A Nightly Build in Berlin with a Proxy-Aware AI CLI

The Berlin build that needed one useful answer

At 8:40 on a Tuesday evening, a developer in a Berlin office is trying to understand why a nightly build fails only on one machine. The logs are long, the deadline is close, and opening a browser to paste in fragments would add friction and risk exposing more context than necessary. A small command-line tool can help: it reads a prompt from the terminal, sends it to a model through an API, and prints the response where the developer is already working.

The useful detail is that the tool is proxy-aware. It can operate in environments where outbound network traffic must pass through a company proxy, while also supporting a model API endpoint that is different from the provider’s default. The idea sounds specialized, but it is mostly a few familiar parts connected carefully.

What the OpenAI SDK contributes

The OpenAI SDK is a client library: it handles the mechanics of making API requests, including formatting messages and parsing responses. A command-line program built on the SDK can focus on its own small job, such as accepting text, selecting a model, and displaying the answer. It does not need to implement the HTTP protocol itself.

“Built on the OpenAI SDK” does not necessarily mean the tool can only talk to OpenAI. OpenAI-compatible services expose a similar request shape, so the SDK can often be configured with a different API base URL and a credential for that service. In practice, the tool needs to make those settings configurable rather than burying a provider URL in its source code.

What proxy-aware means in everyday use

A proxy is an intermediary between an application and the wider internet. Organizations use one to manage outbound access, and individual developers may use one for a particular network or security setup. Instead of connecting directly to an API, the command-line tool sends its request through the proxy.

Many HTTP clients, including the client commonly used beneath the OpenAI Python SDK, can read proxy configuration from environment variables such as HTTP_PROXY and HTTPS_PROXY. That lets the same tool work on a home network and on a managed network without changing its request logic. The exact behavior depends on the SDK version and its HTTP client configuration, so a practical test is to run a small request with the intended proxy settings and check that the response arrives.

Proxy configuration and API configuration solve different problems. The proxy controls the network route. The API base URL tells the SDK which service should receive the request. If either is wrong, a request may fail before a model can answer.

A small tool with a clear boundary

For a first version, keep the command-line interface narrow: accept a prompt, read the API key and base URL from environment variables, choose a model, and print the returned text. Avoid logging credentials or full prompts by default. Add a timeout and a readable error for common failures, such as a missing key, a refused proxy connection, or an API error.

For example, a developer might point the SDK at RelayVano’s API base, https://api.relayvano.com, and provide a RelayVano key through the tool’s normal configuration. The same request path can then reach an available model, such as DeepSeek V4 Flash, without the command-line program needing a separate integration for every model. Keep provider-specific options explicit, because compatible APIs can still differ in supported features.

Testing the route before trusting the answer

Test the tool in layers. First, confirm that it starts and reports missing configuration clearly. Next, send a short request without a proxy on a network where direct access is allowed. Then repeat with the proxy environment variables set. Finally, test an invalid key and an unreachable proxy so failures are useful rather than mysterious.

Also decide what belongs on the command line and what belongs in environment variables. Prompts are convenient as arguments or standard input; secrets should not be typed into shell history. If the tool will handle source code or incident logs, review the service’s data handling terms and avoid sending material that the team is not permitted to share.

Taking the pattern into a real workflow

This design is useful for one-off questions, build summaries, and small automation steps where a full application would be unnecessary. It is not a substitute for retries, access controls, or careful data handling when the tool becomes part of a production workflow.

RelayVano is a pay-as-you-go relay for AI models used by developers in many countries, and it works with OpenAI-compatible SDKs as well as tools such as Claude Code, Codex, and Gemini CLI. For a compact tool that needs one configurable endpoint across model options, RelayVano is a practical place to start; sign up to get an API key and try the request from your terminal.