Skip to content
BranchesTags

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

51 Commits
.github/workflows
.github/workflows
 
 
deploy
deploy
 
 
docs
docs
 
 
scripts
scripts
 
 
tests/smoke
tests/smoke
 
 
.env.example
.env.example
 
 
.gitignore
.gitignore
 
 
LICENSE
LICENSE
 
 
Makefile
Makefile
 
 
PRD.md
PRD.md
 
 
README.md
README.md
 
 
README_CN.md
README_CN.md
 
 

Repository files navigation

Team AI Platform

English | 中文

AI Gateway Chat — A unified multi-model aggregation chat platform for small and medium-sized teams

Whether for enterprise teams or individual users, the private management of LLM conversation data is becoming increasingly important. Driven by recent real-world requirements for private Chatlog Data management, this project delivers a validated AI-Agent-Chatlog aggregation platform for small teams and companies, now fully functional and stably deployed.

While it is essentially an AI Gateway Chat project, we prefer to call it Team AI Platform.

Key Features

  1. Multi-Model Chat Platform — Designed for small/medium teams and companies, supporting all major LLM providers
  2. Unified Auth & i18n — Email/phone cross-verification registration/login with multi-language internationalization
  3. Prompt Engineering & Agent Orchestration — Prompt engineering, AI Native ChatAgent orchestration, and native model hyperparameter tuning
  4. Private Chatlog Management — Multi-Agent Workflow orchestration with cross-reuse and Agent marketplace sharing
  5. Centralized API Key Control — Aggregated official models with API keys managed by the Admin backend; team members focus on conversations and agent outputs
  6. Extensible Channel Support — Supports domestic and international model providers, plus third-party relay interfaces for downstream expansion
  7. 7 Channels, 165+ Model Matrix — Integrated with Zhipu, DeepSeek, Alibaba Cloud Bailian, Kimi, Volcano Doubao, Xiaomi MiMo — 7 API channels with 165+ models on demand
  8. Conversation Operations — Chat History hot-editing, Chat Fork, Chat Export (Picture/Markdown), Chat with Internet (third-party)
  9. RAG/MCP & Artifact Strategy — Chat Platform artifact policies, RAG/MCP, side-by-side model/agent comparison
  10. Fine-Grained Management & Data Isolation — Admin Panel for user/group management, containerized data isolation ensuring data security

Development is nearing completion. Registration will be available once domain approval is finalized.

Screenshots

Unified Sign-In Welcome Page
Unified Sign-In Welcome Page
Casdoor-based unified login with password / verification code and multi-language support LibreChat landing page with file search, web search, code interpreter, and artifacts tools
Multi-Provider Model Selection Agent Marketplace
Multi-Provider Model Selection Agent Marketplace
Switch across 7 provider endpoints (Zhipu, DeepSeek, Aliyun, Kimi, Doubao, MiMo, MiniMax) with 165+ models Browse and use shared AI agents categorized by department (HR, R&D, Finance, IT, Sales, etc.)
Side-by-Side Model Comparison Channel Management Dashboard
Side-by-Side Comparison Channel Management
Compare responses from different models / agents side-by-side in real time NEW-API admin panel showing all 7 provider channels with status, latency, and usage metrics

Acknowledgements

This project stands on the shoulders of giants. Thanks to the following excellent open-source projects:

License

This project is licensed under the Apache License 2.0.

Scope

An internal AI chat integration repo based on NEW-API + LibreChat. The goal is to consolidate company-purchased upstream model capabilities into NEW-API, with LibreChat providing the chat interface for department members. Local entry points: http://localhost:3081 and http://localhost:13001.

Responsibilities

  • Codex handles full implementation, structural organization, documentation, script generation, configuration templates, integration path design, self-testing, and issue resolution.
  • The project owner is responsible for final acceptance only.
  • Users provide valid API keys for Zhipu, DeepSeek, Alibaba Cloud Bailian, Kimi, Volcano Doubao, Xiaomi MiMo, or MiniMax.
  • Final acceptance uses the Zhipu API channel as the primary verification channel.
  • Users are not responsible for engineering integration beyond filling in real keys and performing final acceptance.

