Building with an agent, against your own model

Claude Code reads your files and runs commands, so it wants the situation rather than a specification. Open a terminal in an empty folder and tell it what you are trying to make and what it is allowed to assume:

  I want to build <the thing>. It must talk only to a local
  OpenAI-compatible server at http://127.0.0.1:8081/v1 — no
  cloud API, no API key that matters, nothing leaves this
  machine. Use the openai package pointed at that base URL.
  Start by writing the smallest version that proves the
  connection works, and let me run it before you build on it.

That last sentence is the one that matters. An agent given a whole project will happily write all of it before you have confirmed the first request even reaches your model, and then you are debugging four things at once.

Codex asks fewer questions back, so give it more of the shape up front — the file layout, the endpoint, and the constraint — in one instruction:

  Create a Python project with this layout:
    main.py, requirements.txt, README.md
  It talks to an OpenAI-compatible server at
  http://127.0.0.1:8081/v1 using the openai package with
  api_key="not-needed" and base_url set to that address.
  No cloud calls anywhere. main.py should <do the thing>.
  Handle a connection error by telling the user their local
  server is not running, rather than raising a traceback.

The last line is worth keeping in every instruction you write. A refused connection is the single most common thing that goes wrong on this surface, and an agent will not think to handle it unless you say so.

Read what it gave you

Generated code is a draft. These are the things that go wrong on this particular project, in roughly the order they are worth checking.

• Is the base URL actually local? Search the generated code for "api.openai.com" and for "anthropic". A model trained on thousands of cloud examples will sometimes reach for one out of habit, and the code will look perfectly correct while quietly not being the thing you asked for.
• Is there a real API key in the file? There should not be. Local servers do not check it, so "not-needed" is the honest value. If the agent invented one, it is a habit, not a requirement.
• Does it stream, and did you want it to? Streaming feels much faster on a local model because the first token arrives long before the last. An agent will not add it unless asked.
• What happens when the server is off? Run the code with Pip stopped. A traceback is a bug you have not fixed yet.
• Is the context window respected? Your local model has a much shorter one than the agent that wrote the code. A script that pastes an entire document into one request works on a cloud model and fails on yours.
• Can you explain every line? If you cannot, you have a dependency you do not understand in something you are about to rely on. Ask the agent to explain it, then decide whether to keep it.