Fix Claude Code DeepSeek Config Routing | Generated by AI

Home 2025.08

Got you. Your config is sending Claude Code’s Anthropic-style traffic into OpenRouter’s OpenAI-style endpoint for DeepSeek V3.1. When Claude Code (the IDE agent) tries to do tools/plan/stream, the message schema doesn’t line up → errors.

Below is a clean way to wire it so Anthropic models still use an Anthropic-style gateway, while DeepSeek V3.1 (OpenAI-style) stays on /chat/completions. I’ll show the exact edits you can paste.

1) Verify the DeepSeek V3.1 slug

OpenRouter’s chat model slug is:

deepseek/deepseek-chat-v3.1

(That’s the instruct/chat model—distinct from the “base” model.) (OpenRouter)

2) Add an Anthropic-format proxy for Claude Code

Claude Code expects Anthropic Messages (/messages, tool_use blocks). The simplest fix is to run a tiny proxy that converts Anthropic→OpenAI (OpenRouter) and back again. There’s a ready-made one: maxnowack/anthropic-proxy. Start it like this (env var is your OpenRouter key):

OPENROUTER_API_KEY=YOUR_KEY npx anthropic-proxy

It listens locally and translates schemas for you. (GitHub)

Now, add a second provider that points Claude Code traffic to that proxy.

🔧 Replace your JSON with these focused changes

Add this provider (keep your current openrouter provider as-is):

{
  "name": "anthropic-proxy",
  "api_base_url": "http://127.0.0.1:3000/v1/messages",
  "api_key": "", 
  "models": [
    "anthropic/claude-sonnet-4",
    "anthropic/claude-3.5-sonnet",
    "anthropic/claude-3.7-sonnet:thinking",
    "anthropic/claude-opus-4"
  ]
}

Keep your existing OpenRouter provider (it’s correct to use /chat/completions for DeepSeek and other OpenAI-schema models). OpenRouter normalizes to an OpenAI-compatible schema and supports extras like reasoning tokens, tools, etc. (OpenRouter)

3) Route models to the right provider

Point Anthropic models to the anthropic-proxy; point DeepSeek/GPT/Gemini/Qwen to openrouter.

Example Router block you can drop in:

"Router": {
  "default": "openrouter,openai/gpt-5-mini",

  "background": "openrouter,google/gemini-2.5-flash",
  "think": "openrouter,qwen/qwen3-coder",
  "longContext": "openrouter,moonshotai/kimi-k2",
  "longContextThreshold": 2000,
  "webSearch": "openrouter,mistralai/mistral-medium-3.1",

  "overrides": {
    "anthropic/claude-sonnet-4": "anthropic-proxy,anthropic/claude-sonnet-4",
    "anthropic/claude-3.5-sonnet": "anthropic-proxy,anthropic/claude-3.5-sonnet",
    "anthropic/claude-3.7-sonnet:thinking": "anthropic-proxy,anthropic/claude-3.7-sonnet:thinking",
    "anthropic/claude-opus-4": "anthropic-proxy,anthropic/claude-opus-4",

    "deepseek/deepseek-chat-v3.1": "openrouter,deepseek/deepseek-chat-v3.1"
  }
}

If your router doesn’t support an overrides map, the equivalent is: create separate “profiles” in your app (e.g., “Claude Code via anthropic-proxy” vs “General chat via openrouter”) and pick the profile per action.

4) Avoid reasoning stream collisions on DeepSeek (optional but safer)

DeepSeek V3.1 supports “thinking” tokens. Many wrappers choke if those appear mid-stream. On OpenRouter you can exclude reasoning with a request flag. If your router lets you set per-model params, attach:

"params": {
  "reasoning": { "exclude": true }
}

(or include_reasoning: false depending on your router’s param naming). This prevents <think>/reasoning content from showing up where your Claude Code UI expects plain text. (OpenRouter)

5) Make the minimal edits directly in your file

Here’s your config with only the necessary deltas:

After that, when you pick DeepSeek V3.1 in your tool, it goes through OpenRouter’s OpenAI-style path. When you pick any Claude model, it flows through anthropic-proxy, which converts the schema so Claude Code’s tool use works.

Why this works

If you hit a specific error string after this (e.g., JSON pointer errors, “unknown field tool_use”, or SSE parse issues), paste it and I’ll tweak the per-model params (e.g., forcing reasoning.exclude, capping max_tokens, or adjusting tool schema) to match your router’s exact expectations.

Back Donate