Architecture

Zhipu / DeepSeek / Alibaba Cloud Bailian / Kimi / Volcano Doubao / Xiaomi MiMo / MiniMax / Other Upstream Models
    -> NEW-API
    -> LibreChat
    -> Department Users

This repo uses by default:

  • NEW-API native ZhipuV4 channel + Volcano Doubao native channel + DeepSeek / Alibaba Cloud Bailian / Kimi / Xiaomi MiMo / MiniMax OpenAI-compatible channels
  • LibreChat custom OpenAI-compatible endpoints split by provider: API-zhipu / API-deepseek / API-aliyun / API-kimi / API-doubao / API-mimo / API-minimax, all routing through NEW-API
  • Docker Compose as the unified orchestration for local and single-node production
  • Production defaults to direct-ip port access; Caddy is only enabled for domain HTTPS reverse proxy when COMPOSE_PROFILES=domain-proxy

Version Matrix

  • calciumion/new-api:v0.12.1
  • ghcr.io/danny-avila/librechat:v0.8.5
  • postgres:16-alpine
  • redis:7.4.2-alpine
  • mongo:8.0.20
  • caddy:2.10.0-alpine

Version rationale:

  • NEW-API v0.12.1 from official GitHub Release / Docker Hub tag.
  • LibreChat v0.8.5 from official Git tag and GHCR image tag.
  • All core base images are pinned to specific versions to avoid latest drift; the local Admin Panel is controlled by LIBRECHAT_ADMIN_PANEL_VERSION, defaulting to the upstream latest, and is not enabled in production compose by default.

Quick Start

  1. Copy the environment file and fill in at least one real upstream key: 
    cp .env.example .env
    Required: ZHIPU_API_KEY, DEEPSEEK_API_KEY, ALIYUN_API_KEY, KIMI_API_KEY, DOUBAO_API_KEY, MIMO_API_KEY, or MINIMAX_API_KEY. If you only want DeepSeek, Alibaba Cloud Bailian, Kimi, Doubao, MiMo, or MiniMax and no longer use Zhipu, set ZHIPU_ENABLED=false.
  2. Initialize local directories and auto-generate non-sensitive random keys: 
    make init
  3. Start core services: 
    make up
  4. Run Zhipu integration and main-chain smoke test: 
    make smoke-zhipu
    If DeepSeek is enabled: 
    make smoke-deepseek
    If Alibaba Cloud Bailian is enabled: 
    make smoke-aliyun
    If Kimi is enabled: 
    make smoke-kimi
    If Volcano Doubao is enabled: 
    make smoke-doubao
    If Xiaomi MiMo is enabled: 
    make smoke-mimo
    If MiniMax is enabled: 
    make smoke-minimax
  5. If you later adjust the upstream model matrix or configure LibreChat frontend whitelists: 
    make sync-provider-models

Notes:

  • make smoke-* commands all invoke scripts/bootstrap-new-api.sh first, which automatically initializes NEW-API, writes rate-limit parameters, creates enabled provider channels, and generates LibreChat service user and token.
  • The bootstrap process writes the auto-generated NEW_API_SERVICE_TOKEN back to the local .env — no manual token copying needed.
  • Local compose includes LibreChat Admin Panel by default, controlled by LIBRECHAT_ADMIN_PANEL_PORT; the main local deployment uses http://localhost:3003 to avoid conflicts with running branches and other local processes.
  • Main branch containers use ai-gateway-main-* naming to avoid overlapping with ai-gateway-* branch environments.
  • A default admin account is created (email and password configured via LIBRECHAT_DEFAULT_ADMIN_EMAIL / LIBRECHAT_DEFAULT_ADMIN_PASSWORD in .env), synced to LibreChat local user store and Casdoor team-ai organization; the first non-default registered user is automatically promoted to ADMIN.
  • LibreChat OIDC state/session is persisted to new-api-redis DB 1 — no repeated login prompts after restart.
  • Casdoor login page styles are script-rendered with browser light/dark theme adaptation; login/register pages default to Chinese.
  • Production compose now supports domain-less direct access: ports are exposed via LIBRECHAT_PUBLIC_URL / NEW_API_PUBLIC_URL / CASDOOR_PUBLIC_URL; Caddy is only started for domain HTTPS reverse proxy when COMPOSE_PROFILES=domain-proxy.
  • Production env templates and compose are tuned for Aliyun 2C2G with ~1.6GB usable memory; Admin Panel and Caddy are disabled by default in production.
  • When ZHIPU_ENABLED=true, DEEPSEEK_ENABLED=true, ALIYUN_ENABLED=true, KIMI_ENABLED=true, DOUBAO_ENABLED=true, MIMO_ENABLED=true, MINIMAX_ENABLED=true are all enabled, make bootstrap creates/updates zhipu-primary, deepseek-primary, aliyun-bailian-primary, kimi-primary, doubao-primary, mimo-primary, and minimax-primary channels; LibreChat displays API-zhipu / API-deepseek / API-aliyun / API-kimi / API-doubao / API-mimo / API-minimax as seven endpoints, all sharing the same NEW_API_SERVICE_TOKEN.

