AI 后端基础设施

为您的业务构建 AI 工具。. 让您的业务始终掌握控制权。.

构建客户支持、CRM 助手、文档处理等 AI 功能，无需自建 AI 基础设施。. 一次连接所有主流 AI 模型，同时由您的后端决定 AI 能看什么、能做什么。.

免费开始了解工作原理

免费开始 · 无需信用卡

13 家 AI 供应商 98 款模型 13 个现成模板随时切换供应商

客户请求

AI 智能工具

ModelRiver AI 后端处理中

控制上下文会话允许的工具模型

检查可用性就绪

查找最佳座位就绪

预订座位待处理

发送确认就绪

您的后端

场地 API 定价日历

Best seats secured — Front row, Center

Premium upgrade available

构建客户支持 AI、CRM 助手等

从经过验证的工作流开始，而不是从零搭建一切。.

客户支持 AI

用了解您产品的 AI 解决工单、回答客户问题。.

浏览使用场景

我的发票在哪里？?

已发送 — 请查看 [email protected] 已解决

AI CRM 助手

自动评估线索、更新记录并草拟跟进内容。.

浏览模板

Acme 公司. 潜在客户 · 网站表单

已自动草拟跟进内容

文档处理

通过结构化输出自动提取、分类和总结文档。.

为什么团队选择用 ModelRiver 构建

看看团队如何使用 ModelRiver

先测试 AI，再付费

在向 AI 供应商付费之前，先用示例响应测试工作流。.

打开 Playground

Test workflow

Sample JSON in, structured response out

Send request

Workflow

book_info

{ "messages": [{ "role": "user", "content": "Percy Jackson" }] }

Response

{
  "title": "Percy Jackson & the Olympians",
  "author": "Rick Riordan",
  "genre": "Fantasy"
}

一次连接，使用任意 AI

一次集成即可对接 OpenAI、Anthropic、Gemini 等。直接在控制台切换供应商。.

Structured outputs

JSON schemas for responses

sample 4 properties

book_notes 3 properties

让 AI 成本保持可控

每日预算。实时用量。账单不意外。.

设置预算

customer_support_workflow

Live

Primary model $20/day

gpt-5.4-mini

$7 of $20 used$13 left today

Backup model $10/day

claude-haiku-4-5

$2 of $10 used$8 left today

查看请求的每一步

跟踪 AI 请求在每条后端管道中的流转，直到最终响应。.

查看请求日志

Request details

ticket_creat… OpenAI · gpt-5-mini 140d ago

Failed OpenAI gpt-5-mini — rate limit exceeded 0ms

Failover Retrying Anthropic claude-haiku-4-5 1.2s

Success Structured output delivered to webhook 8.4s

一个连接。所有 AI 模型。. You stay in 控制 of your data.

您的应用只需连接一次 ModelRiver。ModelRiver 负责对接 AI 供应商、重试请求并运行多步工作流。您的业务逻辑始终留在自己的系统中。.

了解更多请求流程

您的应用

POST /v1/chat/completions POST /v1/ai POST /v1/ai/async POST /v1/callback/:channel_id

ModelRiver

后端管道 AI 编排自动故障转移输入护栏支出上限测试缓存成本分析会话记忆

您的后端与数据

您的 CRM 您的数据库（Supabase • Neon • Postgres）您的业务规则您的 API 密钥与机密信息

所有主流 AI 供应商

您的数据和 API 密钥由您掌控。ModelRiver 负责路由、重试和供应商管理。.

为什么不直接把任务交给 AI 代理？?

AI 代理往往需要广泛权限才能发挥最佳效果。ModelRiver 则让您的后端来决定 AI 究竟能获得什么。.

AI 代理

ModelRiver

您的数据流向哪里

可能需要访问您的系统和数据的广泛权限

由您的后端精确决定 AI 接收哪些信息

构建 AI 工作流

您需要自行编排重试、多次 AI 调用和集成

内置后端管道协同 AI 与您的代码

切换供应商

通常需要工程开发和重新部署

在控制台即可切换供应商

AI 成本

往往把每个任务都发给较大的 AI 模型

为每个任务选择合适的 AI 模型，有助于降低成本

AI 费用由谁承担

您直接向 AI 供应商付费

同样如此。您直接向 AI 供应商付费，我们只收取 ModelRiver 平台费用。.

看看 ModelRiver 的实际运行

支撑您 AI 功能的控制台 — 供应商、工作流、预算和日志。.

console.northstar.ai/projects/…/workflows

Backend code Frontend code View integration examples for your workflows

Workflows

Configure AI workflows with provider, model, and response options

Create workflow

mr_chatbot_workflow Live

Customer support chatbot with structured replies and failover.

Test Copy

Primary model

gpt-4o-mini

Budget No budget

Backup models if primary fails

Backup 1

claude-haiku-4-5-20251001

Budget No budget

get_book_suggestion Live

Translate this Sanskrit medical text into Sumerian cuneiform notation. Preserve the original meaning with extreme precision.

Test Copy

Primary model

gpt-3.5-turbo

Budget $1.00/day

0% of budget used

$0.00 / $1.00 per day · $1.00 remaining

Backup models if primary fails

Backup 1

claude-3-haiku-20240307

Budget $1.00/day

0% of budget used

$0.00 / $1.00 per day · $1.00 remaining

summarize_ticket Live

Summarize support tickets into actionable next steps.

