Pick the AI model that reads and edits your scripts
Pageferry works with a cloud AI provider through an API endpoint, with a model that runs on your PC, or with the built-in model. The choice decides where your prompts go, so read the three options before you connect anything.
Three ways to connect a model
Cloud AI provider
A cloud AI provider is reached through a standard API endpoint, and you choose the endpoint and enter its key in the app. The key stays on your PC. Each prompt, including the text of any file the agent reads from your folder, goes to that provider, so use this option only for scripts you are allowed to send there. The Test button checks the endpoint and shows the response time and the context window. A larger context window helps when a script and its output are long. The provider may bill its usage separately, and Pageferry itself needs no account.
Model on your PC
A model that runs locally on your PC is the option for scripts that must not leave the machine: with a local model nothing leaves the PC, and your chats, working folders and API keys stay there as well. You point the app at the local endpoint, press Test, and read the response time and the context window. Your hardware sets the limit, so a long script may not fit into a small window. Press Test before the first real task to see how your machine copes.
Built-in model
The built-in model is ready to use without your own endpoint, so you can open a chat and try the agent on a small Bash script straight after installing. You still choose an approval mode: Auto-run acts at once, Ask first pauses before actions that change something, and Plan only writes a plan and runs nothing. For harder steps you can add a cloud or local model later and send those steps to it with run_on_model.
Check the connection with Test
The Test button checks an endpoint and shows two things: the response time and the context window. Press it after you enter an endpoint and before you give the agent a real task. A long response time tells you the endpoint is slow right now, and a small context window tells you that a long script and its output may not fit into one request. If the check fails, correct the endpoint address or the key and press Test again. Do this once for every endpoint you add, because a cloud model and a local model can differ a lot in both numbers.
Choose a model size for your task
Use a bigger model for refactors and for scripts with tricky logic, and a smaller one for small edits such as adding a -v flag or quoting a variable. The context window matters more as the script and its output grow, so check it with the Test button. When one step is hard, run_on_model sends only that step to a bigger model while the rest of the chat stays on your usual one. A reasonable habit is to start small and move up only when the answer is not good enough. Cloud models receive the prompt, so keep that in mind when a script contains private paths or names.
Quality varies between models and between versions of the same model, so never accept a change without reading it. Under Ask first the agent pauses before actions that change something, and for an edit you see the diff and can accept or reject it. After that, Shell can run the script and show the real exit code, which is a better test than the model's own explanation. Plan only is useful when you want to read a plan before anything is run.
