rosetta-llm 0.1.0

Multi-format bidirectional LLM proxy — translates between OpenAI Chat, OpenAI Responses, and Anthropic Messages

Rosetta

Multi-format bidirectional translation proxy for LLM APIs. Translates between OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages formats — letting any client SDK talk to any provider regardless of which native API the provider speaks.

Quick Start

uvx (no install required)

# Create your config
mkdir -p ~/.rosetta-llm
cp config.example.jsonc ~/.rosetta-llm/config.json
# Edit ~/.rosetta-llm/config.json with your providers and API keys

# Run instantly
uvx rosetta-llm

Or point to a custom config:

uvx rosetta-llm --config /path/to/config.json

# Equivalent via env var:
ROSETTA_CONFIG=/path/to/config.json uvx rosetta-llm

uv tool install (persistent)

uv tool install rosetta-llm
rosetta-llm --help
rosetta-llm --config ~/my-config.json --port 9999

Docker

docker run -p 7860:7860 \
  -v ~/.rosetta-llm/config.json:/app/config.json \
  -e ANTHROPIC_API_KEY=sk-ant-... \
  -e OPENAI_API_KEY=sk-... \
  ghcr.io/lokesh-chimakurthi/rosetta-llm:main

From source

git clone https://github.com/Lokesh-Chimakurthi/rosetta-llm.git
cd rosetta-llm
uv sync
python -m rosetta

Features

• Three endpoint families: /v1/chat/completions, /v1/responses, /v1/messages — all with full streaming support
• Passthrough fast path: zero-overhead when inbound format matches provider's native format
• Canonical IR translation: lossless cross-format translation including thinking blocks, tool calls, and reasoning
• Claude Code gateway: full model picker integration — non-Anthropic models appear in /model via claude-code/ prefixing
• Provider routing: <provider>/<model> prefix scheme (e.g., anthropic/claude-opus-4-7)
• Bearer-token auth: optional proxy-level API key authentication
• Structured JSON logging: request-scoped with configurable log levels
• Docker support: multi-stage build with python:3.13-slim

Endpoints

Method	Path	Purpose
POST	/v1/messages	Anthropic Messages
POST	/v1/messages/count_tokens	Local tiktoken token count
POST	/v1/chat/completions	OpenAI Chat Completions
POST	/v1/responses	OpenAI Responses
GET	/v1/models	Merged model list
GET	/health	Liveness check
GET	/providers	Provider status

Model ID Format

Models are addressed as <provider_key>/<model_name> where provider_key matches a key in your config's providers section.

Examples:

• abc/kimi-k2.5 — routes to the "abc" provider with model "kimi-k2.5"
• anthropic/claude-opus-4-7 — routes to the "anthropic" provider
• openai/gpt-5.4 — routes to the "openai" provider

Claude Code Integration

Rosetta is a fully compatible Claude Code LLM gateway. Point Claude Code at Rosetta and all configured providers appear in the /model picker — including non-Anthropic models.

Setup

export ANTHROPIC_BASE_URL=http://localhost:7860
export ANTHROPIC_AUTH_TOKEN=sk-proxy-XXXX   # if proxy auth is enabled

Or in Claude Code settings (~/.claude/settings.json):

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://localhost:7860",
    "ANTHROPIC_AUTH_TOKEN": "sk-proxy-XXXX"
  }
}

Model picker

On startup, Claude Code queries GET /v1/models with its session headers. Rosetta detects Claude Code (via the X-Claude-Code-Session-Id header) and returns a model list tailored for the picker:

• Models already named claude-* or anthropic/* pass through unchanged
• All other models get a claude-code/ prefix — this ensures they pass Claude Code's built-in model filter (which only shows models starting with claude or anthropic)

For example, if your config has an OpenAI provider with gpt-5.4, it appears in the picker as claude-code/openai/gpt-5.4. When selected, Rosetta strips the claude-code/ prefix internally and routes to the correct provider.

Headers forwarded upstream

Rosetta forwards Claude Code's session headers (anthropic-beta, anthropic-version, X-Claude-Code-Session-Id) to every upstream call, preserving prompt caching and feature detection.

Translation Matrix

The proxy automatically translates between formats:

Client Endpoint	Provider Format	Path
/v1/messages	anthropic	passthrough
/v1/messages	openai_chat	translate via IR
/v1/messages	openai_responses	translate via IR
/v1/chat/completions	openai_chat	passthrough
/v1/chat/completions	anthropic	translate via IR
/v1/chat/completions	openai_responses	translate via IR
/v1/responses	openai_responses	passthrough
/v1/responses	anthropic	translate via IR
/v1/responses	openai_chat	translate via IR

Usage Examples

Anthropic client -> OpenAI-backed model

curl http://localhost:7860/v1/messages \
  -H "Authorization: Bearer sk-proxy-XXXX" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-5.4",
    "max_tokens": 256,
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

OpenAI client -> Anthropic-backed model

curl http://localhost:7860/v1/chat/completions \
  -H "Authorization: Bearer sk-proxy-XXXX" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-opus-4-7",
    "messages": [{"role": "user", "content": "Hello!"}],
    "stream": true
  }'

Environment Variables

Variable	Purpose
ROSETTA_CONFIG	Path to config.json (default: ~/.rosetta-llm/config.json)
Provider-specific	Set via api_key_env in config (e.g., ANTHROPIC_API_KEY)

Development

uv sync --group dev
uv run pytest -q
uv run mypy src/
uv run ruff check src/ tests/