Choose the AI model behind your shell edits

SeedSpindle on Windows needs an AI model to plan and write each change. You can connect a cloud AI provider, a model running on your PC, or the built-in model. The choice decides where your prompts go and how large a task the agent can handle.

Choose your model source

Cloud AI provider

A cloud AI provider connects through a standard API endpoint that you enter in the settings. The agent sends your prompt to that provider, including the text of any file it reads, such as your ~/.bashrc. Plan your use around that. Privacy therefore depends on the provider you pick, and a shell file with secrets in it should not be sent to a service you do not trust. A cloud model suits large tasks that a small local model cannot handle, and you can change the endpoint whenever you like. API keys stay on your PC, and prompts go only to the endpoint you picked.

Local model on your PC

A model running locally on your PC answers on the same machine where SeedSpindle runs. With a local model nothing leaves the machine, because the prompt and every file the agent reads stay on your PC. That makes it a good fit for shell files that hold paths, host names or other details you prefer to keep private. Hardware matters. A small model may be slower or less capable on a long task, so you review each diff with extra care, and you can send a single hard step to a bigger model if you decide to.

Built-in model

The built-in model is ready to use without your own endpoint, which makes it the quickest way to start. You install SeedSpindle for Windows, open a chat, and give the agent a first task such as adding an ll alias. No API settings are needed. Try it first. If a task turns out too hard, you can connect a cloud AI provider or a local model later and keep working in the same chat window, with the same approval modes and the same diff review.

Test your connection

The Test button checks the endpoint you entered and shows the response time and the context window of the model. Run it right after you add a new provider or a new local address, because a wrong key or a wrong address shows up there before you start real work. The context window tells you how much text the model can take in one request, which matters when ~/.bashrc and the chat history are both long. Check once. If the response time looks slow, try a smaller model for small edits and keep the larger one for harder steps. The test changes no files.

Pick the right model for your task

Use a bigger model for refactors and complex logic, and a smaller one for small edits such as adding one alias. A bigger model costs more to run or needs stronger hardware, so it makes sense to save it for the steps that need it. The run_on_model tool sends one hard step to a bigger model while the rest of the chat continues with your usual choice. Think about size. A single alias line needs very little reasoning, but a tidy-up of a long file with several functions and many aliases benefits from more capable help, and you can compare the diffs each model proposes.

Model quality varies between providers and local setups, so read the diff before you accept anything. Ask first pauses before actions that change something, which gives you a chance to reject a line that looks wrong. Plan only shows the plan without touching a file. Trust the diff, not the model. A model can write a plausible alias that does not do what you meant, and only your review and a test in a new shell tell you whether it does.