The API Trap: How I Wasted Hours on a Terminal AI Tool

It all started with a simple automation idea

I was writing a Python script on my Ubuntu environment to scrape Upwork job postings and save them to an AWS S3 bucket. Then, a thought crossed my mind: wouldn’t it be convenient to automatically summarize these English job descriptions into Japanese using AI?

That simple thought was the beginning of a long, exhausting day.

I decided to introduce a CLI tool called OpenClaw, which allows you to interact with Gemini directly from the terminal. The idea of seamlessly integrating AI without leaving my black screen seemed incredibly smart.

The First Wall: The Endless Rate Limit Hell

I installed the tool, configured the API key I got from Google AI Studio, and started it up. However, the screen spat out a heartless error message:

“API rate limit reached. Please try again later.”

I waited minutes, then ten minutes, and even left it overnight. The error never disappeared. The causes were completely ridiculous.

First, a conflict of multiple keys. I had mistakenly passed both an old key and a new key to the environment variables, and the system faithfully kept trying to use the old, rate-limited key.

Second, a runaway invisible Discord bot. OpenClaw kindly included a Discord integration feature by default. While I was struggling with the errors, the system had automatically launched a Discord bot in the background, instantly consuming the strict free-tier API quota for the gemini-2.0-flash model. It was a background process I never asked for, preventing the main task from ever starting.

The Second Wall: Manipulated by Plausible Lies

To solve this, I asked the AI assistant for help. This only made the wound deeper.

The AI confidently provided commands to fix the issue, but they included options that did not exist. Furthermore, it suggested switching to a model with looser rate limits, gemini-1.5-flash, but that model was already unsupported in the API version, leading me straight into a “404 NOT FOUND” wall.

Finally, by clearing the cache of past failures, directly rewriting the configuration file, and specifying gemini-2.5-flash-lite, I was able to escape the maze of errors.

Joy, Followed by Emptiness

“It finally connected!”

When I saw the prompt silently waiting on the terminal without spitting out errors, I felt a solid sense of achievement. The long hours of struggle were rewarded.

But moments later, a cruel reality dawned on me.

My original goal was simply to summarize scraped data inside a Python script. In other words, I did not need to install this complex tool at all. I could have achieved my goal perfectly by just adding a few lines of code to call the Gemini API directly within my Python script.

The CLI tool was merely a wrapper that made calling an API look good on the terminal while maintaining chat context. For someone who can write their own code to send data to S3, it was a completely unnecessary detour.

Conclusion: Are Tools Becoming the Goal?

Sometimes, when we find a tool that looks convenient, we forget our original purpose and become obsessed with just getting the tool to work.

The lesson I learned after wasting several hours is simple: Always start with the most direct method.

If you are trying to integrate AI into your program, you should start by directly calling the simple, official API. Do this before you end up fighting an invisible Discord bot and getting manipulated by non-existent model names.