Voice on a real SFU
Voice runs through a selective forwarding unit, not a peer-to-peer mesh. Each person sends their audio once no matter how many are listening, so a room of twelve sounds like a room of two.
In active development · v0.1.0
Chat that stays yours.
Vocalkeep is a voice and text chat platform you can run yourself. Servers, channels and real-time voice on infrastructure you control — with a desktop app and a browser tab that share one account.
Also on macOS and Linux · Free and open source · Hosted sign-up is coming
Vocalkeep — Vocalkeep HQ
RG
DE
ML
+
Vocalkeep HQ⌄
Text
#welcome
#general
#builds3
#bug-reports
Voice
Lounge3
Focus Room
Voice connected
Lounge · 24 ms
#general
KE
kestrel11:04
pushed the read-state fix — unread counts settle instantly now
NI
nine11:06
confirmed on the Windows build. same account, no re-login.
OR
ora11:09
voice held 12 people last night with zero drops
SFU was barely awake
tam is typing…
Message #general
Members — 5
KE
NI
OR
TA
WR
Features
Everything you expect. None of what you didn't ask for.
No ads, no engagement mechanics, no telemetry pipeline pointed at your conversations. Just the chat app, built well.
Servers, channels, conversation
Create a server, carve it into text and voice channels, invite people in. Messages arrive over a persistent WebSocket — no polling, no refresh, no waiting to find out someone replied.
Unread that is never wrong
Unread state is computed from what you have actually read, never stored as a counter that can drift. A dedicated service keeps every badge in the app correct in one request at launch.
Accounts done properly
Passwords hashed with Argon2. Short-lived access tokens in memory, refresh tokens in an httpOnly cookie that JavaScript cannot read — rotated on every use, so a stolen one is detectable.
One app, two places
The Windows app and the browser app are the same client, built twice. Same account, same servers, same unread state — sign in wherever you are and pick up mid-sentence.
Yours to run
One command brings up the whole stack — services, database, object storage, voice. Nothing phones home, nothing needs a cloud account, and the development keys ship in the repo so day one works.
Self-hosting
Run the whole thing on your own box.
Vocalkeep is not a hosted service with a self-host option bolted on. Local is the primary way it runs — the same compose file that a contributor uses on day one is the one that runs it for your group.
- Postgres for everything durable, Redis for everything ephemeral
- S3-compatible object storage for uploads — MinIO locally, anything you like in production
- Voice runs on a self-hosted SFU. No third-party media servers, no per-minute billing
- Development credentials ship in the repository on purpose: a stack that needs secrets before it boots is a stack nobody runs on their first day
bash
# everything: datastores, services, voice
make dev
# on Windows, where make is not a given
.\make.ps1 dev
# services come up on
api localhost:5000
gateway localhost:5001
read localhost:5002
voice localhost:7880Under the hood
Five services, each in the language that fits it.
Not polyglot for its own sake. Each piece has a different failure mode and a different shape of load, and each one is written in the language that makes its particular problem boring.
Client
TypeScript · React
One codebase, built twice — once as a browser bundle, once inside an Electron shell. The desktop app is a shell, not a fork.
API
Elixir · Phoenix
Auth, accounts, servers, channels. Plain versioned JSON over HTTP, so anything can talk to it.
Gateway
Elixir · OTP
Tens of thousands of long-lived connections, each supervised and restartable on its own. That is what the BEAM was built for.
Read states
Rust
The hottest read-write path in a chat app — every mention writes one, every channel click acks one. It gets its own service and its own budget.
Voice
LiveKit SFU
Self-hosted, and deliberately not written here. Building an SFU is a project, not a feature.
Backed by PostgreSQL 17, Redis, and S3-compatible object storage.
Get Vocalkeep
Install it, or just open a tab.
The desktop app adds native notifications, a tray icon and push-to-talk that works when the window isn't focused. Everything else is identical to the web app — same account, same servers.
Builds aren't published yet. Vocalkeep is still in active development — these links go live with the first tagged release. In the meantime you can run it from source.
FAQ
Straight answers.
Including the ones where the answer is "not yet".
Does the web version work with the Windows app?
Yes — they are the same client. One account, one set of servers, one unread state. Sign in on the web, close the tab, open the desktop app, and you are exactly where you left off.
Do I have to self-host to use it?
Self-hosting is the model Vocalkeep is built around: you run the server, you hold the data. If you are joining someone else’s server you just point the app at it and sign in — you do not need to run anything yourself.
Is it really free?
Yes. It is open source and there is no paid tier, no seat count and no feature held back. The only cost is whatever you spend running it.
Is voice end-to-end encrypted?
Not today. Voice is encrypted in transit with DTLS-SRTP between each participant and the server, which is the standard WebRTC guarantee — but the server forwards decrypted media, so it is not end-to-end. If you host the server yourself, that server is yours.
Can I import my Discord servers?
No, and there are no plans to. Discord does not offer an export path for server history, and anything that claims otherwise is scraping against their terms.
How finished is this?
Accounts, servers, channels, real-time messaging, unread state and voice all work today. It is not yet at a tagged release — there are no published builds, and things will move around before there are.