First, what’s an AI harness?
Most people explain it with the analogy that it’s like the body you connect to your LLM, the brain. But I’d argue it’s more like the nervous system linked with some organs 😅 because of how much it impacts how the LLM actually thinks and behaves.
But however you want to see it, building one can get expensive really fast, especially if you are testing with strong models.
So if you are a student, a founder, or going the indie hacker road, that cost can quickly stack up during the building/testing phase.
I’ve been building one myself for the past 3 months, and it was costing me more than I could afford, so I came up with this little hack that I wanted to share.
I built Local Bridge.
You run it locally and point your API calls to it. Then every call spawns a local Claude or Codex CLI session, telling it to act like an LLM endpoint, read the request file, and output a response JSON file.
Every call fires a new session with no past memory, so it can mimic an actual API call where the model doesn’t know anything about the previous call except what’s included in your request.
That way, you can use your Claude or Codex subscription, which is way cheaper than using the actual API for early-stage testing.
Of course, Claude and Codex have some prebuilt system prompts, so I would trust this more for early-phase testing and debugging, not as a 100% replacement for real API evaluation.
Open source: https://github.com/alisel5/local-bridge
abdelali Selouani
Posted on
AI harness engineering is an interesting field, but testing can get costly really fast!
For further actions, you may consider blocking this person and/or reporting abuse
Top comments (0)