Mutter Web

A browser client for Mumble that speaks the protocol itself, so a machine where you can't install anything can still be a first-class participant — including our screen-sharing extension, which stock Mumble clients don't implement.

Why a bridge

Browsers can't open raw TCP sockets, and Mumble's control channel is TLS over TCP. So a small local process relays the browser's WebSocket to the real server:

Browser (Windows, no install, no admin)
   │  WebSocket → localhost
   ▼
Bridge (Node, in WSL)
   │  TLS/TCP :64738
   ▼
Mumble server — unchanged, no plugins, nothing to install

WSL2 forwards listening ports to Windows' localhost, so the browser reaches a bridge running in WSL at the same address. The server operator does nothing.

Voice goes over UDP, like a native client. The browser tunnels voice frames to the bridge over the WebSocket (localhost never drops a packet); the bridge takes the server's CryptSetup key, encrypts each frame with Mumble's OCB2-AES128 and sends it as a UDP datagram, decrypts the server's datagrams back into tunnel frames, keeps the UDP pings going, asks for a nonce resync when decrypts stall, and falls back to the TCP tunnel by itself if UDP is blocked (--tcp forces that). This matters on Wi-Fi: voice over TCP freezes for a second or two on every lost packet while TCP retransmits, then dumps the backlog; over UDP a lost packet is just a lost 20 ms. The Server tab says which lane is in use and counts delivery stalls, playback underruns and capture stalls, so a choppy call can be diagnosed from the log.

Running it

As a desktop app — download a build from the releases page and run it; no install, no admin, no browser. It is the same client and bridge wrapped in Electron, which matters on a managed laptop: the browser's enterprise policy can block WebRTC, and the app's own Chromium doesn't read it. See desktop/.

From source in a browser — Node 18+ is the only requirement. There are no dependencies here in web/: the WebSocket server is hand-rolled precisely so a locked-down machine needs no npm install.

node web/bridge/server.mjs

That starts the bridge and opens Mutter as its own window — Chrome/Edge app mode: no tabs, no address bar, its own taskbar entry, remembers its size. From WSL it launches the Windows browser (Edge by default; BROWSER=chrome to pick). --no-open just serves. You can also install it from Chrome's address bar ("Install Mutter") for a permanent icon.

Set PORT to use a different port. Open it as localhost, not a LAN address: the microphone and the Opus codec are only available to secure origins, and localhost counts as one.

The bridge binds only to 127.0.0.1, checks the HTTP host and WebSocket origin, and requires a per-process token obtained from its same-origin /bridge-token endpoint. It is intended for a browser on the same machine (including WSL localhost forwarding).

Server certificates are checked before any Mumble credentials or voice are sent. A self-signed certificate needs explicit approval; its SHA-256 fingerprint is remembered for that server. A changed certificate prompts again, even if its new certificate is publicly trusted. Check unfamiliar fingerprints with the server owner before accepting them.

The browser remembers server details and certificate pins. Passwords and TURN secrets stay in memory for the current session and clear on reload. External chat images load only after you click their load button. The desktop app can save passwords with OS credential protection.

Make it feel installed — without installing anything

On a managed laptop that runs Node in WSL but treats new Windows executables with suspicion, this gets you a Start-menu app that just works, using only the browser IT already approved:

1. Install it as an app. With the bridge running, open http://localhost:8788 in Edge or Chrome and use Install Mutter in the address bar (Edge: … → Apps → Install). It gets its own window, taskbar icon and Start-menu entry, stored in your browser profile — nothing new is written outside your user folder.

2. Start the bridge with Windows. Press Win+R, run shell:startup, and create a shortcut there whose target is

   wsl.exe -d Ubuntu -- bash -lc "cd ~/mutter && exec node web/bridge/server.mjs --no-open"
   

   (your distro name in place of Ubuntu, your clone's path in place of ~/mutter). Set the shortcut to run minimised. From then on the bridge is up by the time you sit down, and the app icon opens straight into Mutter. No admin, no installer, no new binary — the only executables involved are wsl.exe and the browser.

The one thing this cannot fix is a browser policy that blocks WebRTC; see Settings → Screen share → Test my connection. Everything else — voice, chat, images — is unaffected.

Voice needs Chrome or Edge (WebCodecs + AudioWorklet). Other browsers get chat only.

What works

The desktop layout has a server rail on the left with your saved servers, a channel sidebar with the voice panel and your user panel at the bottom, the chat with a channel header, and a member list on the right. Under 880px it becomes the phone layout — one pane at a time behind a tab strip.

