Where it lives
The Code agent is in the right panel: CmdShiftA shows or hides it, Cmd3 puts the cursor in its box, and CmdL (or + at the top) starts a new chat. Chat history keeps every conversation, per project. You can pull the agent into its own window, and Piye Studio runs it across many projects.
What it can do
On Piye credits, the Code agent works the way a senior engineer at your keyboard would:
- Read and search the project, and any other folder on your computer except the places that hold secrets.
- Edit, create, rename and delete files, with every change listed and undoable.
- Run commands in the terminal (in the command sandbox), start dev servers and watch their output, and run the project's own type check, lint, tests and build.
- Use the language tools: outline, go to definition, find references, rename across files.
- Look at your app: open the preview, read its errors, click through it; run it on simulators and phones and look at the screen.
- Search the web and read docs pages.
- See the computer: the system, memory and disk, the apps installed and the command-line tools they carry, so it finds and uses the programs you have (a virtual machine's tool, a cloud CLI) instead of telling you to.
- Use your connections: MCP servers such as GitHub, a database or Figma. See MCP and the Marketplace.
- Read Operations and deploy, with your approval.
- Work in another project on your computer from the same chat. The first change or command there asks once for the chat; each change there has its own Undo.
- Hand work to helpers that run in parallel, in this project or another one Piye knows.
On your own API key or a local model, the agent has the standard IDE tools: files, the terminal, the editor and the chat. See Models, effort and keys.
Four modes
| Mode | What it does |
|---|---|
| Build (the default) | Does the work: edits, commands and checks. Every change can be undone. |
| Ask | Looks wherever the answer is (this project, your other repos, this computer, the web) and changes nothing. |
| Architect | Reads the project and writes a plan you agree before anything changes. |
| Debug | Reproduces the bug, finds the cause, fixes it and shows it gone. |
See Ask, Build, Architect, Debug.
Giving it context
- Type
@to reference files. - Attach Files & Folders, a Screenshot of a window, Source Control changes, or Recently opened files (up to 24). Drop images straight in: they go to the model.
- Settings, AI, Context sets how much of the editor goes with each message by default: the message only, the active file, or the file and its problems.
The agent also reads the project's rules (AGENTS.md, CLAUDE.md, .piye/rules.md and others) and its own notes. See Rules, memory and skills.
While it works
The chat stays calm: the work folds into a live line with the current step, which you can open for every detail (files read, commands and their output, the thought process).
- Correct or add to it: type and press Enter; it reads your message at its next step, without stopping.
- Queue the next task: it runs after the current one finishes (CmdEnter waits).
- Stop it at any time. What's done stays, with its Undo.
One agent works in a project at a time; if it's busy in another chat or window, the box says where.
When it finishes
The reply says what changed and what's left, with the files it touched (each with Undo) and the credits it used. Retry asks whether to restore the files it wrote first or keep them. Big pieces of work get a choice before they start: Plan it first (Architect) or Go straight.
What it asks before doing
Pushing, deploying, publishing, deleting database data, rm and sudo, secret files, and commands that change folders outside the project always wait for you, in every mode. In Review mode (the default), every file change and command waits too. See Review, Auto and undo and Security.