TIDESa civilization in progress

PaperFern

How a private team chat is put together: the design choices behind Fern

Abstract. A design note on Fern, the team chat app made by TIDES. It explains the main choices in the product, including peer-to-peer calls, admin control, on-screen privacy and deployment options, and the trade-off behind each one. Features that are planned but not yet released are marked as such.

Key points

  • Messages and call media have different needs, so they travel by different routes.
  • Calls connect peer-to-peer: audio, video and screen sharing go directly between participants and are encrypted in transit.
  • Signaling only arranges a call. It does not carry the call itself.
  • Admins control who is a member and who may post, because a work chat belongs to the organisation.
  • End-to-end encryption for messages, customer-held data and peer-to-peer messaging are planned and not yet released.

1. Scope of this note

Fern is a desktop app for a company's internal conversations: direct chats, private groups, a company-wide channel, voice and video calls, screen sharing and file sharing. This note describes how a product of this kind is put together and why Fern makes the choices it does.

It is a design note, not a test report. It contains no measurements. Where a capability is planned and not yet shipped, the text says so plainly.

2. Two kinds of traffic

A team chat moves two very different kinds of data. Messages and files are small or occasional, and they must be kept. A person who opens the app on Monday expects to see what was said on Friday. That calls for a server that stores the conversation and delivers it to each member when they come online.

Call media is the opposite. Audio and video are a heavy, continuous stream that is worthless a moment later. Nobody needs last Tuesday's video frames to be delivered late. What matters is that they arrive quickly.

Fern treats these two kinds of traffic differently. In the deployment available today, messages and files go through a hosted server, database and file store. Call media goes directly between the people on the call.

3. Why calls go peer-to-peer

In a peer-to-peer call, each participant's device sends audio and video straight to the other participants' devices. There is no media server in the middle receiving the stream and passing it on.

The first reason is delay. The shortest path between two laptops in the same office, or the same city, is the direct one. Sending the stream out to a distant server and back adds distance, and distance adds lag that people notice as talking over each other.

The second reason is exposure. A stream that never passes through a central media server is not sitting on that server to be seen, logged or stored. For a company discussing salaries, clients or legal matters, fewer places where a conversation exists is a real benefit.

The trade-off is that the work moves to the participants. In a direct design, each device sends its stream to every other participant, so the load on each computer and each internet connection grows as people join. That is a general property of direct connections and the reason they suit team-sized calls better than very large broadcasts. Fern's plans set a limit on call participants at each tier.

4. What signaling does and does not carry

Two devices cannot simply start sending video to each other. They first need to find each other and agree on terms. The exchange that arranges this is called signaling, and it does need a go-between that both devices can already reach.

In the standard approach to peer-to-peer calls, known as WebRTC, signaling carries a small set of things.

  • The invitation: who is calling whom, and whether the call was accepted or declined.
  • A session description: which audio and video formats each device supports.
  • Network candidates: the addresses at which each device might be reachable, so a direct path can be found through home and office routers.
  • Key fingerprints: a short value each side uses to check that the encrypted connection it sets up is really with the other participant.

Signaling does not carry the voice, the video or the shared screen. Once the direct connection is up, the media no longer goes through the go-between.

5. Encryption in transit

Fern calls are encrypted in transit. In the standard peer-to-peer design this is done with two protocols that work together. DTLS performs a handshake directly between the participants and produces the encryption keys. SRTP then uses those keys to encrypt every packet of audio and video.

The important detail is where the keys are made. They are negotiated between the participants' devices over the direct connection, not handed out by a central server. Someone watching the network between two participants sees encrypted packets, not a conversation.

This is a different thing from end-to-end encryption of stored messages and files. That is on Fern's roadmap and has not been released. Section 11 covers it.

6. Calls as part of the conversation

A call in Fern starts from any chat with the phone or camera button. The design reason is simple. The people you need to call are usually the people you were just writing to, so the call should begin from where the conversation is.

The call opens in its own window, and a compact call bar stays in the main app so you can keep chatting. People on a call often need to look something up, paste a link or answer someone else. A call that takes over the whole app forces a choice between talking and working.