• Connect, saved servers, reconnect with backoff after a drop (fresh roster, no ghost "old me").
• Channel tree with people nested under their channel, search, collapse (remembered per server), counts, a join arrow on hover, status-dot avatars (speaking ring, muted, deafened, sharing). Members grouped by channel on the right. Click anyone for a profile card: banner, badges, connection stats, message / mute-for-me / volume / watch / join them.
• A voice view (the speaker icon in the header, the channel row in the voice card, or the voice pill on a phone): a tile per person in your channel with a speaking ring, mute and deafen glyphs and a LIVE badge, and a bar with hold to talk, mute, deafen, camera, share screen and disconnect. Cameras play inside the tiles (640×360, capped at 900 kbit/s, peer to peer like screen share) and only stream to people who have the voice view open; a switch button appears when you have more than one camera.
• Voice both ways over the TCP tunnel: 20 ms Opus at 16–96 kbit/s, per-user jitter buffer that reorders by sequence number, conceals lost packets with a mirrored decaying copy of the last one, and sizes itself from what it actually saw: it grows on a near miss before anyone hears a gap, and only trims latency after a stretch of real audio with real cushion. Cleaning: the browser's echo cancellation and gain control, then RNNoise — the neural denoiser Mumble desktop uses, compiled to WebAssembly and run in the worklet (−22 dB on realistic background noise, voice within 1 dB; web/vendor/rnnoise) — or the iOS app's spectral suppressor (Light / Strong, with a keyboard-click ducker). Voice activity uses RNNoise's voice probability over the room's noise floor; push to talk on any key or mouse button you pick in Settings (Space by default; the browser only hears keys while the window is focused); always on. Mute, deafen, local per-user mute and volume.
• Chat the Discord way: grouped by author within seven minutes, day dividers, a "New messages" line, hover toolbar (quote, copy, message them, mute for me), Enter sends and Shift+Enter breaks. Send to the channel, a channel and its subtree, or one person. Inbound HTML goes through a whitelist. Images: paste, drop or pick; they're shrunk to the server's limit and sent as the well-formed XHTML murmur insists on for long messages. A refused message is marked "Not delivered".
• Eleven shared themes, each with light, dark, and system appearance; preview cards and smooth palette transitions. The same catalog supplies iOS, Android, the lock-screen widget, and the desktop picker. Settings and a diagnostics log live in the sheet.
• Screen share, which stock Mumble clients don't have: the share button in the dock opens the browser's picker (screen, window or tab, with audio where the browser offers it). People in your channel see a card and a green badge on your name; Watch opens the video in a third column with live resolution / fps / bitrate / codec, full screen and picture-in-picture. Video goes peer to peer over WebRTC (AV1 › VP9 › H.264 › VP8, up to 1080p); only the signaling crosses the Mumble server, in PluginDataTransmission. A TURN server can be set in Settings for networks that block direct connections. Protocol: docs/screen-share.md.

Layout

Field numbers match the Swift and Kotlin protocol implementations. Cross-platform tests exercise voice formats and screen signaling against the shared test server.

Testing

Nothing here talks to a real server. test/fake-server.mjs is a small Mumble server (TLS, handshake, roster, text, voice relay, plugin data with murmur's limits) and test/browser.mjs drives headless Chromium over the DevTools protocol — both with Node's built-ins only.

node scripts/generate-themes.mjs --check           # shared palette freshness + foreground contrast
node --test web/test/appearance.test.mjs            # all themes, persistence, reduced motion, responsive layouts, picker
node web/test/webcodecs.test.mjs                    # does this Chromium do Opus the way we assume?
node --test web/test/bridge.test.mjs                # bridge access controls, malformed input, certificate trust
node --test web/test/peer-certificate.test.mjs      # CA/hostname validation and certificate pin precedence
node --test web/test/certificate.test.mjs           # browser certificate approval, decline, and changed pins
node --test web/test/persistence.test.mjs           # settings survive browser + bridge restarts on a fixed port
node --test web/test/native.test.mjs                # macOS: native reconnect and exact loss-concealment duration
node web/test/e2e.test.mjs                          # two tabs, voice both ways over UDP, chat, images, reconnect
FAKE_VERSION=1.4.287 node web/test/e2e.test.mjs     # same, legacy voice format
FAKE_UDP=0 node web/test/e2e.test.mjs               # same with UDP blocked: voice must stay on the TCP tunnel
node web/test/ocb2.test.mjs                         # the cipher against Mumble's test vectors, loss/replay/resync rules
node web/test/share.test.mjs                        # screen share between two tabs, WebRTC + signaling
node web/test/signal.test.mjs                       # the plugin-message fragment codec, in Node
node web/test/dsp.test.mjs                          # the spectral suppressor: FFT, SNR gain, click ducking, block-size independence
node web/test/rnnoise.test.mjs                      # the RNNoise wasm: loads without imports, −22 dB on noise, voice kept, speed
node web/test/voice.test.mjs                        # voice packet codec round-trips and missing-packet arithmetic, in Node
node web/test/jitter.test.mjs                       # the playout buffer's grow/trim policy: quiet calls, near misses, damping
node web/test/quality.test.mjs                      # a tone through two tabs: SNR, clicks, dropouts, underruns; FAKE_LOSS=0.05 FAKE_JITTER=40 to impair
node web/test/fake-server.mjs                       # keep one running to click around against
node web/probe.mjs <host> [port] [username]         # handshake against any server, read-only

SHOTS=dir saves screenshots from the end-to-end run; VERBOSE=1 shows page console output.

Run timing-sensitive audio quality checks without concurrent emulator builds or other audio benchmarks. Host scheduling delays can change the adaptive jitter buffer even on loopback.