llm-babel

Translate an LLM request, response, error, or stream between provider wire formats, and get the exact same result whether it runs in Python, Rust, or TypeScript.

Every provider speaks a slightly different dialect on the wire. llm-babel routes everything through one neutral intermediate representation, so a format is one decoder plus one encoder rather than a converter per pair. The transforms are authored once as a small pure spec and compiled to plain Python, Rust, and TypeScript, proven byte-identical against a shared golden corpus every release has to pass in all three.

Formats: openai-chat, openai-responses, anthropic-messages, bedrock-converse, gemini, mistral, deepseek, cohere, plus the OpenAI-compatible providers (Groq, Cerebras, xAI, Z.AI, OpenRouter, HuggingFace) as routes over the openai-chat codec.

Install

pip install llm-babel                # the runtime: pure stdlib, no dependencies
pip install "llm-babel[validation]"  # adds Pydantic for the typed, validating IR surface

The import name is llm_transform.

Use

from llm_transform.registry import translate, decode, encode

translate("openai-chat", "anthropic-messages", openai_request)  # source wire -> target wire
decode("gemini", gemini_request)                                # any wire -> IR
encode("bedrock-converse", ir)                                  # IR -> any wire

Responses, errors, and streams have the same shape of API (translate_response, translate_error, fold_stream), and reconcile fits a request to a target's declared capabilities before encoding.

The same transforms are published for TypeScript as @pydantic/llm-babel on npm.

The golden corpus is not in this wheel, since it proves the transforms rather than running them. To re-run it yourself, take the sdist (pip download --no-binary :all: llm-babel) or the repository, where every backend runs it.