This proxy is intended for local developer use with ZCode and NVIDIA NIM OpenAI-compatible chat completions. Keep the default loopback binding unless you have separately configured TLS termination, firewall rules, and client access controls.

• Do not commit API keys.
• Do not commit .env; it is intentionally ignored by Git.
• In default env mode, set NVIDIA_API_KEY in the local environment.
• In client mode, put each NVIDIA API key in the corresponding ZCode provider API key field.
• In pool mode, keep real NVIDIA keys only in .env and put only NIM_PROXY_CLIENT_KEY in ZCode.
• Generate a random local proxy key, keep it different from every NVIDIA key, and rotate it if exposed.
• Restrict .env filesystem permissions to the Windows account that runs the proxy.
• The .env loader accepts only the documented credential names and never executes file content.
• Upstream URLs containing credentials, query parameters, or fragments are rejected before the proxy binds.
• The proxy never logs API key values, local proxy credentials, or full message content in any mode.
• /health exposes aggregate queue/key-state counts only. It does not expose raw keys or fingerprints.

env mode cannot bind to a non-loopback address. client and pool modes still require explicit --allow-remote; that flag does not add TLS or firewall protection.

• 408 and 429 cool down only the affected key.
• 401 and 403 quarantine only the affected key until restart.
• A quarantined key should be removed or replaced in .env before restarting.
• Streaming requests are never replayed after response streaming begins.

When reporting issues, include:

• proxy version
• Python version
• sanitized stripped key names
• HTTP status code
• queue mode and aggregate /health counts

Do not include .env, API keys, local proxy credentials, key fingerprints, or full prompt/message content.