OfflinAI Agent Lite
State a bigger task and get a plan you can change. The agent works one task at a time, shows every step, and asks before anything leaves the app.
Coming soon to the Microsoft Store
About the app
Agent Lite is for tasks bigger than one question, such as write a short report comparing these three options, or make a checklist for moving house. You state a goal in plain words. The agent first shows a short plan of one to eight tasks, each with a done-when check that you could make by reading the result. You can change titles and checks, reorder, add and remove tasks, choose the tools and budgets, and give it text files to work from, before anything runs.
The agent then works one task at a time and shows each step and tool call as it happens, with how long it took. After each try it checks the result against the task's done-when. A failed check is told to the model on the next try. A task that fails twice is set aside and reported, so the run never loops.
Budgets of steps and time are always visible as bars. When one is used up, the run stops cleanly with a result, and you can add more steps or minutes to carry on. Stopping loses nothing. The run is saved after every task, and Run again continues from the first task not done.
Every run ends with a result: the answer pulled together, the files produced, the tasks done, and each task not done with the reason. A small model can make plans that look right and do little, so the app says so plainly and points you to the not-done list.
The tools are limited to what you allow: a calculator, today's date, listing and reading the run's files, and writing a file inside the run. Saving a file to your own folder always asks you first and never replaces an existing file. Running programs and using the internet are not offered. The step log is sealed on your device and can be read later or exported to Markdown.
Agent Lite uses OfflinAI Server running on your own PC and needs a chat model that can call tools. Without it, the plan is your goal as one task, and running it stops with a note on how to get a model.
Everything stays on your device. Nothing is sent to OfflinAI, there is no account, and no data is collected.
Features
- State a goal and review a plan of up to eight tasks, each with a done-when check, before anything runs
- Works one task at a time, shows every step and tool call, and checks each task
- A task that fails twice is set aside and reported, so the run never loops
- Step and time budgets always visible; stopping loses nothing and you can run again
- Every run ends with what was produced, what was done, and what could not be done and why
- Only the tools you allow; saving a file outside the run always asks you first
- Step log sealed on your device and exportable to Markdown