Providers¶

A provider is where your agents' models run. Open Models on the left rail to see them. Each provider is a card that shows how it is paid for (subscription for a CLI, this machine for a server on your own computer, API key otherwise) and what it has left. A click on a card opens its window.

The Models page with one provider, a model server on this machine.

Kinds of provider¶

Adding a provider¶

Press Add a provider ("A CLI with your subscription, an API key, or a server of your own"). Give it a Name and a Kind, then:

• For a server: its Address and, if it needs one, an API key. "Keys are kept in plain text in KOLONİ's local database on this machine."
• For the Anthropic API: an API key and a Default model.
• For a CLI: Path to the command, if it is not found by its name. KOLONİ never signs in for you. If the CLI is missing or not signed in, the window shows the command to install it or to sign in, with links to the vendor's docs.

Press Add provider to save it.

The provider window¶

The window has the tabs Models, Connection and Limits. A server on this machine also has Server.

Connection holds what you filled in when you added it, and:

• Default model: the model used when nothing else names one.
• At most this many agents at once: how many agents may run on this provider at the same time.
• How it runs (CLIs only):

• Sign-in (Claude Code and Codex): My own uses the sign-in in your home folder; Its own for KOLONİ keeps a separate one, and the window shows the command to log in to it. "KOLONİ never sees the token."

The buttons at the bottom:

• Save saves the changes.
• Test checks the provider. For a CLI it only says whether the command was found: "It was not run, so sign-in is not checked." For the Anthropic API it asks for the model's details, which spends nothing. For an OpenAI-compatible server it sends one small request ("Call the ping tool.") to see whether the model can use tools; on a paid service that costs a few tokens. It then says "Reachable" and the model.
• Try a real run (CLIs): runs the CLI once on a one-word question. This counts against your plan.
• Make default: makes this provider the default, the first one in the list.
• Remove key: shown when a key is saved.
• Delete provider: deletes it at once, without asking. The last provider cannot be deleted ("At least one provider is needed"). Chats and roles that used it go back to having no choice.

Switching a provider off¶

Each provider has an on / off switch, on its card and at the top of its window. "Off: nothing new starts on it - no chat, no agent, no learning; its key and settings stay. A turn already running goes on until it ends; Stop ends it now." Its card then says "switched off".

The models of a provider¶

The Models tab lists the models that agents and you can pick, at most 30. Each row has:

Open a row for more, such as Its note for agents (what the model is for, up to 120 characters) and Used in 30 days. Each change saves at once.

Add a model adds one by name. Fill from the catalogue proposes a list: your rows first, then what the provider lists with your key, then the bundled models.dev catalogue's models, with their details; nothing changes until you press Save these models. At the top of the page, Update catalogue downloads a newer catalogue, and Keep the catalogue fresh lets KOLONİ do that by itself.

Which model a chat uses¶

The model button under the message box opens Model for this conversation: a Provider (the first option is the default) and a Model name. Use this model applies it from the next turn. "A role with its own model keeps it."

An agent's model is chosen in this order:

1. the models the role lists under Try first, in your order, unless the agent that opens it gave the job a tier;
2. the role's own model (One model);
3. the first model of the role's tier, or of the tier the opener gave the job;
4. the chat's model;
5. the default provider, the first one in the list.

A model past its limit, down, switched off or not one agents may use is passed over, and the next is tried. Roles and their models are set on the Agents page.

Where an agent runs: the full order, tiers, limits and moves. See also Tiers and Limits.

Plan meters¶

For Claude Code, Codex, agy and Copilot, the top bar shows how much of the plan is used, from what each CLI last reported. Click a meter to open Usage. On a provider's card, Read now reads the plan again; for Claude Code that is one short request.

A model server on this machine: the Server tab¶

For an OpenAI-compatible server on this machine, the Server tab starts and stops your llama.cpp llama-server. Set Program to the full path of llama-server in your llama.cpp build (the one on PATH is filled in when there is one), choose the Model from the models folder next to it, set the context, GPU layers and the other values, then press Start. Server output shows what the server prints. "It starts only when you press Start, and stops when KOLONİ closes. No agent can start it; Start runs exactly what is shown here."