OpenAI to Claude Migration

Point this agent at your OpenAI-integrated codebase and ask it to plan your migration to the Claude API.

What it does

1. Map every OpenAI call site: chat completions and responses API usage, embeddings, function calling, structured outputs, streaming, batch jobs, and assistants — with the model and parameters used at each site.
2. Plan the migration: model mapping to current Claude models, parameter mapping (max_tokens, temperature, stop sequences, system prompt placement), tool/function schema translation, response-format strategy, and cost/latency tier guidance per workload.
3. Scaffold the code changes: SDK swap per language, messages-array translation including system placement and multimodal blocks, tool-use schema and tool_result loop translation, and streaming event mapping.
4. Parallel-run eval — run a paired-prompt harness on a sample set against both APIs, diff outputs with rubric scoring, and report PASS/PARTIAL/FAIL drift evidence per workload.

Prerequisites

• ANTHROPIC_API_KEY — Claude API key used for scaffold verification and the eval harness
• OPENAI_API_KEY — optional; enables side-by-side parallel-run diffing against the source workloads

Current scope

This solution migrates chat/completions-style workloads, tool use, structured outputs, streaming, and batch jobs. It does not migrate embeddings (the Claude API has no embeddings endpoint — call sites are flagged and a third-party provider is recommended), does not translate the assistants API mechanically (it is re-mapped to agent patterns), and does not translate fine-tuned GPT models (they need re-evaluation on Claude, not translation). Prompt quality tuning beyond the eval harness's drift report stays with the customer.