Skip to content
sorryhyunPublic

About

You Are Absolutely Right!

Resources

Contributing

Stars

6 stars

Watchers

0 watching

Forks

Repository files navigation

YAAR

Software you can rewrite, keep, and share — in a shape any agent can drive.

TypeScript React Bun MCP Claude Agent SDK License

한국어

YAAR Desktop

YAAR is a local desktop where the agent you already use — Claude Code or Codex — builds the apps, and you change them by talking to the app you're looking at. Every app is a folder on your disk with its own git history. Nothing is hosted, nothing is rented.

"Build me a Tetris game"                 → writes it, builds it, opens it
"Make the pieces fall faster"            → edits the running app in place, redeploys, restores it
"Undo that"                              → rolls back to the previous deploy
"Publish it"                             → ships the source to YAAR Market

Chat assistants regenerate; app stores make you rebuild and resubmit. YAAR is the thing in between: an app you're using is an app you can rewrite, right there, with the agent that wrote it.

Install

Codex or Claude Code authentication is required.

For Linux, Mac, Android (termux) users:

curl -fsSL https://github.com/sorryhyun/yaar/releases/latest/download/install.sh | bash
yaar                # Browser opens automatically

On macOS the installer puts the app in ~/Applications/YAAR.app (its own window, and the microphone permission apps need) and yaar on your PATH starts it. How it runs: docs/installations/mac.md. On Linux, yaar opens the desktop in Chrome or Chromium in app mode: docs/installations/linux.md.

For Windows users:

irm https://github.com/sorryhyun/yaar/releases/latest/download/install.ps1 | iex
yaar

Android (Termux, Claude only): With no Android binary to download, it installs Bun's Android build, checks out the source to ~/yaar, and puts a yaar launcher on your PATH. Guide

Once running, start with something like "install essential apps".

Other install options

Pin a version / custom install path:

VERSION=v0.25.10 curl -fsSL ... | bash             # Specific version (default: latest)
INSTALL_DIR=/usr/local/bin curl -fsSL ... | bash  # Install path (default: ~/.local/bin)

Windows: You can also download yaar.exe directly from the Releases page.

Bundled apps ship separately as yaar-apps.tar.gz; the install scripts extract them next to the binary automatically.

Build from source (requires Bun >= 1.4):

git clone https://github.com/sorryhyun/yaar.git && cd yaar
bun install
make dev          # Browser opens automatically

Why YAAR

  • You own the app. One folder = one app: app.json, source, and an optional agent prompt, compiled to a single self-contained HTML file. Drop it in to install, delete it to uninstall, git it wherever you like. Every deploy is snapshotted to a shadow git repo, so any edit the agent makes is a diff you can read and a version you can restore.

  • You edit it in place. Dev Tools is an app that edits other apps: it clones the source, changes it, previews the result live, and redeploys — without leaving the desktop and without starting over. "Customize this for me" is a first-class operation, not a new conversation.

  • Every app speaks a contract. An app publishes a manifest — typed commands, state keys, event channels — so the agent that edits or drives it works against a schema, not a screenshot. Any agent that can speak yaar:// can drive any app.

  • You bring your own agent. YAAR drives Claude Code or Codex as a subprocess using the login you already have. No account, no API key, no hosted service; your conversations go wherever your provider already sends them, nowhere else.

  • The agent answers with UI, not paragraphs. Ask for an analysis and you get a chart window. Windows persist, are addressable, can be messaged individually, and stay live with their data — the server pushes updates when a yaar:// resource changes, no polling, no re-asking.

  • Five verbs, flat prompt. Everything — windows, files, apps, config, other agents — is a yaar:// URI, and describe · read · list · invoke · delete operate on all of it. Capability discovery happens at runtime, so the system prompt stays ~8K tokens with 3 apps or 100.

Curious about the reasoning — why a GUI instead of a TUI, why it's shaped like an OS, why the web? See the FAQ.

The whole desktop is an input

Input What happens
Typing Send a message
Paste image / drag & drop Send an image to the agent
Right-click drag Sketch — the agent turns it into code or diagrams
Button click inside a window Execute that window's action
Right-click → select window Talk to one specific window
Drag a file/selection to an app Move data between apps

Build apps

Apps are plain TypeScript with batteries included:

  • Bundled libraries — Solid.js, lodash, Three.js, Konva, Chart.js, D3, Tone.js and more via @bundled/*, no npm install
  • appDb — per-app isolated SQLite with Mongo-style filters and FTS5 full-text search (reference)
  • App agents — add agent/prompt.md and the app gets its own agent; declare controls and it can drive other apps
  • Gated SDKs — declare them in app.json to unlock yaar-dev (compile/deploy), yaar-web (browser automation), yaar-ml (in-browser ONNX inference)
  • YAAR Market — install from the catalog or publish your own; the market ships source, and installs compile locally

See Building Apps with the YAAR SDK and, for how an app is built and run, Apps: from source to a running window.

Share apps

  • As a file — every app compiles to one self-contained dist/index.html. Send it and it opens in any browser, no YAAR needed; whatever runs in the page works, saving and loading data does not.
  • As source — copy the app folder into another YAAR, or publish it to YAAR Market; the receiving machine compiles it and can keep changing it.
  • As a live window — the wifi button in an app window's titlebar copies an MCP URL. Hand it to Claude Code or any MCP client and that agent drives the window you're looking at, with the window's own app-agent authority. Closing the window revokes it.

Guide

Trust model

YAAR lets an agent write and run code on your machine, so it is built assuming you don't trust it:

  • Scoped filesystem — the agent sees storage/, config/, apps/, session_logs/; anything else must be mounted explicitly (read-only supported)
  • One access gate — every route resolves who is calling and what URI + verb they want through the same check
  • App permissions — an app touches its own namespace plus whatever its app.json declares; installs prompt for the rest
  • Origin isolation — apps run on a different browser origin than the desktop, so they can't forge desktop requests or reach its DOM
  • Agent tiers — the dangerous namespaces (yaar://session/*, including real-browser control) are reachable only by the privileged session agent
  • Network allowlist + SSRF guard — outbound HTTP is limited to approved domains; internal addresses are blocked

Full detail, including what the sandbox does not cover: Security and the OS Architecture Map.

Also

  • Multiple desktops — each monitor has its own agent and history; a session agent coordinates across them. Sessions survive a closed tab; rejoin with ?sessionId=X.
  • Remote access — make claude / make codex print a QR code; connect from your phone over Tailscale Serve. Guide
  • Hooks — event-driven automation in config/hooks.json. Guide
Browser (UI) ←→ Local Server ←→ Claude Code / Codex

Development setup and architecture: CLAUDE.md.


YAAR: You Are Absolutely Right — the phrase every heavy agent user has read a few hundred times. If the agent is going to say it anyway, it may as well run the desktop.

About

You Are Absolutely Right!

Resources

Contributing

Stars

6 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages