Discord,without the browserinside.

1,279 MB to say good morning.

An Electron client is a whole browser dressed up as an app. One square is 5 MB of RAM.

larperCord: 125–140 MB in one account's normal use and about 70 MB right after opening. Yours depends on your servers, members and images. Idle CPU was measured across all 11 themes with GIF autoplay off, and an animated GIF on screen costs about 2.5%. About 35 MB of the PSS is the NVIDIA driver, libraries and fonts. Vesktop: average since launch, measured with ps. Both on the same PC: Arch Linux, Hyprland, i5-2400S, GTX 1050 Ti, 1280×800. Discord Phoenix: not measured by us and not a side-by-side test. It's an early Windows 11 build, from a Task Manager screenshot posted by Discord's CTO. How this was measured →

The interface is a file. Go on, break it.

These are the real layouts from themes/*.kdl. Change a width, move a node, rename a component. In the app, saving the file reloads the window through inotify. An error points to file:line:column and the previous theme stays on screen.

Full theme language reference →

Eleven themes. Not eleven color schemes.

Same binary, same data, different file. Components move, change shape and swap variants. Move your pointer across a screenshot to see light mode. Click one to dress this whole page in its colors (or press F2, like in the app).

It speaks the real gateway. It's a real client.

Real calls. Encrypted end to end.

• A mixer with one volume per person. Each person is their own audio stream on the same output device.
• No call, no cost. The voice thread is only created when you join. Before that, audio only starts if you open the voice settings.
• Your mic, your rules. Pick your devices, even mid-call. Voice activity with an automatic threshold or push-to-talk, a high-pass filter, a noise gate and RNNoise noise suppression, a live meter and a mic test.
• Honest limits. Push-to-talk only works while the window is focused. No screen sharing or camera. Voice guide →

Lua, in a cage.

-- @name CVE notes
-- @capabilities messages, decorate, http
-- @hosts cveawg.mitre.org
larper.on("message", function(msg)
  local id = msg.content:upper():match("CVE%-%d+%-%d+")
  if not id then return end
  larper.http.get("https://cveawg.mitre.org/api/cve/" .. id, function(res)
    local cve = res.ok and larper.json.decode(res.body)
    local cna = cve and cve.containers and cve.containers.cna
    if cna and cna.title then
      larper.decorate(msg.id, { note = id .. ": " .. cna.title })
    end
  end)
end, { contains = { "cve-" } })

larperCord

Demo data, but the CVE and GitHub notes are real answers from those services. Nothing goes out until you approve the host on the plugin's card.

• Scam guard. Free Nitro links, dlscord-nitro.ru, addresses in another alphabet, a link that shows one site and opens another: marked in red. It runs on your computer, with no network.
• CVE radar. Someone writes CVE-2021-44228, and the severity, the CVSS score and a summary appear under the message, from the official record.
• GitHub cards. owner/repo#123 and GitHub links get the title, the state and the stars. A renamed repository shows its current name.
• What did I miss. /missed sums up the unread messages: who talked, who called you, which questions nobody answered, links and files, with a click to each one.

No magic. Ten boring decisions.

1. It sleeps when nothing happens. The main loop waits in SDL_WaitEvent. Network, inotify and the clock wake it up with events. No animation, no redraw.
2. No browser. Everything is drawn with Dear ImGui's ImDrawList on top of SDL_Renderer.
3. Fonts are mapped, not copied. Each font file is one mmap shared by every weight, so only the pages read stay in RAM.
4. Only the video driver it needs. On NVIDIA-only machines, glvnd would also load Mesa and LLVM for nothing. The app detects that and avoids it.
5. Markdown is parsed once. When the message arrives. After that, the UI only lays out spans, with a height cache per width.
6. A virtualized list. Binary search for the first visible item, then draw only what's on screen.
7. Memory has ceilings. 1,500 messages per channel and 3,000 in total. Forgotten conversations get unloaded, and the image cache has a fixed 64 MB budget.
8. fontconfig gets freed. After fonts are resolved: FcFini, then malloc_trim.
9. No blinking cursor. A blinking cursor would keep the CPU awake for no good reason.
10. Things start when you ask. The voice thread on your first call, FFmpeg on your first video. With no plugin on, there's no plugin thread and no Lua, and nothing is spent per message, key or frame.

The UI never changes data directly. It only reads the Store and asks for actions, which is why swapping the demo gateway for the real one doesn't touch the UI. Architecture →

What works today. And what doesn't.

Roadmap →

Get it. Run it.

No account needed to look around: the demo mode connects to nothing. After that it updates itself, and every update is signed.

All versions and checksums · What changed

Three commands. No account.

# Arch Linux
sudo pacman -S zig sdl3 freetype2 fontconfig curl zlib libwebp opus openssl \
    inter-font ttf-jetbrains-mono noto-fonts-emoji

# Build (Zig 0.17)
zig build -Doptimize=ReleaseFast

# Demo mode: fake data, connects to nothing
./zig-out/bin/larpercord --demo --theme subsolo