Skip to main content

What is Agivar

Agivar is a desktop AI assistant that runs on your computer. It has two abilities that feed each other:

  • It gets things done for you — it understands what you want, plans the steps itself, and then actually drives your desktop GUI, runs commands, and works with files to finish the task.
  • You can teach it — show it how to operate a particular app or website, your business rules, or your personal preferences. In particular, screen-record yourself doing something once and it absorbs that know-how and reuses it later when running related tasks.

In short: the more you use it and teach it, the better it understands your computer and how you like things done.

Two working modes​

Task ModeTeach Mode
What you're doingAsking it to execute a taskTeaching it knowledge, a workflow, or a preference
What it doesPlan → drive the GUI / run commands → verify → reportUnderstand your explanation / demo → distill it into a reusable entry → write it to the memory base
Typical input"Compress all the images in this folder"
"Give 2 coins to this creator on Bilibili"		"Let me show you how to submit an invoice in our expense system"
(with a screen recording) "follow this workflow"
OutputTask done + summaryA memory entry (filed under platform/topic)

Both modes live in the same chat window — pick the mode when you start a new conversation. Once a conversation has any messages, its mode is locked and can't be changed.

The "teach → remember → execute" loop​

   Teach Mode                       Memory base                    Task Mode
┌────────────────┐  distilled into  ┌──────────────────────┐  read before  ┌──────────────┐
│ you explain /  │ ───────────────▶ │ platform/topic +      │ ────────────▶ │ follows the  │
│ screen-record  │                  │ operation description │   executing   │ workflow,    │
│ a demo         │                  │ + (optional) recording│               │ fewer detours│
└────────────────┘                  └──────────────────────┘               └──────────────┘
  • A screen recording is processed automatically: keyframes are extracted, a short video summary is generated, and a structured "operation description" is produced (Overview / Initial state / Step-by-step / "stage result" after each step).
  • What it learns goes into the memory base, organized as platform/topic (e.g. bilibili/coins, general/expense-rules).
  • Before running a GUI task, Task Mode consults the memory base; if an entry references a recording, it can even ask targeted questions about details inside that recording.

Who it's for​

  • People who repeatedly do the same operations on certain sites/apps and want the AI to take over once it has learned them.
  • Teams with internal workflows that "only you know" (expenses, approvals, listing products…) that you'd like to turn into reusable know-how.
  • Anyone who wants the AI to follow their preferences and rules when working on their machine, instead of re-explaining every time.

Next​