Integration · Model Providers

Connect model providers to Evose. Three categories: SaaS API, self-hosted, and OpenAI-compatible.

Three Integration Categories

Category	Examples	Integration point
SaaS API	OpenAI · Anthropic · Google · Aliyun · Tencent · ByteDance · Azure	Interface platform
Self-hosted	Llama · Qwen · DeepSeek · Mistral self-hosted	Model deployment
OpenAI-compatible	Any service that follows the OpenAI API format	Interface platform (pick OpenAI-compatible vendor)

OpenAI

Standard Setup

Platform: OpenAI US-West
Vendor: OpenAI
Base URL: https://api.openai.com/v1
API Keys:
  - sk-xxx (description: primary, weight: 1)
  - sk-yyy (description: backup, weight: 1)

Models: gpt-4-turbo / gpt-4o / gpt-4o-mini / text-embedding-3-large / ...

Anthropic Claude

Platform: Anthropic
Vendor: Anthropic
Base URL: https://api.anthropic.com/v1
API Keys: sk-ant-xxx
Models: claude-sonnet-4-6 / claude-opus-4-7 / claude-haiku-4-5

Azure OpenAI

Platform: Azure OpenAI East China
Vendor: Azure
Base URL: https://<your-resource>.openai.azure.com/openai
API Keys: <azure-key>
API Version: 2024-06-01
Models: gpt-4-turbo (matches the Azure deployment name)

Aliyun Tongyi

Platform: Aliyun Tongyi
Vendor: OpenAI-compatible
Base URL: https://dashscope.aliyuncs.com/compatible-mode/v1
API Keys: sk-xxx
Models: qwen-max / qwen-plus / qwen-turbo / text-embedding-v3

Self-Hosted Llama / Qwen / DeepSeek

Via Model deployment:

Model: deepseek-v3
Deployment:
  Instance 1: 10.0.0.5:8000 (replicas 4, GPU 4×A100)
  Instance 2: 10.0.0.6:8000 (replicas 4, GPU 4×A100)
Strategy: Round Robin
Health check: every 30s

Inference engine

Evose calls go through the OpenAI-compatible protocol. The inference engine can be vLLM / TGI / SGLang / LMDeploy, etc.

Multi-Vendor Failover

Model gpt-4-turbo:
  → OpenAI (priority 1)
  → Azure (priority 2, used when SaaS down)
  → Self-hosted OpenAI-compatible (priority 3, used when all SaaS down)

→ Interface platform · Routing

Credential Management

All API Keys should be registered in Credentials with field-level encryption.

Cost Attribution Under Private

No platform billing under Private, but costs can be attributed inside Evose:

[Model] Configure per-token cost (per your provider contract):
  gpt-4-turbo: in 0.01 / 1k tokens, out 0.03 / 1k tokens
[Usage] After Workflow execution, auto-bill by tokens × unit price
[Reports] In [Observability], view cost per Workspace / user / Agent

Anti-Patterns

Sharing the production key across all workspaces — issue per-workspace / per-project keys
No Failover — a single SaaS hiccup affects the whole company
Keeping the long-context model at 128K by default — costs spike; choose by scenario

Next Steps

Data connectors → Data connectors
Channel deployment → Channels

Integration · Model Providers

On this page