FluxCode

English · Bahasa Indonesia

A native code editor for Windows, macOS and Linux, written in Rust and drawn by the GPU. One small program with an AI agent panel, a Python-friendly terminal, built-in Source Control, and read-only views for files other editors refuse to open. No telemetry. It opens its window in 0.22 s and sits at 1.4 MB idle, in one process (benchmark).

A MuaraAI project. Open source, and still in development.

The agent added a docstring (line 23), the Changes card lists the file with Undo, and because access is set to Read & write, it has to ask before running a command.

Benchmark

On one Windows 11 PC, FluxCode put its window on screen in 0.22 s and sat at 1.4 MB in Task Manager once idle, in one process. Zed took 0.82 s and 66 MB; VS Code about 3 s and 929 MB over nine processes.

The gap is widest where an editor sits all day: idle memory. FluxCode uses 1.4 MB, Zed 66.2 MB, VS Code 929.2 MB. On disk it is 16 MB against 261 MB and 1,044 MB. Its window is on screen in 221 ms, against 818 ms for Zed and about 3 s for VS Code. Lower is better in every chart below.

Median of 5 runs each on a PC with an AMD Ryzen 5 7535HS and 15 GB of RAM, 10 November 2026. Every app started from a fresh, empty profile with telemetry and update checks off and no folder open.

Median, with the lowest and highest of the runs in brackets:

• Window: from launch to the first visible window, runs 2 to 5 (run 1 reads the program from disk cold). VS Code's figure is the time until the workbench is usable, about 3 s: its window shows before the workbench inside it has loaded.
• Memory in Task Manager: the private working set, 9 s after launch, summed over every process of the app. FluxCode hands memory it isn't using back to Windows 6 s after start, so this is its idle figure; it is about 5 MB after a click and about 30 MB while the agent install animation runs. Zed and VS Code keep theirs.
• Memory reserved: private bytes (commit), the memory the app has claimed whether or not it is in RAM right now.
• CPU while idle: CPU time from 9 s to 39 s after launch, as a share of one core. Windows counts it in 15.6 ms steps, so FluxCode's 0.05% is at the limit of what it can tell apart from zero.
• Size on disk: the install folder; for FluxCode, its one program file.
• Not measured here: typing latency, opening very large files, working in a big project, or anything with extensions and language servers loaded. Zed and VS Code do more in those areas (see the comparison table below: FluxCode has no debugger, extensions or language servers yet).
• Reproduce: close all three apps, then run powershell -ExecutionPolicy Bypass -File tools/bench/bench.ps1 -FluxExe <path to fluxcode.exe> and python tools/bench/chart.py. The raw runs are in tools/bench/results.json.

Download

Get the newest build from Releases.

The macOS build is a preview. We have not tested it on a real Mac yet. Try it and open an issue if something breaks.

The macOS build isn't notarized by Apple yet (that takes an Apple Developer ID), so the first open asks once: on macOS 15 and newer, open it, then System Settings > Privacy & Security > Open Anyway; on older ones, right-click FluxCode > Open. Uninstall: Settings > Apps on Windows, the Trash on macOS, ./install.sh --uninstall on Linux.

There are no version numbers yet: every change merged into main becomes a build named after the moment it was made, like BuildDev-2026-11-09-18-19-54. Installed copies offer the new one in a card and update themselves in one click. Settings > Updates turns the check off.

A closer look

A Python error becomes a card with the file, the line and a hint; for a missing module, Fix installs it. A SQLite database opens as its tables, SQL and row counts, like Word, Excel, PDF and ZIP files do, instead of an error.

Source Control stays fast with thousands of new files and offers to ignore the folders that hold them. Go to File (Ctrl+P) finds a file from a few letters of its name.

Six color themes, three interface sizes and your own background image. The interface also comes in Bahasa Indonesia, shown here with the update card in the FluxCream theme.

What makes it different

• Laid out and driven like VS Code. Activity bar, side bar, tabs, panel and status bar sit where VS Code puts them, and the keys are VS Code's: Ctrl+Shift+P for the Command Palette, Ctrl+P to find a file, Ctrl+G to go to a line, Ctrl+/ to comment, Alt+Up to move a line. If you know VS Code, you already know where things are. It is not a VS Code clone: there is no debugger yet, and the extensions are build tasks rather than a catalog (see the table below).
• Light like Zed. The whole editor is one Rust program drawn by the GPU with GPUI, the framework Zed is built on. In the benchmark it opens its window in 0.22 s, reserves 72 MB (Zed: 154 MB), and once idle shows 1.4 MB in Task Manager, because it hands the memory it isn't using back to Windows.
• Built for AI agents. The agent panel speaks the Agent Client Protocol and lists every agent in the ACP registry, each with its own icon. Missing ones install with one click into FluxCode's own folder. You choose the model, effort, mode and permissions under the message box; the agent can ask you multiple-choice questions and has to ask before it runs a command or changes a file; each turn's file changes can be undone.
• GitHub Copilot inline completions. A moment after you stop typing, Copilot suggests the next lines as ghost text; Tab accepts, Esc dismisses. The Copilot language server runs as its own process and does the model work, so the editor stays light. Install it once with npm install -g @github/copilot-language-server, then sign in under Settings > AI agents and GitHub > GitHub Copilot (a Copilot plan is needed; a free plan exists).
• For beginners and experts. A beginner can pick an agent from a list, answer its questions by clicking, and keep it read-only. An expert gets the palette, the shortcuts, per-agent instruction files and every editor option in Settings: font, size, tab size, word wrap, whitespace, line numbers, folding, bracket closing, smart indent, cursor padding, scrolling past the end, interface size, themes and a background image.

Live share

Code together in real time, with your PC as the server. The people icon in the activity bar opens the Live share page in the editor area: its cards sit two to a row on a wide window and one under the other on a narrow one.

1. The host opens a folder, clicks Start hosting and sends the invite code (FLUX1-...) to the others, in a chat or by mail.
2. Each guest pastes it and clicks Join, either into their own clone of the same repository or into a new folder that fills as they open files.
3. Everyone types in the same files at once. The host's PC orders the edits (operational transformation), so everyone ends with the same text.

The panel lists who is in the session (by GitHub name when signed in) and which file and line each person is on. Its activity list shows each person's edits with lines added and removed and a diff; an edit made by an AI agent carries an AI label with the agent, model, effort and mode it ran with, so you can see which model wrote what.

Edits go between the host's PC and each guest, encry