TokenLink — AI Model API Relay Platform

Protocol adaptation is TokenLink's core capability. It lets you call any model with any Agent tool — whether or not they are natively compatible.

How it works

Client (Claude Code)              TokenLink                     Upstream (DeepSeek)
        │                            │                               │
        │  Anthropic Messages API    │                               │
        │  /anthropic/v1/messages    │                               │
        ├───────────────────────────►│                               │
        │                            │  Convert to OpenAI Chat API   │
        │                            │  /chat/completions            │
        │                            ├──────────────────────────────►│
        │                            │                               │
        │                            │       OpenAI Chat Response    │
        │                            │◄──────────────────────────────┤
        │                            │                               │
        │  Convert to Anthropic      │                               │
        │  Message Response          │                               │
        │◄───────────────────────────┤                               │
        │                            │                               │

After TokenLink receives a client request:

Detect protocol — determine the client's format from the URL path (/anthropic/*, /openai/*, /gemini/*)
Match upstream — find the corresponding upstream channel by model name
Convert format — if the client format ≠ the upstream format, automatically convert the request and response
Forward transparently — completely invisible to the client, just like calling the native API

Supported conversion paths

Request conversion

Client format	→	Upstream format	Scenario
Anthropic Messages	→	OpenAI Chat	Claude Code calling DeepSeek / GPT-5
OpenAI Chat	→	Anthropic Messages	Codex calling Claude
OpenAI Chat	→	Gemini	OpenAI SDK tools calling Gemini
Gemini	→	OpenAI Chat	Gemini CLI calling DeepSeek
OpenAI Responses	→	OpenAI Chat	Responses API calling a Chat model
OpenAI Chat	→	OpenAI Responses	Chat request converted to the new Responses format

Streaming conversion

Streaming (SSE) requests are also supported. Each SSE event is converted from the upstream format to the client format as it is forwarded.

Specific supported conversions:

Anthropic SSE → OpenAI SSE: Claude streaming events converted to OpenAI format
OpenAI SSE → Anthropic SSE: OpenAI streaming converted to Anthropic format
Gemini SSE → OpenAI SSE: Gemini streaming converted to OpenAI format
OpenAI Chat SSE → OpenAI Responses SSE: Chat streaming events converted to the Responses format (including event: lines)

Native protocol (Passthrough)

If the upstream channel natively supports the protocol format of the client request, TokenLink passes it through directly without any format conversion.

For example: when Claude Code sends an Anthropic-format request to call a real Claude model, the upstream is already native Anthropic format → forwarded directly, with zero overhead.

A channel's protocol support is configured via the protocols field:

Protocol	Description
`chat`	OpenAI Chat Completions API
`responses`	OpenAI Responses API
`anthropic`	Anthropic Messages API
`gemini`	Google Gemini generateContent API

Model compatibility

Model type	Chat API	Embedding	Image	Audio
GPT-5	✅	✅	✅	✅
Claude	✅	—	—	—
Gemini	✅	—	—	—
DeepSeek	✅	—	—	—
text-embedding-*	—	✅	—	—
dall-e-*	—	—	✅	—
tts-/whisper-	—	—	—	✅

What this means for developers

You don't need to worry about any of these conversion details. Just remember one thing:

Use Claude Code to call DeepSeek, Codex to call Gemini, Gemini CLI to call GPT-5 — it all works out of the box.