A coding agent for your terminal

Explore, plan, edit, test.
In your terminal.

RoomKit Code is a coding agent built in Rust on RoomKit. You choose the model and how much it can do on its own, and the sidebar shows the files it changes.

Linux and macOS MIT licensed Written in Rust

RoomKit Code after a fix: the test suite exits with 0, the plan shows two of two steps done, and the sidebar lists the one modified file and the context used.
Real capture of roomkit-code 0.1.0. The tools, the files and the test run are real. The model’s replies are scripted.
Early version · v0.1.0

Commands and settings can still change between versions.

Release notes

How much it does is up to you

Three modes.
One keystroke apart.

01 / Plan

Read first.
Change nothing.

In Plan mode the agent explores the project and saves a plan you can follow in the sidebar. The application blocks file changes, commands and MCP calls, whatever the model asks for.

Web research still asks for your approval.

Plan mode: the agent has read the failing test, explained the cause and saved a two-step plan. The sidebar says no files changed.

02 / Ask

You see the change.
You decide.

In Ask mode, a change or a command waits on a card that shows the exact arguments. Allow it once or deny it, and the agent carries on. Yolo mode runs them directly, when you trust the task.

Shift+Tab changes the mode.

An approval card asks to allow an edit, showing the old text, the new text and the file path, with the choices allow once and deny.

03 / Tools

Give each agent
what it needs.

Search and outline code, edit files, run commands and long-lived processes, read the web and keep plans. /tools lists the 29 tools available, and /agents chooses them for each agent.

The tools panel lists the 29 registered tools with a filter, starting with code_outline, list, glob, grep and read.

04 / Delegate

Hand off a task.
Keep working.

An agent can hand a task to another agent of the project. That agent works in the background, in a conversation of its own, and its result comes back to the first. /tasks follows each task as it runs, stops it, or opens its room so you can write to its agent.

A delegated agent works in the same mode. In Ask, its approvals come to you.

The Code agent hands the review of its change to a Review agent. A line says the task completed in 2 seconds, then Code passes on what Review found.

What comes with it

Your model.
Your project’s rules.

Any compatible model

Use any OpenAI-compatible Chat Completions endpoint, local or in the cloud, or sign in with ChatGPT. The model must support function calling.

Your instructions

It reads your project’s AGENTS.md. Skills stay off until you enable them for an agent.

MCP and hooks

Connect MCP servers over stdio or Streamable HTTP. Run your own commands around tool calls and turns with command hooks.

Sessions that last

Each project keeps its conversations in SQLite, and /resume reopens them. When the model’s window is known, a long conversation is summarized before it fills it, or at once with /compact.

Scripts and CI

--prompt sends one message without the interface and exits when the agent’s turn ends. Add --json to get every event as one JSON line.

You stay in the loop

Write to the agent while it works, or press Esc to interrupt it and keep your draft. The sidebar shows the context used, the plan, the files changed and, when the model’s price is known, what the conversation cost.

Take the parts you need

Built from parts
you can reuse.

No Rust needed

Install it.
Open a project.

01 Download the installer and run it

curl -fsSLO https://github.com/roomkit-live/roomkit-code/releases/latest/download/install.sh sh install.sh

02 Open your project

roomkit-code /path/to/project

03 Add a provider with /providers, then an agent with /agents

Linux x86-64 and ARM64 with glibc 2.35 or newer. macOS 12 or newer, Intel or Apple Silicon. The installer checks the binary’s SHA-256 and installs in ~/.local/bin, without sudo. macOS builds have an ad-hoc signature and are not notarized.

The RoomKit Code welcome screen, with the roomkit. wordmark, the version and a message input.

Before you install

A few things
to know.

Which models work?

Any model behind an OpenAI-compatible Chat Completions endpoint, for example a local server, as long as it supports function calling and streaming. You can also sign in with ChatGPT. The RoomKit model catalogue shows which models declare tool calling, with their context and list prices.

What leaves my machine?

Model requests go to the endpoint you selected and can include the file contents its tools return. The web tools reach the sites they search or read. Update checks contact GitHub at most once a day, send no project, conversation or provider information, and can be turned off.

Where are my API keys stored?

In a private local file with 0600 permissions, not in an operating system keychain. They are never shown in the interface or stored in chat history.

Is it a sandbox?

No. Commands run with your own user permissions, and the project directory is where they start, not a limit. The modes decide what runs and when, not what a command can reach.

Does it run on Windows?

Not today. Linux and macOS are the supported platforms.

Open source, from the engine up

Try it on
your own project.

Start in Plan mode. It reads your code and changes nothing until you say so.