cli4api

Wrap Google Antigravity CLI (agy) as an OpenAI-compatible HTTP API.

Use any OpenAI SDK or tool (curl, OpenClaw, Cursor, etc.) to send prompts to agy's Gemini models — no API key needed, just agy's OAuth login.

Why

agy uses OAuth (not API keys), so you can't point standard OpenAI-compatible clients at it. cli4api bridges the gap:

[OpenAI client] → POST /v1/chat/completions → agy -p → Gemini

Requirements

• Bun >= 1.2
• Antigravity CLI (agy) installed and authenticated (agy once interactively)

Install

npm install -g cli4api

Or run directly:

npx cli4api

Usage

# Start server (default port 8080)
cli4api

# Custom port and host
PORT=9090 HOST=127.0.0.1 cli4api

Then send requests:

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Gemini 3.5 Flash (Medium)",
    "messages": [{"role": "user", "content": "Say hello"}]
  }'

Streaming

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Gemini 3.5 Flash (Medium)",
    "messages": [{"role": "user", "content": "Say hello"}],
    "stream": true
  }'

List models

curl http://localhost:8080/v1/models

OpenAI Compatibility

| Endpoint | Method | Status | |----------|--------|--------| | /v1/chat/completions | POST | ✅ Streaming + non-streaming | | /v1/models | GET | ✅ |

API

POST /v1/chat/completions

{
  "model": "Gemini 3.5 Flash (Medium)",
  "messages": [
    { "role": "system", "content": "You are a helpful assistant." },
    { "role": "user", "content": "What is the capital of France?" }
  ],
  "stream": false
}

• model — agy model name (default: Gemini 3.5 Flash (Medium))
• messages — array of { role, content } objects (roles: system, user, assistant)
• stream — boolean, SSE streaming (default: false)

Development

git clone <repo> && cd cli4api
bun install
bun run dev      # start server with hot reload
bun test         # run tests
bun run build    # bundle to dist/
bun run typecheck

How it works

cli4api spawns agy -p as a subprocess for each request, captures stdout, and wraps the response in OpenAI chat completion format. Session continuation uses agy's --conversation flag. Streaming is simulated by splitting the response into word-sized SSE chunks.

License

MIT