT: 1000
# /v1 API ratelimit for each token
GLOBAL_RELAY_RATE_LIMIT: 1000
# Whether to ratelimit for channel, 0 is unlimited, 1 is to enable the ratelimit
GLOBAL_CHANNEL_RATE_LIMIT: 1
# (optional) REDIS_CONN_STRING set REDIS cache connection
REDIS_CONN_STRING: redis://100.122.41.16:6379/1
# (optional) FRONTEND_BASE_URL redirect page requests to specified address, server-side setting only
FRONTEND_BASE_URL: https://oneap
8000
i.laisky.com
# (optional) OPENROUTER_PROVIDER_SORT set sorting method for OpenRouter Providers, default is throughput
OPENROUTER_PROVIDER_SORT: throughput
# (optional) CHANNEL_SUSPEND_SECONDS_FOR_429 set the duration for channel suspension when receiving 429 error, default is 60 seconds
CHANNEL_SUSPEND_SECONDS_FOR_429: 60
volumes:
- /var/lib/oneapi:/data
ports:
- 3000:3000The initial default account and password are root / 123456.
You can update the used quota using the API key of any token, allowing other consumption to be aggregated into the one-api for centralized management.
Each chat completion request will include a X-Oneapi-Request-Id in the returned headers. You can use this request id to request GET /api/cost/request/:request_id to get the cost of this request.
The returned structure is:
type UserRequestCost struct {
Id int `json:"id"`
CreatedTime int64 `json:"created_time" gorm:"bigint"`
UserID int `json:"user_id"`
RequestID string `json:"request_id"`
Quota int64 `json:"quota"`
CostUSD float64 `json:"cost_usd" gorm:"-"`
}
By default, the thinking mode is automatically enabled for the deepseek-r1 model, and the response is returned in the open-router format.
By default, the thinking mode is not enabled. You need to manually pass the thinking field in the request body to enable it.
Supports two URL parameters: thinking and reasoning_format.
thinking: Whether to enable thinking mode, disabled by default.reasoning_format: Specifies the format of the returned reasoning.reasoning_content: DeepSeek official API format, returned in thereasoning_contentfield.reasoning: OpenRouter format, returned in thereasoningfield.thinking: Claude format, returned in thethinkingfield.
support gpt-4o-search-preview & gpt-4o-mini-search-preview
- BUGFIX: Several issues when updating tokens #1933
- feat(audio): count whisper-1 quota by audio duration #2022
- fix: Fix issue where high-quota users using low-quota tokens aren't pre-charged, causing large token deficits under high concurrency #25
- fix: channel test false negative #2065
- fix: resolve "bufio.Scanner: token too long" error by increasing buffer size #2128
- feat: Enhance VolcEngine channel support with bot model #2131
- fix: models API returns models in deactivated channels #2150
- fix: Automatically close channel when connection fails
- fix: update EmailDomainWhitelist submission logic #33
- fix: send ByAll
- fix: oidc token endpoint request body #2106 #36