Directory Structure

docs/                     Requirements, architecture, deployment, acceptance, self-test docs
deploy/                   Local/prod compose and config templates
scripts/                  Init, start/stop, health check, bootstrap, backup/restore
tests/smoke/              Smoke test request templates
.github/workflows/        Basic CI checks
runtime/                  Local and production runtime data directory (not tracked)

Core Commands

  • make init: Initialize local directories, copy env files, generate random key placeholders.
  • make up: Start local compose.
  • make down: Stop local compose.
  • make restart: Restart local compose.
  • make bootstrap: Initialize NEW-API and configure enabled provider channels.
  • make bootstrap-librechat-admin: Create or fix LibreChat default admin and promote the first non-default registered user to ADMIN.
  • make sync-provider-models: Detect provider model APIs, sync NEW-API channels, and re-render LibreChat model list.
  • make sync-librechat-models: Compatibility alias, equivalent to make sync-provider-models.
  • make health: Check NEW-API and LibreChat application-layer health.
  • make smoke: Run generic smoke test.
  • make smoke-zhipu: Run Zhipu verification channel smoke test.
  • make smoke-deepseek: Run DeepSeek verification channel smoke test.
  • make smoke-aliyun: Run Alibaba Cloud Bailian verification channel smoke test.
  • make smoke-kimi: Run Kimi verification channel smoke test.
  • make smoke-doubao: Run Volcano Doubao verification channel smoke test.
  • make smoke-mimo: Run Xiaomi MiMo verification channel smoke test.
  • make smoke-minimax: Run MiniMax verification channel smoke test.
  • make doctor: Diagnose dependencies, ports, env, compose config.
  • make verify-no-secrets: Scan tracked files for obvious key risks.

Default deployment notes:

  • MeiliSearch has reserved variables but is disabled by default and not part of the acceptance main chain.
  • To enable search, restore the meilisearch service in compose and set LIBRECHAT_SEARCH_ENABLED=true.

Zhipu Configuration

Fill in at least the following fields in the root .env:

ZHIPU_ENABLED=true
ZHIPU_API_KEY=__FILL_BY_USER__
ZHIPU_API_BASE_URL=https://open.bigmodel.cn
ZHIPU_DEFAULT_MODEL=glm-5.1
ZHIPU_TEST_MODEL=glm-5.1
ZHIPU_EXPOSED_MODEL=glm-5.1,glm-5,glm-5-turbo,glm-4.7,glm-4.6,glm-4.5-air,glm-4.5
ZHIPU_LIBRECHAT_ENDPOINT_NAME=API-zhipu

The main config detects available models from the Zhipu model API, then syncs sorted ZHIPU_EXPOSED_MODEL to NEW-API and LibreChat's API-zhipu endpoint, achieving:

  • Frontend never holds official procurement keys
  • LibreChat only shows Zhipu models under API-zhipu
  • NEW-API can extend other upstreams without changing the frontend

NEW-API's ZhipuV4 channel auto-appends /api/paas/v4, so ZHIPU_API_BASE_URL must be https://open.bigmodel.cn — do not include the full path.

