Case study

JSONMeet. Slack, Meet and TeamSpeak in one, on your own server.

Case Study · JSON Crew in-house product · Self-hosted messenger

JSONMeet. Slack, Meet and TeamSpeak
in one, on your own server.

4 → 1four tools replaced by one
8 weeks.from design to production
4K/60screen sharing
JSONMeet In-house product 1 serwer
Product
JSONMeet: a company messenger with voice channels, chat, client meetings and an AI prompter. Built by JSON Crew, used every day by the JSON Crew team.
Challenge
  • Team split across 4 tools: voice, chat, meetings, recordings
  • Sales call recordings in someone else's cloud
  • Clients should join from a link, with no account or installation
  • AI should help during the call, not after it
Solution
  • Voice and text channels, DMs, search
  • Meetings via link with a waiting room and recording
  • HQ presenter with no server-side recompression
  • AI prompter: live transcription and suggestions
Result
  • One server in Warsaw, ~10 ms ping from Poland
  • Zero per-user subscriptions
  • Recordings and transcripts on our side, with retention
  • Desktop app with global push-to-talk

Four tools, four subscriptions, zero control over recordings.

Our team works remotely, and we treat sales conversations as data we don't hand over to anyone else. All day we sat in a TeamSpeak voice channel, wrote on Slack and Discord, and scheduled client meetings in Google Meet. Each tool did its part well. The problem was the sum.

Four logins, four sets of notifications, four places where decisions got lost. Sales call recordings ended up in the provider's cloud, on its terms and with its retention. We wanted to transcribe them, analyze them and learn from them. Every such attempt meant an export, an upload to yet another tool and yet another data processing agreement.

The biggest gap was in samej rozmowie. A salesperson hears a client's objection and has a few seconds to respond. All the wiedza firmy o tej obiekcji (what worked, what to avoid, the lead's context from the CRM) sits in documents nobody opens during the call. AI that summarizes the meeting after the fact changes nothing here.

We set the requirements: one app, on our server, the client joins from a link without an account, screen sharing quality good enough to present interfaces in 4K, recording with a clear banner for all participants. And a prompter that listens to the call and coaches the salesperson live, invisible to the client.

What we built. JSONMeet's four modules.

Module 01

Messenger: channels, chat, DMs

A Discord-style layout: workspaces, text and voice channels in categories, a member panel with live statuses. Persistent chat with mentions, files, reactions and replies. Direct messages. Full-text search across channels and your own conversations. All over a single WebSocket connection: messages, presence, “typing…”, unread counters.

Module 02

Client meetings and recording

A room via link. A team member joins right away, a client's guest lands in a waiting room and is admitted with one click. The guest doesn't touch the media server until the host approves. Server-side recording to MP4 with a “Recording” banner visible to everyone, a list of recordings with retention and automatic deletion.

Module 03

HQ Presenter and screen sharing up to 4K

In the browser: presets 1080p60, 1440p60, 1440p30 “sharp text” and 4K30, switchable during the call. For studio presentations, a separate desktop app: OBS or GStreamer encodes the video in hardware and sends it via the WHIP protocol straight to the media server, which forwards the stream without recompression. The server uses almost no CPU, and viewers see exactly what came out of the presenter's graphics card.

Module 04

AI prompter for salespeople

The live call transcript goes to a language model with the company playbook: a matrix of 26 objections, price negotiation rules, a company brief. The prompter shows the salesperson the detected objection, a ready response, questions to ask and a readiness-to-close score on a scale of 0 to 100. Before the call, you select a lead from the CRM and the prompter gets its context. The client never sees this panel. After the call: analysis, scorecard, search across transcripts.

Under the hood

Frontend in React 18 with Vite and Tailwind CSS 4. Backend in Node.js on Fastify 5 with Prisma 6 and PostgreSQL 17, Redis 7 for sessions, presence and queues. Media is handled by a self-hosted open-source LiveKit server: SFU for audio and video, Egress for recording, Ingress for receiving the WHIP stream from the presenter. The camera runs in VP9 with three quality layers, the screen in H.264 simulcast, so a weaker viewer gets a lighter layer without degrading the picture for others.