Test Copy

Primary model

gemini-2.5-flash

Budget No budget

Backup models if primary fails

Backup 1

gpt-4o-mini

Budget No budget

content_moderator Live

Classify user content and route low-confidence cases.

Test Copy

Primary model

grok-4.3

Budget $1.00/day

12% of budget used

$0.12 / $1.00 per day · $0.88 remaining

Backup models if primary fails

Backup 1

claude-haiku-4-5

Budget No budget

Provider credentials

Add your API keys for OpenAI, Anthropic, and other providers

+ Custom provider

Anthropic

Claude 4.5 Sonnet, Claude 4.5 Opus, Haiku 4.5

Connected

Update

OpenAI

GPT-5, GPT-5 pro, GPT-5 mini

Connected

Update

Cohere

Command-A-03-2025, Command R, Embed

Not configured

Add API Key

DeepSeek

DeepSeek-V3 (Chat) and DeepSeek-R1 (Reasoner)

Not configured

Add API Key

Google

Gemini 2.5 Flash, Gemini 2.5 Pro, 2.5 Flash-Lite

Not configured

Add API Key

Groq

Ultra-fast LPU inference - Llama 3.1/3.3, GPT OSS

Not configured

Add API Key

Mistral AI

Mistral Large 24.11, Small 3.2, Codestral

Not configured

Add API Key

Qwen

Qwen-Max, Qwen-Plus, Qwen-Turbo, Qwen-Long (Alibaba Cloud)

Not configured

Add API Key

xAI

Grok 4, Grok-4-fast-reasoning

Not configured

Add API Key

Async mode (background job)

Process AI requests in the background with real-time updates via WebSocket. Works even if you refresh the page!

Workflow

my_workflow

LIVE Production settings

Choose how this playground request runs before you send it.

Edit workflow

Run with model

Pin one model for this test, or use Auto to follow your configured order.

Auto Follow workflow order (primary, then backups)

claude-haiku-4-5-20251001 Primary model

gpt-5.4-mini Backup 1

grok-4.20-0309-non-reasoning Backup 2

Pipeline steps

Backend pipeline flow for this test — pin a model on linked workflows when needed.

1

movie_suggestions BACKEND WEBHOOK

Backend webhook + auto callback

2

movie_ratings AI WORKFLOW

json_extractor · gpt-5.4-mini

Run linked workflow with model

Auto Follow workflow order (primary, then backups)

gpt-5.4-mini Primary model

gemini-2.5-flash-lite Backup 1

3

final_movies FINAL CALLBACK

Backend callback, then deliver

Response format

book_notes

CACHE

15m cache window

Cache skipped — 15m cache window disabled in playground for fresh responses

Backend pipeline test — Backend steps use auto callbacks; linked workflow models follow your selections above

JSON payload

1{

2 "messages": [

3 {"role": "user", "content": "Hello from Northstar AI"}

4 ]

5}

Response

Raw JSON Preview

1{

2 "data": {

3 "author": "Northstar AI",

4 "title": "The Midnight Library",

5 "notes": [

6 "Philosophical fiction exploring life's infinite possibilities",

7 "Strong character development and emotional resonance"

8 ],

9 "schema": {

10 "type": "json_schema",

11 "strict": true,

12 "properties": { "type": "object" }

13 },

14 "structure_id": "cddd219a-6e55-42b3-8f98-d343fe5f7f72"

15 },

16 "status": "success",

17 "channel_id": "628f6a0d-ec75-4a8c-be7c-4208cafe6b72",

18 "customer_data": {}

19}

Back to request logs

Request details

OpenAI gpt-5-mini 140d ago

ticket_creat... Success

Input tokens

468

Output tokens

368

Duration

8.4s

Price

$0.000853

Cache reuse

0

No downstream cache hits

Response source

Provider

Cached from request ID

N/A

Estimated price without cache

N/A

Timeline logs

Failed

anthropic/claude-haiku-4-5

661ms • 141d ago

Success

openai/gpt-5-mini

7.7s • 141d ago

Error

Trigger webhook

141d ago • 3ms • Callback: success

Success

Backend callback

141d ago

Success

WebSocket output

141d ago

Webhook Deliveries

Failed WebSocket 3ms Attempt 1

URL: cli://localhost CLI

No CLI client connected

Members

15 members

+ Invite member

Michael Carter Owner

[email protected]

Joined Nov 14, 2025

Emily Johnson Member

[email protected]

Joined Nov 18, 2025

James Wilson Member

[email protected]

Joined Dec 2, 2025

Sarah Martinez Member

[email protected]

Joined Dec 15, 2025

Robert Thompson Member

[email protected]

Joined Jan 8, 2026

Jennifer Lee Member

[email protected]

Joined Jan 22, 2026

David Brown Member

[email protected]

Joined Feb 3, 2026

Amanda Davis Member

[email protected]

Joined Feb 19, 2026

Christopher Miller Member

[email protected]

Joined Mar 1, 2026

Jessica Taylor Member

[email protected]

Joined Mar 14, 2026

Matthew Anderson Member

[email protected]

Joined Apr 2, 2026

Ashley Thomas Member

[email protected]

Joined Apr 18, 2026

Ryan Garcia Member

[email protected]

Joined May 6, 2026

Nicole White Member

[email protected]

Joined May 21, 2026

Brandon Clark Member

[email protected]

Joined Jun 3, 2026

Pending invitations

No invitations awaiting acceptance

No pending invitations

实时演示 — 查看 AI 请求如何流经供应商、故障转移、预算和日志

开始构建 AI 驱动的业务工具。.

免费开始