@i1hwan/apex-compatible

AI SDK provider adapter for ApexRoute. Extends @ai-sdk/openai-compatible with cache write token support via MetadataExtractor.

Why

When using Anthropic Claude through an OpenAI-compatible proxy like ApexRoute, the standard @ai-sdk/openai-compatible provider loses cache_creation_input_tokens (cache write) because the OpenAI spec has no field for it. This adapter extracts cache data from the raw response — before Zod strips unknown fields — and injects it into providerMetadata where consumers like OpenCode can read it.

What it fixes

| Token Data | @ai-sdk/openai-compatible | @i1hwan/apex-compatible | |---|---|---| | inputTokens.total | ✅ | ✅ | | inputTokens.cacheRead | ✅ | ✅ | | inputTokens.cacheWrite | ❌ always undefined | ✅ via providerMetadata | | outputTokens.reasoning | ✅ | ✅ |

Install

npm install @i1hwan/apex-compatible

Usage

import { createApexCompatible } from "@i1hwan/apex-compatible";

const provider = createApexCompatible({
  baseURL: "https://your-apexroute-instance.com/v1",
  apiKey: process.env.APEX_API_KEY,
});

const model = provider("claude-sonnet-4-20250514");

With OpenCode

In your opencode.json:

{
  "provider": {
    "apexroute": {
      "npm": "@i1hwan/apex-compatible",
      "options": {
        "baseURL": "https://your-apexroute-instance.com/v1",
        "apiKey": "{env:APEX_API_KEY}"
      }
    }
  }
}

How it works

ApexRoute already emits prompt_tokens_details.cache_creation_tokens in OpenAI-compatible responses. The standard @ai-sdk/openai-compatible provider drops this field during Zod schema validation.

This package uses the official metadataExtractor extension point to:

1. Read cache_creation_tokens and cached_tokens from raw response data (before Zod parse)
2. Inject them into providerMetadata under the anthropic key (for OpenCode compatibility)

ApexRoute → { prompt_tokens_details: { cache_creation_tokens: 152 } }
    ↓
MetadataExtractor.processChunk(rawChunk)  ← pre-Zod raw data
    ↓
providerMetadata: { anthropic: { cacheCreationInputTokens: 152 } }
    ↓
OpenCode getUsage() → Cache Write: 152 ✅

No server-side changes required. No fork of @ai-sdk/openai-compatible. ~50 lines of code.

API

createApexCompatible(options)

Creates an OpenAI-compatible provider with ApexRoute metadata extraction pre-configured.

interface ApexCompatibleSettings {
  baseURL: string;
  name?: string;           // default: "apex-compatible"
  apiKey?: string;
  headers?: Record<string, string>;
  queryParams?: Record<string, string>;
  fetch?: FetchFunction;
  supportsStructuredOutputs?: boolean;
  transformRequestBody?: (args: Record<string, any>) => Record<string, any>;
}

createApexMetadataExtractor()

Standalone metadata extractor — use if you want to configure createOpenAICompatible yourself:

import { createOpenAICompatible } from "@ai-sdk/openai-compatible";
import { createApexMetadataExtractor } from "@i1hwan/apex-compatible";

const provider = createOpenAICompatible({
  baseURL: "https://your-proxy.com/v1",
  name: "my-provider",
  includeUsage: true,
  metadataExtractor: createApexMetadataExtractor(),
});

Related

• ApexRoute — Private OmniRoute fork with lexical rewrite, thinking effort control, and cache token translation
• Vercel AI SDK — The AI SDK this adapter extends
• OpenCode — CLI that consumes this provider

License

MIT