DeepSeek Configuration

Fill in at least the following fields in the root .env to enable DeepSeek:

DEEPSEEK_ENABLED=true
DEEPSEEK_API_KEY=__FILL_BY_USER__
DEEPSEEK_API_BASE_URL=https://api.deepseek.com
DEEPSEEK_DEFAULT_MODEL=deepseek-v4-flash
DEEPSEEK_TEST_MODEL=deepseek-v4-flash
DEEPSEEK_EXPOSED_MODEL=deepseek-v4-pro,deepseek-v4-flash
DEEPSEEK_LIBRECHAT_ENDPOINT_NAME=API-deepseek

Notes:

  • This repo uses type=1 OpenAI-compatible channel for DeepSeek official API.
  • make bootstrap creates/updates deepseek-primary channel when DEEPSEEK_ENABLED=true.
  • Bootstrap corrects deepseek-primary and other enabled provider channels' balance to NEW_API_PROVIDER_CHANNEL_BALANCE; no per-channel balance limits in this project.
  • DeepSeek official docs specify OpenAI-compatible base_url as https://api.deepseek.com — no need to manually append /v1.
  • make sync-provider-models dynamically refreshes DEEPSEEK_EXPOSED_MODEL from the DeepSeek model API; current real API returns deepseek-v4-pro / deepseek-v4-flash.

Alibaba Cloud Bailian Configuration

Fill in at least the following fields in the root .env to enable Bailian:

ALIYUN_ENABLED=true
ALIYUN_API_KEY=__FILL_BY_USER__
ALIYUN_API_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode
ALIYUN_DEFAULT_MODEL=qwen-plus
ALIYUN_TEST_MODEL=qwen-plus
ALIYUN_EXPOSED_MODEL=qwen3.6-max-preview,qwen3-max,qwen-max-latest,qwen-max,qwen-plus-latest,qwen-plus
ALIYUN_LIBRECHAT_ENDPOINT_NAME=API-aliyun

Notes:

  • This repo uses type=1 OpenAI-compatible channel for Alibaba Cloud DashScope.
  • make bootstrap creates/updates aliyun-bailian-primary channel when ALIYUN_ENABLED=true.
  • ALIYUN_API_BASE_URL is set to https://dashscope.aliyuncs.com/compatible-mode; model detection is configured separately via ALIYUN_MODEL_LIST_URLS=https://dashscope.aliyuncs.com/compatible-mode/v1/models.
  • make sync-provider-models dynamically refreshes ALIYUN_EXPOSED_MODEL from the Bailian model API, sorted by ALIYUN_MODEL_ORDER (higher-tier first).

Kimi Configuration

Fill in at least the following fields in the root .env to enable Kimi:

KIMI_ENABLED=true
KIMI_API_KEY=__FILL_BY_USER__
KIMI_API_BASE_URL=https://api.moonshot.cn
KIMI_DEFAULT_MODEL=kimi-k2.6
KIMI_TEST_MODEL=kimi-k2.6
KIMI_EXPOSED_MODEL=kimi-k2.6,kimi-k2.5,moonshot-v1-128k,moonshot-v1-32k,moonshot-v1-8k
KIMI_LIBRECHAT_ENDPOINT_NAME=API-kimi

Notes:

  • This repo uses type=1 OpenAI-compatible channel for Kimi Open Platform.
  • make bootstrap creates/updates kimi-primary channel when KIMI_ENABLED=true.
  • Kimi official SDK Base URL is https://api.moonshot.cn/v1; NEW-API channel uses https://api.moonshot.cn, with model detection via KIMI_MODEL_LIST_URLS=https://api.moonshot.cn/v1/models.
  • make sync-provider-models dynamically refreshes KIMI_EXPOSED_MODEL from the Kimi model API, sorted by KIMI_MODEL_ORDER (higher-tier first).

