Integrating DeepSeek API into a Minimal Q&A Script

The most practical starter task for the DeepSeek API Open Platform is to get a "program sends one sentence, model replies with one sentence" loop working. On day one, grab your key; by day three, hook it into a terminal script; after a week, you'll understand what API calls are actually doing.

Day One: Get a Key That Opens the Door

Beginners should first grasp two terms. An API is an interface—think of it as a door where programs say hello to each other. A key (API Key) is a string that acts like a door card. Without the card, no matter how big the room behind the door is, you can't get in.

After logging into the DeepSeek platform, find API Keys or "Create API Key" in the left-hand menu. Click it, enter test-local in the notes field, and confirm. The page will give you a long string of characters, usually warning you that it's only displayed once.

My first mistake was screenshotting the key and sending it to a group chat asking if it worked. That's basically taking a photo of your door card and broadcasting it to everyone. The correct approach is to copy it locally immediately—don't paste it into chat boxes, don't paste it into public repositories.

I always ask myself: Is this data source reliable? For beginners, key management is more important than model capability.

If the platform requires account verification, follow the on-page prompts. Don't just top up randomly to make things work. The only goal for day one is to get a key and ensure it exists only on your local machine.

Day Three: Put the Door Card into a Terminal Script

On day three, I tried writing a minimal script. Assuming you've never touched code before, let me break down the jargon.

1. Create a new file named ask.py.

2. Copy the example from the DeepSeek docs into it; don't change the model name yet.

3. Replace the base_url in the example with the DeepSeek address provided in the platform documentation. Don't type it manually; copying is safest.

4. Put the key in an environment variable, e.g., DEEPSEEK_API_KEY. Do not hardcode it directly in the code.

5. Type python ask.py in the terminal. If it says python isn't found, install Python first.

6. When the cursor waits for input, type "Hello, please introduce yourself in one sentence."

7. Expected result: The terminal returns a response—not an error, and not blank.

The most common error involves environment variables. Many people write the key directly into the code, and it leaks as soon as they upload to GitHub. A safer method is letting the program read from the system environment. For instance, Python examples often use os.environ["DEEPSEEK_API_KEY"], which means "look for the value associated with that name in the computer's environment."

The second pitfall is compatibility format. DeepSeek API docs mention it uses an API format compatible with OpenAI/Anthropic. This sounds friendly, but beginners often misunderstand it as meaning any SDK works out of the box. In reality, you still need to change the base_url and the key. I initially only changed the key and forgot the address, so requests went to another service, resulting in confusing errors.

After a Week: You'll See the Ledger Behind "Calls"

Looking back at this task after a week, the feeling is different. The first time you got it running, you thought the model was magical. After asking a dozen questions in a row, you start caring about something else: Who processed this sentence? What resources were consumed? Where are the logs?

It's like conducting an investigation. On the surface is an answer; behind it is a chain. Keys, model names, addresses, request content, return results—every step leaves a trace. For developers, this is functionality; for regular users, this defines responsibility boundaries.

You can do three small things while you're at it: Separate keys by purpose—one for local testing, one for online programs; Save every question and its returned result to a local file for easy review; For uncertain capabilities, don't just look at the marketing—run the same question through different models and verify from multiple sources.

What tutorials ultimately aim to solve is teaching programs to ask questions safely.

Don't rush into building Agents for the next step. First, create a "Question List" script. Prepare ten questions you genuinely care about, have the program call them one by one, and record the answers, time taken, and failure reasons. If you can run these ten stably, you'll avoid the pitfalls that trip up most beginners.