CA Server

Lightweight streaming proxy for Anthropic and OpenAI, focused on computer-use (CUA) loops via the Messages/Responses APIs.

Setup

1. Configure server env

   export ANTHROPIC_API_KEY=your-key
   export OPENAI_API_KEY=your-key
   # Redis (required)
   export REDIS_URL=redis://localhost:6379

2. Install & run

   npm install
   npm run build
   npm start

3. Provision keys (CLI)

npm run key:new -- alice
npm run key:list

Keys are stored in Redis under keys:rec:<pk> (and indexed in keys:all).

Test Manually

# Test Anthropic (Messages API)
curl -X POST http://localhost:3000/anthropic/v1/messages \
  -H "Authorization: Bearer pk-<user-key>" \
  -H "Content-Type: application/json" \
  -H "anthropic-version: 2023-06-01" \
  -d '{"model": "claude-3-haiku-20240307", "max_tokens": 10, "messages": [{"role": "user", "content": "Hi"}]}'

# Test OpenAI (Responses/Chat)
curl -X POST http://localhost:3000/openai/v1/chat/completions \
  -H "Authorization: Bearer pk-<user-key>" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-3.5-turbo", "max_tokens": 10, "messages": [{"role": "user", "content": "Hi"}]}'

Note: Streaming (SSE / chunked JSON) is forwarded transparently.

## Status

- Health: `GET /health` → `{ status: 'ok' }`
- Version: `GET /version` → `{ version, commit }`
- Status: `GET /status` → Probes Redis and upstream providers (OpenAI/Anthropic) using configured credentials and returns per-target `ok`, `status/latencyMs`, plus overall status.

## CUA Loop Test (Simple)

Minimal commands; ensure `.env` contains `OPENAI_API_KEY` for OpenAI official.

- Ticketek (find Ticket Delivery FAQs):
  - `npm run cua:ticketek`
- AccuWeather (find Monthly forecast for Manchester, GB):
  - `npm run cua:accuweather`

Notes:
- Uses Playwright Chromium (downloaded automatically).
- Switches to Node 20 with `nvm` if available.
- Always acknowledges tool safety checks in this harness to keep the loop simple.

## ACUA Loop (Anthropic)

Run the Anthropic computer-use loop locally using the Messages API tool-calls.

```bash
# Ensure ANTHROPIC_API_KEY is set (or use the proxy base URL and a pk-...)
export ANTHROPIC_API_KEY=sk-ant-...

# Optional: route via this server proxy
# export ANTHROPIC_BASE_URL=http://localhost:3000/anthropic

# Ticketek and AccuWeather tasks
npm run acua:ticketek
npm run acua:accuweather

# Generic
ACUA_TASK=tasks/accuweather.json npm run acua:run

Environment variables:

• ANTHROPIC_API_KEY: API key for Anthropic (required unless using proxy with pk-...).
• ANTHROPIC_BASE_URL (optional): Set to http://localhost:3000/anthropic to route via this server.
• ANTHROPIC_MODEL (optional): Defaults to claude-sonnet-4-20250514.
• ACUA_THINKING_BUDGET (optional): Enable thinking for Claude 4/3.7 (e.g. 1024).

Proxy One-Liners

Run against OpenAI (uses .env OPENAI_API_KEY):

• Ticketek: npm run cua:ticketek
• AccuWeather: npm run cua:accuweather

Notes:

• All scripts auto-install Chromium and use Node 20 via nvm if available.

1:1 Drop-in (Swap Base URL)

• OpenAI SDKs: set baseURL to http://<your-proxy>/openai and use your issued pk-... as the API key. Example (Node):

  import OpenAI from 'openai'
  const client = new OpenAI({ baseURL: 'http://localhost:3000/openai', apiKey: 'pk-...' })
  const res = await client.chat.completions.create({ model: 'gpt-4o-mini', messages: [{ role:'user', content:'hi'}] })

• Anthropic SDKs: set baseURL to http://<your-proxy>/anthropic and use your issued pk-... as the API key. You may omit anthropic-version; a default is set if missing.

Usage Accounting

• Per-key counters are updated after each proxied request (total/success/error/lastUsedAt) in Redis (keys:rec:<pk>).
• Additional Redis usage metrics (hashed by key id) are in usage:key:<sha16>.