AI2P is what Trello or Jira usually are: projects, tasks, executors, a board, a history. There is one difference, and it changes everything: the executor of a task may be an AI. Whatever the AI can do itself it does itself, at the right moment and without a reminder; the rest is left to people and waits for them, as in any planner.

The system works on your computer (Windows 10–11, Linux, macOS), the interface is in the browser. The data goes nowhere: the database, the project files and the keys lie next to the program.   

What it does

• It leads the work, not just a to-do list. A task has an executor, a due date, a priority, acceptance criteria, blocking tasks and subtasks. Finishing a task starts the next ones by itself — the ones that were waiting for it.

• It connects AI of any kind. Text models over an API (Claude, GPT, Gemini, DeepSeek, Qwen, GigaChat, YandexGPT and others), Claude Code by subscription, local models on your video card, and also pictures, video, sound and 3D. All of them are ordinary executors: each has a nick, skills, a price and limits.

• It picks an executor itself. By the skills of the task and the "cost ↔ quality" setting of the project the system chooses whom to give it to — the AI first and then a person, or the other way round. An executor that is busy or has used up its limit is skipped, and the task waits for one to free up.

• It splits big work into parts. An agent may create subtasks and start them by priority; the last task of the group sums it up and, if something did not add up, returns its neighbours for fixing.

• It remembers experience. The lessons gained in the work are written down to the project, to a template node or to the whole organization and are put into the next job by themselves — their own for every skill.

• It repeats typical processes. A template is a tree of tasks with descriptions, skills and an order; a new working process is deployed out of it by a single action.

• It works on a schedule — from "every Monday" to the system's own actions, automatic archiving for example.

• It lives on several computers. The servers are joined into a cluster and replicate the data to each other; every task has an owner server on which it is carried out — that is how a local model computes where the video card stands.

• It takes tasks from external systems — Trello, GitHub, GitLab.

• It keeps the AI within bounds. The security rules decide what the agent is allowed to do, what needs a person's confirmation and what is forbidden; the agent does not go outside the project folder.

• It speaks your language — the interface, the messages and the agents' prompts are taken from dictionaries;

• ​

What it looks like in work

1. You create a project and name its folder — that is where the agent will read and write files.

2. You write a task the same way you would write it for a person: the title, the description, the acceptance criteria.

3. You press "start". The task goes to an AI executor, its work is visible in the history, and it asks its questions in the task chat — and that is where you answer it.

4. The finished result lands as the task's artifact, and the task itself moves to review.

 

What is needed

 

* Windows 10/11, Linux or macOS; administrator rights only for the installation.

* The **ASP.NET Core 8.x** runtime — either installed separately or taken from the full installation

  package (`AI2P_v_1_NN_full_…`), which already holds it inside.

* An API key of the AI you are going to use — or a Claude Code subscription, or a video card for

  local models. Without a model the system works as an ordinary planner for people.

​

Quick start

​

For the first run in a few minutes see  https://github.com/mmf007/AI2P/blob/main/doc/en/man/quickstart.md

The distributions live here   https://github.com/mmf007/AI2P/releases
Main page on github   https://github.com/mmf007/AI2P

​​​

​

​

​