Glossary¶

KOLONİ's own words, as they appear on screen.

Agent

One worker in a turn: a role running on a model. The canvas draws each agent as a small box: its name and model, its state, and what it did. Its tools, commands and files are in the Inspector.

Agents may use

A switch for each model on the Models page. When it is on, agents may pick that model by themselves, not only you.

Always run

A button on a command's card. It keeps a rule for that chat, so the same kind of command runs without a card next time. See Permission levels.

Assistant PO

After a turn that did something, one small model request with no tools that suggests at most three follow-up tasks for the To-do list, each with the words it came from. On the canvas it is a small box under the turn. See Follow-up tasks.

Board

A page with every chat of the project in three columns: at work, waiting for you, and done today.

Branch (a team's)

A git branch koloni/<name> that the Orchestrator can give a team's work. Stopped, the agent's work stays committed there; told to go on, it goes on on the same branch. The card offers Merge or Open pull request. See Branches.

Candidates

A recipe step given to several agents at once, each in its own working copy. Each one's work is scored from 0 to 100 by the project's Measure command, and the best one is kept.

Card

A box in the chat that waits for your answer, for example "Run this command?" with Run once, Always run … and Reject. There are cards for file changes, commands, web searches, secret files, opening a sub-agent, plans and more.

Compare

Compare… under More next to the message box. One message goes to two models side by side, each in its own working copy. Nothing reaches your project until you keep one.

Coordinator

A team's own lead, shown as "Team coordinator". It runs its team's stages and reports to the Orchestrator.

Core

The facts in a project's memory: what you approved or wrote yourself. See Memory.

Decided for you

Where you see the questions the Orchestrator answered while you were away, if you let it (Settings › This project › Asking and away › When you're away).

Discussion

Several agents talking one question through in rounds, with nobody changing anything meanwhile. An agent starts one; it has a token budget.

Fact

A short piece of memory that you approved or wrote, usually pointing to lines of a file. It is given to an agent whose task names its subject.

Journal

The notes in a project's memory: what agents wrote while they worked. Agents find notes only when they look.

Knowledge

A page of expert topics, each read once from web pages or project files into pages and a guide. Roles that know a topic use it while they work.

MCP server

A program on your machine that gives agents more tools. You add it on the Tools page, under Servers. Each tool asks before it runs unless you say otherwise.

Measure command

A command, set under Settings › This project › Checks and git, whose last line of output is a score from 0 to 100. It scores candidates.

Memory

What a project remembers between chats: your preferences, the project's rules, facts and notes. See Memory.

Note

A piece of memory an agent wrote with its Remember tool. See Memory.

Observer

A built-in role that never works on its own. If you switch on Have a model look over each sub-agent's report, its model reads each sub-agent's report and says what the record shows. Its words are an opinion, not a fact.

Orchestrator

The agent you talk to in a chat. It does small jobs itself and hands separate parts of the work to sub-agents. (Behind the scenes its role is named Lead.)

Outside content

Text that comes from outside your messages: web pages and searches, MCP tools' results, Knowledge pages and GitHub issues. KOLONİ frames it as data, not instructions, for the agents. See Outside content.

Page key

A secret that the KOLONİ page sends with every request, so that a program on your machine cannot approve cards or change settings in your place. It changes each time KOLONİ starts. Open KOLONİ from the link in its window.

Part

In the project map, a folder with its own project file (such as pyproject.toml or package.json), with its language and test command. In a recipe drawing, a step or a group of steps.

Permission level

What agents may do without asking: Ask before changes, Auto-edit files or Full auto. See Permission levels.

Plain chat

A chat with no folder ("No folder · a plain chat"). It is only talk: no files and no commands.

Plan first

A switch next to the message box. Each new message first gets a plan to approve before anything changes.

Project

A folder your agents work in. Each project has its own settings, memory and To-do list.

Project notes

A Knowledge topic that KOLONİ fills from what you tick on the Add to Project notes? card at the end of a turn.

Provider

Where a model runs: an OpenAI-compatible server, the Anthropic API, or a CLI you are signed in to (Claude Code, Codex, GitHub Copilot CLI, Qoder, agy). You add providers on the Models page.

Ready

A page on the rail that lists everything waiting for your word, with a count.

Ready for you

The panel and page that list everything waiting for your word, in three parts: Needs your answer, Ready to take in and Told you.

Recipe

The drawn steps of a piece of work: who does each step, what runs at once, where it loops back and how it ends. Recipes are on the Agents page.

Role

What an agent is for: a description, instructions, the tools it may use and, if you like, its own model. KOLONİ comes with built-in roles such as Planner, Coder, Tester, Reviewer, Researcher, Debugger, Architect and Docs writer. You can add your own.

Room

Two to six agents working on one goal at the same time, in one working copy. They share the work out first, and each file has one holder. You open one with Work together… under More.

Skill

A folder with a SKILL.md and its files, in the form Claude Code reads. You add skills on the Tools page and give them to roles.

Stop everything

The button in the top bar that stops every turn in every chat and holds triggers, To-do lists that run by themselves, and work while you're away. While it holds, it reads Held · Resume. See Chat and canvas.

Sub-agent

An agent that the Orchestrator (or another agent) opens for a part of the work. It works in its own working copy.

Take in

Bringing finished work from a working copy into your project. Every card that does this says Take in.

Team

A named group of roles with stages: Plan, Build, Check, and Write & keep up. A chat can work with teams, picked under Teams for this chat: Orchestrator decides, Manual or No teams.

Teams' work is ready

The card where a recipe's work ends. It asks whether to take the work into the project, with Take in, Fix the findings and Ask for changes.

Thinking

How hard a role's model thinks, set on its Model card: Auto (the agent that opens it picks a level for each task, up to a most) or One fixed level, on a ladder from Off to Max. See Roles.

Tier

How strong you say a model is: fast, balanced or strongest. You set it for each model on the Models page. A model with no tier is never picked by itself.

To-do

A project's tasks, in your order. Agents, and the Assistant PO after a turn, propose tasks; you decide.

Trigger

Something that starts a turn by itself: the clock, a change to the project's files, or a request from outside (a webhook). Nobody watches such a turn, so it may do only what the trigger allows.

Turn

Everything that happens for one message: the Orchestrator's answer, the agents it opened, their tools, commands and changes. The canvas draws one turn at a time.

Where the work is

The panel the git button in the top bar opens ("⑂ main · 4 changed"): your branch, the agents' copies in this chat, and other branches, each saying how the work stands. See Branches.

Workspace

The side window with the project's files, what agents changed, wrote or showed, and what waits for you.

Working copy

A copy of your project, outside your folder, where a sub-agent makes its changes. Its work comes back to you on a card to take in or discard.

How a CLI runs¶

A CLI provider has a setting How it runs: