# Space AI

A workspace with an agent around your repositories — it reads files, answers with links to lines, opens issues and proposes changes that you accept or decline.

## Starting a session

1. **New session** — tick the repositories you are working with (up to five). Only repositories
   you can see and where **AI** is enabled are listed.
2. Write a request and pick a model — the list of your keys and your organizations' keys, each
   with the model from its settings.

Space AI opens from the menu. **Ctrl+Enter** sends the request. The old `/chat` address leads
here too.

## The screen

* **Left — sessions.** **×** next to a session deletes it (this cannot be undone). Sessions of an
  external agent over MCP are marked **MCP**.
* **Centre — the feed.** Your request, the **agent's steps** as it works — «Collected context»,
  «Reading notes/a.md», «Searching …», «Proposed a change» — the answer and proposal cards. While
  the agent works, the feed updates by itself, without reloading and without losing your scroll
  position.
* **Right — the session.** Attached repositories (**×** removes one, **+ Add a repository** adds
  one), what was sent to the model last time and how much of its window is used, money and
  results: how much was spent, how many turns, **how many changes were accepted** and what one
  accepted change cost.
* **The model** is switched next to the input field — for this session.

On a phone the columns stack: the feed, the session, the list of sessions.

## What the agent can do

The agent is a model with tools, like the Claude, ChatGPT and Copilot agents. It decides itself
what to read:

* **read** — the list of files, a whole file or a part of it, text search, issues and pull
  requests;
* **open issues** — an idea, a hypothesis, a problem it found;
* **reply** in issues and pull requests;
* **change issues** — title, description, close or reopen, labels;
* **write notes** into the project context repository — see [MCP server](https://skillok.org/docs/en/account/mcp#where-the-project-context-lives);
* **propose a change** — new or modified files;
* **draw images** — a PNG proposed into the repository;
* **make drills and challenges** from a markdown note — see [From a note](https://skillok.org/docs/en/learning/from-a-note).

Issues, replies and notes are applied at once. **File changes only happen when you decide:** a
card with the files and **Accept** and **Decline** buttons appears in the feed. An accepted change
becomes an `llm/…` branch and a pull request — nothing is written to the default branch. When you
decline, you can say why: the session remembers it.

The agent reads the project rules — `AGENTS.md` at the repository root — and the context
repository if one is linked. What the agent may do is decided by the owner in **Repository
settings → AI → Agents**.

The agent acts only when asked and only where you have the right: issues where issues are enabled,
changes where you have write access. Everything it does is signed «Through Space AI».

## What it reads

Repositories you can see that have **AI** enabled (**Repository settings → AI**). Access is checked
before every turn and in every action. The text of the fragments goes to the provider of the
connected model.

**What the agent never sees.** List paths in a `.skillokignore` file at the repository root, one
per line, like `.gitignore`: `data/raw/`, `*.csv`, `!public.csv`. Secrets are closed even without
the file: `.env` (but not `.env.example`), keys (`*.pem`, `id_rsa`), `.netrc`, `.ssh/`. A closed
path does not show up in the agent's file list or search, cannot be read and cannot be changed by
it — the same rule in Space AI, over MCP and in Harness on your computer, where a closed path is
hidden from the agent's commands too. The agent reads `.skillokignore` but never changes it.

## Memory

The model is replaceable; memory stays in SKILLOK: project rules in `AGENTS.md`, results as notes
in the context repository, the session history in its log. Switch the model or the client and
nothing is lost.

## What it costs

A turn is several model calls: an initial search, then steps with tools (no more than eight). The
price of a turn and of a session is shown in dollars and is paid by the connection — see
[Connecting models](https://skillok.org/docs/en/account/ai-keys). Limits: 50 turns a day and one turn at a time per
session.

## Your own agent and automation

SKILLOK offers the same tools to external agents — Claude, ChatGPT, Cursor, VS Code — over the MCP
protocol, connected in one click: [MCP server](https://skillok.org/docs/en/account/mcp). Requests then go through your
subscription, and the agent's changes arrive in Space AI for your decision.

For scripts there is the `/api/v1/space/*` API with a [personal token](https://skillok.org/docs/en/account/access-tokens):
sessions, requests, the event log and decisions on proposals, and `POST /api/v1/space/repositories`
creates a repository.

SKILLOK models are also available in the OpenAI format — `/api/v1/llm` (`/models`,
`/chat/completions`, streamed answers): this is how the Space AI — Harness desktop app uses them. The
same connection pays, within the same limits as a Space AI turn; without a confirmed email AI is closed.

The agent takes challenges with you: `start_challenge` creates the attempt repository,
`attempt_status` shows the task of the open stage, `check_attempt` checks the pushed solution and
says what failed and why. In the Space AI — Harness desktop app it is one phrase: "commit and
check" — commit, push and the check run by themselves.

---

https://skillok.org/docs/en/repositories/chat