✨ ShadRox v0.0.6 · Windows desktop beta See what's new →
Beta v0.0.6 · Windows 10 / 11 x64

Your project, your code, your AI —
one desktop workspace.

ShadRox brings a focused code editor, AI chat that runs on your own key, live project preview, Git tools and a real terminal into one native Windows app. Your files stay on your computer; only the prompt and context you choose are sent to your AI provider.

Windows 10 / 11 · 64-bit · Google sign-in at first launch · Bring your own AI key · Auto-updates

1// src/App.tsx
2import { useWorkspace } from "./useWorkspace"
3
4export function App() {
5 const { ready } = useWorkspace()
6 // Ctrl+K on a selection → inline AI edit
7 return <Workspace ready={ready} />
8}
@App.tsx add a loading state
Reading 2 files · round 3/12
- return <Workspace ready={ready} />
+ if (!ready) return <Skeleton />
+ return <Workspace ready={ready} />
✓ Accept Reject 2 files · agent
PREVIEW
localhost:5173 · live preview
⎇ main · 1 modified 0 errors 0 warnings TypeScript Gemini · BYOK

Works great with

0 project files kept in cloud storage — the app opens the folder on your disk
0 AI provider types: Gemini, Groq, OpenAI-compatible
0 agent roles in a multi-step team run
0 built-in agent tools in the bounded loop, from reads and search to one guarded shell command, a confirmed terminal answer and a paused-debugger snapshot

Everything in one place

Stop switching windows.
Keep every tool in one app.

Four things ShadRox does end to end — and the rest of the workspace, listed.

How it works

From folder to commit,
in three steps.

Open a local folder, choose your AI provider, and review every proposed change.

1

Open a folder

Pick any local project. ShadRox builds a compact repo map and loads your rules automatically.

2

Ask, then review

Chat, press Ctrl+K or run an agent. Every proposed change shows up as a diff first — you accept or reject.

3

Run & commit

Preview it live, test in the terminal, and commit with an AI-written message right inside the app.

Download for Windows

The agent

An AI that shows its work.

Bounded, reviewable and honest about its limits.

Reads first.
Then proposes.

The agent reads your project — files, search, diagnostics, Git — and proposes changes as diffs you accept or reject. Requests can include a compact repo map so it understands the codebase before touching anything.

  • Read-only tools: list, read, search, diagnostics, Git, outlines
  • Five tool rounds by default, with up to six calls per round
  • Edits go through a patch → review → apply pipeline
  • Unsafe edits are held for review, never silently written
1
Read the fileslist_dir · read_file · search
2
Check diagnostics & Gitknow what is broken first
3
Propose a patchnever an arbitrary write
4
You review & acceptalways your call

Plan. Patch.
Verify. Repair.

Big goals become a task DAG of milestones. Each is verified with diagnostics, and failures get a repair brief. Still beta — review the results.

  • Verified with diagnostics and an optional approved command
  • Configurable repair attempts per milestone
  • Unrecoverable failure rolls back to the checkpoint
G
Goalwhat you asked for
P
Plan (task DAG)milestones with dependencies
V
Verifydiagnostics + approved command
R
Repair or rollbackbounded by maxAttempts

A small team,
with boundaries.

A Coordinator splits the goal and assigns specialists. Each brief carries file ownership so agents stay out of each other's way.

  • Shared bounded tool loop and guarded patch path
  • The Tests role will not weaken tests to make them pass
  • Reviewer reads and reports — changes nothing
C
Coordinatorplans and assigns, never edits
F
Frontendcomponents, styling, client state
B
BackendAPIs, data access, services
T
Tests · Reviewerfixtures and checks · read-only review

Your tools.
Your rules.

Connect Model Context Protocol servers with approval for every external tool call, and teach the agent your conventions once via rule files.

  • User-configured MCP servers, approval-gated
  • Rules and skills load automatically from your project
  • Supports .cursorrules, AGENTS.md, SHADROX.md
MD
SHADROX.mdyour project conventions
M
MCP Serversapproval-gated external tools
S
Skills.shadrox/skills/*.md
R
Rules.shadrox/rules/*.md

Why not just…

Aider and Cline are powerful.
Neither one is the whole workspace.

Same review-and-accept discipline and bring-your-own-key pricing — plus the editor, live preview, Git UI and terminal in the same window.

Aider · Cline

Aider drives from a terminal and Cline is a plugin inside another editor — strong agents, but the workspace lives somewhere else. ShadRox keeps the same bounded, review-and-accept loop and adds the explorer, preview, Git tools and terminal in one app.

Cloud AI editors

Requests go from your machine to the provider you configure — not through a ShadRox server, and your key is never sent to one. Point it at Ollama or LM Studio and the prompt stays on your PC.

Editor + plugin stacks

One native app instead of six extensions. Open a folder and the explorer, chat, preview and shell are already wired together.

Privacy

Local-first.
By design.

ShadRox edits real files on your disk. Only the prompt and context you choose go to the provider you picked. Your key never touches ShadRox servers.

Your computerfiles · key · rules
Your providerprompt + context
1

Files stay on your computer

No ShadRox cloud workspace and no project sync. The app opens the folder you pick.

2

Keys in OS-backed encrypted storage

Your provider key is sent with the request to the provider you configured — never to ShadRox.

3

Guarded commands

Approval is default-deny. Only these exact forms run without a prompt: npm test, pytest, python -m pytest, tsc --noEmit, npx tsc --noEmit, git diff, git status. One extra argument or a shell operator breaks the match, so everything else waits for a native approval dialog — and the workspace is re-checked between approval and execution.

4

Reviewable changes

Edits arrive as diffs. Unsafe patches are held for review instead of written in place.

Beta, and honest about it.

ShadRox is beta software, not a mature replacement for a general-purpose IDE. Agent output and model-generated edits can be incorrect, so review changes and keep backups for important work.

  • Review every AI diff and keep Git or backups for important work
  • Google sign-in is required to enter the app
  • Agent, autonomous run, team, plugin, and automation features are evolving
  • Preview support depends on your project and installed runtimes
  • Turns on your own key are neither metered nor capped; the plan and credit-meter screens are previews, and public payments are not enabled

Download for Windows

Windows shows a SmartScreen warning for unsigned builds — More → Run anyway.

FAQ

Quick answers.

Everything you need to know before you download.

Open a folder.
Start building.

Download the Windows beta, add a provider key, and bring your local project workflow into one window.

Download for Windows

Free during beta · Google sign-in at first launch · Unsigned build: open SmartScreen's More → Run anyway.

Built by Sourav Maurya, a solo developer from Delhi, India 🇮🇳
Download for Windows