Set up a CLI¶

A CLI provider runs an agent CLI that you installed and signed in to yourself: Claude Code, Codex, GitHub Copilot CLI, Qoder or agy. It runs on your plan with that vendor. KOLONİ never installs a CLI and never signs in for you. Read Subscription rules first.

Steps¶

1. Install the CLI in your own terminal. If it is missing, the provider's window shows the install command for your system, with a link to the vendor's install docs.
2. Sign in in your own terminal (see the table). If it is not signed in, the window shows the sign-in command.
3. Models → Add a provider, and pick its Kind. Set Path to the command only if KOLONİ does not find it by its name.
4. Choose How it runs (see below), and press Add provider.
5. Test only checks that the command is found; it does not run it. Try a real run runs it once on a one-word question, on your plan.

At the first start, KOLONİ adds Claude Code, Qoder, Codex and agy by itself if they are installed. It never adds Copilot by itself. It adds agy switched off: Google's terms forbid using Antigravity through another app (see Subscription rules). Its card on the Models page, switched on or off, says "Google's terms forbid using Antigravity through another app; your Google account may be suspended." KOLONİ adds providers only at the first start, so an agy already on the Models page is left as it is.

The CLIs¶

How it runs¶

• One session per agent and One run per model step use KOLONİ's tools and cards, with the CLI's own tools off.
• Native uses the CLI's own tools in your folder. Claude Code, Qoder and agy ask KOLONİ before every tool they use; Codex before each command and patch, except the commands its own rules count as safe, which it runs without asking; Copilot before each command, write and read.
• Reads with its own tools (agy): it reads, and is started without leave to write or run.

A trigger never starts a CLI that works with its own tools (Native, or agy's reads): with nobody watching, only an agent that works with KOLONİ's tools runs.

The modes, KOLONİ's tool bridge, and how each CLI asks. (full size)

Going on from the last turn¶

Only Claude Code, run Native with Sign-in: Its own for KOLONİ, goes on in its own session from one turn to the next: it is sent only what is new, and compacts that session itself. Every other CLI lead starts each turn anew and is sent the whole chat. KOLONİ compacts that, as it does for an API model, past Summarise a long context at (%) of the window set on its role; with no window set there, only Compact now compacts it. A CLI run One run per model step is never compacted: each step is sent a trimmed history.

Compact now is in the top bar's context button (see Chat and canvas).

Qoder's free models¶

Qoder puts a free model's request in a queue when it is busy. In tests the wait was about 1 to 3 minutes, once longer. While it waits, the agent's window says so, for example "Qoder: in the model's queue for 2 min 43 s (waits up to 60 min)", and its box on the canvas says "queued 2 min". A sub-agent still queued after 90 seconds is moved to the next model of its tier, where KOLONİ picked its model and the tier has another; a move that may cost money is asked on a card first (see Ready and approval cards). In a turn you watch, a lead, or a sub-agent that cannot be moved, waits as long as Qoder keeps it queued; in a trigger's or a job's turn while you're away it gives up after 90 seconds. Run Native, Qoder shows no queue line, and is not moved.

Sign-in for Claude Code and Codex¶

Sign-in offers My own (the sign-in in your home folder) or Its own for KOLONİ: a sign-in folder for this provider alone, so that several accounts can run side by side. For its own, save the provider, then the window gives the command to log in to it in your own terminal. "KOLONİ never sees the token."

Claude Code and Codex can also Use a model server of your own instead (optional): a Model server address on this machine or your network.

GitHub Copilot CLI¶

• API key: leave it empty to use your own GitHub sign-in (gh), or give a fine-grained GitHub token with the Copilot Requests permission. A token has not been tried on a real account yet.
• Thinking: KOLONİ sends Copilot's levels Off to Max (Off as none), whether it runs as a session or Native; the Models page shows "Off–Max". A level the model does not take gives no error.
• Model auto: Copilot picks the model for each request. It is in Copilot's model list (tier fast) and takes no thinking level, so KOLONİ sends none. Pick a model for Copilot, auto or another: with no model and Medium thinking, Copilot's default spent 7.12 AI credits for the same short answer that auto answered for 0.18 (measured once, 7 Nov 2026, Copilot CLI 1.0.105). The agent's node and window show the model auto picked (auto → gpt-6-luna), and its Details say why; an agent that goes on from it asks for auto again.
• Usage: Copilot's tokens and AI credits (or premium requests) are counted from its usage checkpoint, run as a session as in Native. As a session that was measured with copilot -p (CLI 1.0.103), not yet seen in a live session. See Usage and budgets.
• Copilot can also use a model server of your own, with Server type, Most prompt tokens and Most output tokens. Never connect to GitHub then means no sign-in, telemetry or update check, and web tools are off.

Plan meters¶

For Claude Code, Codex, agy and Copilot, the provider's card shows what is left of the plan, and Read now reads it again (for Claude Code, one short request). For Claude Code, the Limits tab also offers Keep it fresh from your Claude Code sessions: a status line you add to your own Claude Code settings, so that your sessions report what is left of your plan.

See CLIs for what has been measured.