Claude Anthropic Proxy

A small TypeScript service that looks like Anthropic to Claude Code and talks to an OpenAI-compatible or Ollama backend underneath.

This repo is intentionally MVP-first:

• POST /v1/messages
• POST /v1/messages/count_tokens
• GET /health
• Text-only request/response translation
• Anthropic-style SSE streaming for stream: true

What is not supported yet:

• Tool calling
• Images
• Prompt caching semantics
• Exact Anthropic token accounting

Why this shape

Claude Code expects Anthropic-flavored JSON and Anthropic-flavored server-sent events. This proxy keeps Claude Code untouched and concentrates the protocol weirdness in one service:

• Anthropic request validation and normalization
• Model mapping from Anthropic names to backend names
• Provider adapters for OpenAI-compatible servers or Ollama
• Anthropic response and SSE event translation
• Rough token counting for /v1/messages/count_tokens

Quick start

1. Install dependencies:

npm install

2. Create your env file:

cp .env.example .env

3. Point the provider at your backend and adjust the model map.

Example for an OpenAI-compatible local server:

PROVIDER=openai
PROVIDER_BASE_URL=http://127.0.0.1:8000/v1
MODEL_MAP_JSON={"claude-sonnet-4-5":"qwen2.5-coder-32b-instruct"}

Example for Ollama:

PROVIDER=ollama
PROVIDER_BASE_URL=http://127.0.0.1:11434
MODEL_MAP_JSON={"claude-sonnet-4-5":"qwen2.5-coder:32b"}

4. Run the proxy:

npm run dev

5. Point Claude Code at it:

ANTHROPIC_BASE_URL=http://127.0.0.1:8080 \
ANTHROPIC_API_KEY=dummy \
claude --bare --tools "" "hello"

Config

Supported environment variables:

• PORT
• HOST
• LOG_LEVEL
• PROVIDER=openai|ollama
• PROVIDER_BASE_URL
• PROVIDER_API_KEY
• DEFAULT_MODEL
• MODEL_MAP_JSON
• MODEL_MAP_FILE
• PROVIDER_TIMEOUT_MS

MODEL_MAP_JSON should be a JSON object that maps the Anthropic model names Claude Code sends to the provider model names your backend understands.

API behavior

POST /v1/messages

Accepts Anthropic-style messages requests and returns an Anthropic-style message response.

Streaming requests emit the Anthropic event order Claude Code expects:

1. message_start
2. content_block_start
3. repeated content_block_delta
4. content_block_stop
5. message_delta
6. message_stop

POST /v1/messages/count_tokens

Returns a rough input_tokens estimate. This is good enough for an MVP, but it is not exact Anthropic accounting.

GET /health

Returns a simple readiness payload:

{
  "status": "ok"
}

Project structure

src/
  app.ts
  server.ts
  config/
  providers/
  routes/
  schemas/
  services/
  translators/
  tests/
  utils/

Roadmap

MVP is working when:

• Claude Code can send text-only prompts through this proxy
• streaming works
• a mapped backend model answers successfully

The next major step is tool-use support:

• map Anthropic tools to backend tools
• map backend tool calls back to tool_use
• accept tool_result blocks on follow-up turns
• preserve tool call IDs exactly

After that:

• improve token counting with model-specific tokenizers
• broaden provider coverage
• add request tracing and richer observability

Development

Useful commands:

npm run dev
npm run typecheck
npm run test
npm run build