OpenClaw OpenAI Proxy

A small OpenAI-compatible gateway that routes requests through OpenClaw agents.

What it does

• Exposes OpenAI-style endpoints:
  • GET /v1/models
  • POST /v1/chat/completions
  • POST /v1/completions
• Routes public models to different OpenClaw agents
• Adds debug headers for actual upstream/provider visibility
• Records per-request usage to local JSONL
• Works well as a shared gateway for SillyTavern and other custom apps

Supported public models

By default the example routing exposes:

• openclaw/st
• openclaw/main
• openclaw/writer
• openclaw/coder
• openclaw/research

Config files

Tracked in git:

• config.example.json
• routes.example.json
• .env.example
• claw-proxy.service.example

Local-only files (ignored by git):

• config.json
• routes.json
• .env
• data/usage.jsonl

Quick start

1. Prepare local config

cp config.example.json config.json
cp routes.example.json routes.json
cp .env.example .env

Edit them for the current machine.

2. Install service

chmod +x deploy-install.sh
sudo ./deploy-install.sh /opt/claw-proxy

If you keep the project somewhere else, pass that directory instead.

Example config.json

{
  "host": "0.0.0.0",
  "port": 8780,
  "apiKey": "change-me",
  "corsOrigin": "*",
  "agent": "st",
  "publicModel": "openclaw/main",
  "upstreamModel": "bailian/kimi-k2.5",
  "thinking": "",
  "defaultMode": "light",
  "streamChunkSize": 24,
  "timeoutSeconds": 600,
  "includeOpenClawMeta": false
}

Example routes.json

{
  "defaults": {
    "agent": "main",
    "mode": "light"
  },
  "clients": {
    "default": {
      "apiKey": "change-me",
      "defaultModel": "openclaw/st",
      "allowedModels": [
        "openclaw/st",
        "openclaw/main",
        "openclaw/writer",
        "openclaw/coder",
        "openclaw/research"
      ]
    }
  },
  "models": {
    "openclaw/st": {
      "agent": "st",
      "mode": "light",
      "configuredUpstreamModel": "bailian/kimi-k2.5"
    },
    "openclaw/main": {
      "agent": "main",
      "mode": "light",
      "configuredUpstreamModel": "998code/gpt-5.4"
    }
  }
}

Debug headers

Each generation response may include:

• X-Claw-Proxy-Mode
• X-Claw-Proxy-Agent
• X-Claw-Proxy-Session
• X-Claw-Proxy-Configured-Upstream-Model
• X-Claw-Proxy-Actual-Upstream-Model
• X-Claw-Proxy-Actual-Provider
• X-Claw-Proxy-Prompt-Tokens
• X-Claw-Proxy-Completion-Tokens
• X-Claw-Proxy-Total-Tokens

Usage endpoint

GET /debug/usage?limit=50

Notes:

• Only requests made after usage logging was enabled are available
• Existing historical requests are not backfilled

Current limitation

Streaming is still pseudo-streaming for now: the proxy waits for the full result and then replays SSE-style chunks.