An agent is the coding assistant that does the work in a thread. CodeInOven coordinates it rather than replacing it: the app owns the project, the conversation, and the review, and the agent owns the reading, the commands, and the edits.
Pi comes with the app
Pi is bundled. A fresh install can send a message without installing another CLI and without signing in anywhere first.
The first-run guide walks through the agent list once, so a new install can pass through it without choosing anything.
The other agents the app knows are listed in the same place, and the agent list marks Pi as the one that ships with the app.
Bring an account, a key, or a local server
Open Settings from the rail, then Harnesses. The app wraps the agents you already use, so the account you connect is the account you already pay for.
- Sign in with the account you use for a provider.
- Paste an API key when you would rather connect that way.
- Connect a model on your own machine through a base URL.
Accounts and keys go into the app's encrypted vault on this computer, and the Accounts tab shows what is connected for each agent. The model picker in the composer lists the models those accounts can reach.
The Base URL providers tab connects a local server. Ollama, LM Studio, and llama.cpp are pre-filled, and any server that speaks an OpenAI-compatible or Anthropic-compatible API works the same way. When the model answering you runs on your own machine, the prompt and the reply stay on your own network. A saved server appears in the model picker like any other provider.
Provider usage and agent subscriptions still apply. CodeInOven is free and open source, and it does not stand between you and the provider. If a run needs to leave this computer, it leaves through the account you connected, under that provider's terms.
The other agents
Pi, Codex CLI, Claude Code, OpenCode, Cline, Antigravity, and Muse Code are the agents the app knows. Each one is a real agent with its own models and its own accounts, and the app coordinates whichever one a thread uses.
Connecting one takes a single step. Pick it from the agent list in Settings. If this machine already has it, its accounts open, and you choose the account to use when sending a message. If this machine does not have it, the page that installs it opens instead. Nothing is installed without that step.
Each agent reports its own state. The Harnesses tab shows whether an agent is installed, what version it reports, and how many accounts are connected to it. An agent that is installed and signed in can answer a thread; one that is not offers the way to get there.
Model and thinking level
The composer holds four decisions: the chat options menu, attach, the permission level, and the model with its thinking level.
The model control names the model that answers and lists what the connected agents can reach. An empty list means no account is connected yet, and the picker offers to connect one. The thinking level sits beside the model, and appears for models that offer one, so a hard problem can be given more room and a small edit less.
Set the model before you send. It belongs to the turn you are about to make, and every message after it stays in the same thread with the same history.
Where the agent settings live
Settings from the rail is the one place for all of it.
- Harnesses holds the agents, their accounts, their background models, and the base URL providers. Its tabs split connected agents from the accounts attached to them.
- Auxiliary Agents assigns the model each agent uses for background work, so the model that answers you is not always the model doing the small jobs.
- Agents holds the roles the app uses to plan, implement, and review work.
Nothing here is required before a first message. Pi is already available, and an account can be connected when you decide which provider you want to use.
Next
Projects covers where a thread lives, the composer covers the decisions you make before sending, and Getting started ties the first four moves together.