Volcano Doubao Configuration

Fill in at least the following fields in the root .env to enable Doubao:

DOUBAO_ENABLED=true
DOUBAO_API_KEY=__FILL_BY_USER__
DOUBAO_API_BASE_URL=https://ark.cn-beijing.volces.com
DOUBAO_DEFAULT_MODEL=doubao-seed-1-6-250615
DOUBAO_TEST_MODEL=doubao-seed-1-6-250615
DOUBAO_EXPOSED_MODEL=doubao-seed-1-6-250615,doubao-seed-1-6-flash-250828,doubao-1-5-pro-32k-250115
DOUBAO_LIBRECHAT_ENDPOINT_NAME=API-doubao

Notes:

  • This repo uses type=45 NEW-API Volcano Ark native channel for Doubao.
  • DOUBAO_API_BASE_URL is set to the root https://ark.cn-beijing.volces.com; the NEW-API adapter appends /api/v3/chat/completions.
  • Model detection is configured via DOUBAO_MODEL_LIST_URLS=https://ark.cn-beijing.volces.com/api/v3/models, filtering out disabled, embedding, image, audio/video, and other non-chat models.
  • Volcano Ark accounts must enable the corresponding model service or create a callable inference endpoint in the console; otherwise, the API key can list models but real chat requests will return "model service not enabled".

Xiaomi MiMo Configuration

Fill in at least the following fields in the root .env to enable MiMo:

MIMO_ENABLED=true
MIMO_API_KEY=__FILL_BY_USER__
MIMO_API_BASE_URL=https://api.xiaomimimo.com
MIMO_DEFAULT_MODEL=mimo-v2.5-pro
MIMO_TEST_MODEL=mimo-v2.5-pro
MIMO_EXPOSED_MODEL=mimo-v2.5-pro,mimo-v2.5,mimo-v2-pro,mimo-v2-omni,mimo-v2-flash
MIMO_LIBRECHAT_ENDPOINT_NAME=API-mimo

Notes:

  • This repo uses type=1 OpenAI-compatible channel for Xiaomi MiMo.
  • MiMo official SDK Base URL is https://api.xiaomimimo.com/v1; NEW-API channel uses https://api.xiaomimimo.com, with model detection via MIMO_MODEL_LIST_URLS=https://api.xiaomimimo.com/v1/models.
  • make sync-provider-models dynamically refreshes MIMO_EXPOSED_MODEL from the MiMo model API, filtering TTS/voiceclone/voicedesign non-chat models, sorted by MIMO_MODEL_ORDER (higher-tier first).

MiniMax Configuration

Fill in at least the following fields in the root .env to enable MiniMax:

MINIMAX_ENABLED=true
MINIMAX_API_KEY=__FILL_BY_USER__
MINIMAX_API_BASE_URL=https://api.minimaxi.com
MINIMAX_DEFAULT_MODEL=MiniMax-M2.7
MINIMAX_TEST_MODEL=MiniMax-M2.7
MINIMAX_EXPOSED_MODEL=MiniMax-M2.7,MiniMax-M2.7-highspeed,MiniMax-M2.5,MiniMax-M2.5-highspeed,MiniMax-M2.1,MiniMax-M2.1-highspeed,MiniMax-M2
MINIMAX_LIBRECHAT_ENDPOINT_NAME=API-minimax

Notes:

  • This repo uses type=1 OpenAI-compatible channel for MiniMax.
  • MiniMax official OpenAI SDK Base URL is https://api.minimaxi.com/v1; NEW-API channel uses https://api.minimaxi.com, with model detection via MINIMAX_MODEL_LIST_URLS=https://api.minimaxi.com/v1/models.
  • make sync-provider-models dynamically refreshes MINIMAX_EXPOSED_MODEL from the MiniMax model API, defaulting to MiniMax-M2.7, MiniMax-M2.7-highspeed, MiniMax-M2.5, MiniMax-M2.5-highspeed, MiniMax-M2.1, MiniMax-M2.1-highspeed, MiniMax-M2, sorted by MINIMAX_MODEL_ORDER (higher-tier first).