During a call, a participant can share their screen live for walkthroughs and reviews, switch to a focus layout, pop a participant out into a separate view, or record the call. Screen sharing travels the same direct route as audio and video.

7. Why admins control posting and membership

A consumer chat app treats every user as the owner of their own account. A work chat is different. The workspace belongs to the organisation, and the organisation is responsible for who is in it and what is announced there.

Fern gives admins a set of controls for this. Each one answers a specific, ordinary problem.

  • Lock a chat so only admins post. A company-wide channel used for announcements stays readable. An important notice is not pushed off the screen by forty replies.
  • Slow mode. A busy group can be limited so each person posts less often, which calms a heated thread without silencing anyone.
  • Member management. When someone joins a team or leaves the company, their access changes in one place.
  • A command center for users, roles and sessions. Admins can see who has access and what role each person holds, and manage signed-in sessions.

8. The trade-off in admin control

Central control has a cost. Members have less freedom than they would in a personal app, and a careless admin can make a workspace feel policed.

The alternative is worse for a company. Without membership control, a person who has left the organisation keeps reading its conversations. Without posting control, the channel everyone is supposed to read becomes the one everyone mutes.

Fern also puts some of this control in each member's own hands. A person can sign out a single device, or use "Forget this device" so a computer they no longer use stops being a way into their account.

9. Privacy on the screen itself

Much of what leaks from a company does not leak through the network. It leaks through a screen. A laptop is open in a shared office. A notification pops up while someone is presenting to a client. A colleague stands behind a chair at the wrong moment.

Encryption does nothing about this, so Fern has features aimed at the screen.

  • Privacy mode frosts messages on screen, so the app can stay open without its contents being readable at a glance.
  • Notification previews can hide the sender or the content. You still learn that something arrived, without the message being displayed to the whole room.
  • Read receipts and typing indicators can be switched off, so a person can read and think without broadcasting it.

These are convenience trade-offs, which is why they are switches and not fixed behaviour. A frosted screen is safer and slower to read. A hidden preview is more private and less informative.

10. Files and themes

Fern lets people share documents, images and voice notes, up to 250 MB each on the Starter plan, with larger limits on higher plans. Every chat keeps its own Files, Media and Links library. The reasoning is that people remember where something was shared more easily than what it was called. "The contract Ayesha sent in the finance group" is a location, and the library makes that location searchable by eye.

Appearance is set per person: ten themes, nine accent colours, wallpapers, density, bubble shape and font options. This can look like decoration, but there is a practical argument. People spend hours a day in a work chat, on different screens and with different eyesight. One person wants a dark, compact layout on a small laptop. Another needs larger text and more spacing.

11. Three ways to deploy

Where the data lives is the biggest decision for any company adopting a chat tool. Fern describes three deployment models. They are at different stages, and the difference matters.

  • Fern Cloud (available today). Fern hosts the server, database and file storage. There is nothing to install beyond the app, and updates roll out automatically. The trade-off is that the company's chats and files sit on infrastructure it does not operate.
  • Your database, your storage (launching next, not yet released). The company connects its own database and storage so chats and files stay in infrastructure it owns. It is planned for the Enterprise plan. The trade-off is that the company takes on running and backing up that infrastructure.
  • Peer-to-peer (live for calls). Voice, video and screen sharing already connect directly between participants. Peer-to-peer messaging is planned and not yet released.

12. What is planned and not yet released

A design note should be clear about its edges. The following items are on Fern's roadmap. None of them should be assumed to exist in the app today.

  • End-to-end encryption for messages and files.
  • Encryption at rest with customer-managed keys.
  • Bring your own database and storage.
  • Peer-to-peer messaging.
  • Audit log export and retention policies.
  • Single sign-on.

Today's privacy position is this: calls are peer-to-peer and encrypted in transit, and the on-screen and admin controls described above are in the app. Messages and files are not yet end-to-end encrypted.

13. Summary

Fern's design follows from a few plain ideas. Send each kind of data by the route that suits it. Keep call media off central servers by connecting participants directly. Give the organisation control over membership and posting, and give each person control over their own screen, notifications and appearance.

Each choice has a cost, and this note has tried to name it alongside the benefit.

Learn to build this.

Four months, hands-on, in Lahore. Admissions are open.