claude-rt

Transparent proxy for Claude Code. Anthropic traffic passes through unchanged, gpt-* models get routed to OpenAI.

Architecture

graph LR
  subgraph Claude Code Agents
    O["Orchestrator<br/><small>Claude Opus</small>"]
    S1["SubAgent<br/><small>Claude Haiku</small>"]
    S2["SubAgent<br/><small>GPT-5</small>"]
    T["Team Member<br/><small>GPT-5</small>"]
  end

  RT["<b>claude-rt</b><br/><small>:8787</small>"]

  A["Anthropic API<br/><small>passthrough</small>"]
  OA["OpenAI API<br/><small>transform</small>"]

  O --> RT
  S1 --> RT
  S2 --> RT
  T --> RT
  RT -->|raw bytes| A
  RT -->|format convert| OA

Typical setup: Opus orchestrates, subagents and team members run on cheaper/specialized models to cut cost.

Per-agent routing is controlled via routing rules (model pattern match) or MODEL_ROUTE magic (system prompt override).

Quick Start

# Docker (recommended)
docker run -d --name claude-rt \
  -p 8787:8787 \
  -e OPENAI_API_KEY=sk-... \
  ghcr.io/soju06/claude-rt:latest

# Local
cp .env.example .env     # add your OPENAI_API_KEY
bun install && bun run dev

Then connect Claude Code:

ANTHROPIC_BASE_URL=http://127.0.0.1:8787 claude

To make this permanent, add to ~/.claude/settings.json:

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://127.0.0.1:8787"
  }
}

Examples

Local

Create config/runtime.local.json:

{
  "providers": [
    {
      "id": "anthropic-main",
      "type": "anthropic",
      "baseUrl": "https://api.anthropic.com",
      "mode": "passthrough"
    },
    {
      "id": "openai-main",
      "type": "openai",
      "baseUrl": "https://api.openai.com/v1",
      "apiKey": "sk-...",
      "mode": "transform"
    }
  ],
  "routing": {
    "version": 1,
    "defaultProvider": "anthropic-main",
    "routes": [
      { "id": "gpt-to-openai", "modelPattern": "^gpt-", "providers": ["openai-main"] }
    ]
  }
}

CONFIG_FILE=config/runtime.local.json bun run dev

Docker

Mount your config file into the container:

docker run -d --name claude-rt \
  -p 8787:8787 \
  -v $(pwd)/config/runtime.local.json:/app/config/runtime.local.json \
  -e CONFIG_FILE=config/runtime.local.json \
  ghcr.io/soju06/claude-rt:latest

Claude models → Anthropic (passthrough), gpt-* models → OpenAI (transform). Routes are regex, first-match-wins.

Subagent Routing

Claude Code subagents can target a specific provider/model by adding a magic line as the first line of their system prompt:

[MODEL_ROUTE provider=openai-main model=gpt-5.3-codex]

The proxy strips this line before forwarding. See .claude/agents/ for examples.

Configuration

Runtime config: config/runtime.example.json (copy to runtime.local.json for local overrides)

{
  "providers": [
    { "id": "anthropic-main", "type": "anthropic", "baseUrl": "https://api.anthropic.com" },
    { "id": "openai-main",    "type": "openai",    "baseUrl": "https://api.openai.com/v1", "authEnv": "OPENAI_API_KEY" }
  ],
  "routing": {
    "defaultProvider": "anthropic-main",
    "routes": [
      { "id": "gpt-to-openai", "modelPattern": "^gpt-", "providers": ["openai-main"] }
    ]
  }
}

| Variable | Description | |---|---| | PORT | Listen port (default 8787) | | CONFIG_FILE | Runtime config path | | OPENAI_API_KEY | OpenAI API key for gpt-* routing | | ROUTE_MODEL_ALLOWLIST | Comma-separated allowlist for MODEL_ROUTE targets |

Development

bun run dev              # dev server (port 8787)
bun run check            # typecheck + all tests
bun run test:e2e         # E2E tests (replay)
bun run test:e2e:live    # E2E against live OpenAI
docker compose watch     # Docker watch mode