Connect an AI model to BenchQuorum

BenchQuorum needs a model to plan its steps and write files. You choose where that model runs: a cloud AI provider through a standard API endpoint, a model on your own PC, or the built-in model. The choice decides which text leaves your machine.

Three ways to connect a model

A cloud AI provider

A cloud AI provider suits hard multi-step work, and you connect it through a standard API endpoint with your own key. Your key stays local. Your prompts go to that provider, so do not paste anything from a machine you may not share, and note that the provider may charge separately for its usage.

A model on your own PC

A model running locally on your PC suits private work and machines without internet access. With a local model nothing leaves the machine, because prompts never reach another endpoint. The trade-off is hardware: a small model may need more of your review, and a large one needs a PC that can run it.

The built-in model

The built-in model needs no endpoint setup, which makes it a place to begin when you have no provider yet. Begin small. You still review every diff it proposes and every command before you accept it, exactly as with a cloud or local model, because model quality varies and nothing is correct by default.

Press Test before you start

The Test button checks the endpoint you entered. It shows the response time and the context window, which is how much text the model can hold at once. A slow response or a small window tells you early that a long script and a long results file may not fit.

Match the model to the step

Use a bigger model for refactors and logic that spans several files, and a smaller one for small edits such as fixing one line in bench.py or adding a column to results.csv. Size should follow the task. When a single step is hard, run_on_model sends just that step to a bigger model, so the rest of the chat stays on the smaller one.

Quality varies between models, so read every diff. Ask first pauses before each change, which gives you a moment to check a command or a file edit against the numbers you measured. Never accept a result you did not look at.