Model Sync

The default strategy splits LibreChat model endpoints by provider:

  • API-zhipu: Zhipu models only, sorted by ZHIPU_MODEL_ORDER (higher-tier first).
  • API-deepseek: DeepSeek models only, sorted by DEEPSEEK_MODEL_ORDER (higher-tier first).
  • API-aliyun: Alibaba Cloud Bailian models only, sorted by ALIYUN_MODEL_ORDER (higher-tier first).
  • API-kimi: Kimi models only, sorted by KIMI_MODEL_ORDER (higher-tier first).
  • API-doubao: Volcano Doubao models only, sorted by DOUBAO_MODEL_ORDER (higher-tier first).
  • API-mimo: Xiaomi MiMo models only, sorted by MIMO_MODEL_ORDER (higher-tier first).
  • API-minimax: MiniMax models only, sorted by MINIMAX_MODEL_ORDER (higher-tier first).
  • All endpoints share the same NEW_API_SERVICE_TOKEN for NEW-API /v1/* requests.

Recommended default settings:

  • NEW_API_SERVICE_TOKEN_QUOTA=1000000000000
  • NEW_API_SERVICE_TOKEN_UNLIMITED=true
  • NEW_API_PROVIDER_CHANNEL_BALANCE=999999999999
  • NEW_API_RATE_LIMIT_ENABLED=false
  • NEW_API_TOKEN_MODEL_LIMITS_ENABLED=false
  • NEW_API_SYNC_CHANNEL_MODELS_FROM_ENV=true
  • LIBRECHAT_SPLIT_PROVIDER_ENDPOINTS=true
  • LIBRECHAT_FETCH_MODELS=true
  • LIBRECHAT_VISIBLE_MODELS= (leave empty)

Daily dynamic sync:

make install-model-sync-cron

Default cron is 17 4 * * *, executing make sync-provider-models: detect provider model APIs, update *_EXPOSED_MODEL, replay bootstrap, and restart LibreChat.

LibreChat Default Admin

A local admin account is enabled by default for first-time login to LibreChat and Admin Panel:

LIBRECHAT_DEFAULT_ADMIN_ENABLED=true
LIBRECHAT_DEFAULT_ADMIN_EMAIL=<YOUR_ADMIN_EMAIL>
LIBRECHAT_DEFAULT_ADMIN_PASSWORD=<YOUR_ADMIN_PASSWORD>
LIBRECHAT_DEFAULT_ADMIN_CASDOOR_ENABLED=true
LIBRECHAT_FIRST_USER_ADMIN_ENABLED=true

Notes:

  • make up and make bootstrap auto-execute scripts/bootstrap-librechat-admin.sh.
  • The default admin is written to both LibreChat MongoDB and Casdoor team-ai organization, supporting both LibreChat local login and Casdoor unified auth login.
  • If auto-redirected to the unified auth page when logging into LibreChat, open http://localhost:3081/login?redirect=false to use the default admin account (see .env config).
  • The first non-default registered user is automatically set to ADMIN; subsequent users remain as USER for user group and permission debugging in Admin Panel.

Smoke Tests

  • General check: 
    make smoke
  • Zhipu primary verification: 
    make smoke-zhipu
  • DeepSeek primary verification: 
    make smoke-deepseek
  • Alibaba Cloud Bailian primary verification: 
    make smoke-aliyun
  • Kimi primary verification: 
    make smoke-kimi
  • Volcano Doubao primary verification: 
    make smoke-doubao
  • Xiaomi MiMo primary verification: 
    make smoke-mimo
  • MiniMax primary verification: 
    make smoke-minimax
  • Application-layer health check: 
    make health

Documentation

Recommended Reading Order

If you are new to this project, read in the following order:

  1. docs/architecture/requirements.md
  2. docs/architecture/architecture.md
  3. docs/architecture/implementation-plan.md
  4. docs/architecture/deployment-local.md or docs/architecture/deployment-cloud.md
  5. docs/architecture/admin-new-api.md and docs/architecture/admin-librechat.md
  6. docs/architecture/runbook.md
  7. docs/architecture/acceptance-criteria.md

About

Team AI Platform, the new chatbot hub.

Resources

Readme
Apache-2.0 license
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

Packages

Contributors

Languages