Noise suppression (RNNoise) and virtual background (MediaPipe) run locally in the participant's browser. Behind corporate firewalls, a built-in TURN over TLS and a fallback TCP transport save the day. The Electron desktop app connects to the company's instance and receives every interface update automatically; it adds a tray icon, a global push-to-talk shortcut, an unread counter on the icon and deep links to meetings.

Transcription and the prompter's language model run on our own GPU servers, connected to the application server over a private network. Production is a single VPS at OVHcloud in Warsaw (8 vCPU, 16 GB RAM) with services under systemd, no Docker. Deployment: an archive from git, build on the server, migrations, service restart.

FrontendReact 18, Vite, Tailwind CSS 4, komponenty LiveKit
BackendNode.js, Fastify 5, Prisma 6, PostgreSQL 17, Redis 7
MediaLiveKit OSS (SFU, Egress, Ingress), VP9 SVC, H.264 simulcast, wbudowany TURN
PrezenterOBS or GStreamer, WHIP, hardware encoding with no recompression on the server
DesktopElectron, global push-to-talk, deep links, RNNoise and MediaPipe running locally
AI and hostingTranscription and language model on our own GPUs, a single OVHcloud Warsaw VPS under systemd

Before and after. Same team, one tool.

❌ Przed JSONMeet
  • 4 tools: TeamSpeak, Slack/Discord, Google Meet, osobne nagrywanie
  • Call recordings in the provider's cloud, retention on its terms
  • The client installs an app or logs in to join a meeting
  • Objection knowledge in documents, outside the call
  • Screen shared at whatever quality the tool sees fit
  • Per-user subscription in each tool
✅ With JSONMeet
  • 1 aplikacja: voice channels, chat, DMs, meetings, recordings
  • Recordings and transcripts on our own server, retention set per plan
  • The client joins from a link in the browser, through a waiting room
  • The prompter gives hints during the call, the client can't see it
  • Presets up to 4K30 and 1440p60, presenter without recompression
  • A single VPS, fixed cost regardless of headcount

Liczby projektu.

8 weeks. from design to production
99 commits
48 tys. linii TypeScript
frontend, backend, desktop, prezenter
~30 participants per room
test: 10 senders < 50% CPU
26 objections in the prompter playbook
plus negocjacje cenowe

Plus: 22 database migrations, 26 backend test files, 4 plans with enforced limits (FREE, STARTER, PRO, ENTERPRISE), 2 parallel recordings on the server, deep links jsonmeet://join/<kod> from the desktop app.

It wasn't about having our own Meet on principle. It was that sales conversations are the most valuable data in a company, and we were handing them to someone else. Now the recording, the transcript and the suggestions are ours. And they work during the call, not after it.

Szymon Stelmach, JSON Crew, autor JSONMeet

How the build went. Eight weeks.

Week
1

Design and foundation

Design document: requirements, risks, data model, phase plan. Foundation: multiple isolated companies on one instance, accounts and sessions, roles with permissions resolved at login, invitations via a copyable link. Media server set up on a VPS with built-in TURN.

Week
2

Spotkania pod linkiem

Room with a code, a server-side guest waiting room, spotlight and grid, screen sharing with presets. Recording of the full room composition to MP4 with a banner for participants and a list of recordings with retention.

Week
3

Prezenter HQ

A separate desktop app for the presenter: OBS or GStreamer encodes in hardware, the stream goes via WHIP to Ingress and onward with no recompression. Decision: WHIP instead of RTMP, because RTMP forces re-encoding on the server.

Week
4–5

Komunikator

Fixed app layout: workspace bar, channels in categories, content, member panel. Persistent chat with mentions, files, reactions and replies, DMs, search. Voice channels that work independently of navigation: you click around the app and the call continues. Push-to-talk without renegotiating the connection.

Week
6

Aplikacja desktopowa

Electron connecting to the company's instance: tray icon, global push-to-talk shortcut, unread counter, custom window picker for sharing, deep links to meetings. First launch means entering the server address and accepting the certificate; interface updates come from the server.

Week
7

Transcription and prompter

Live transcription on our own GPU. A prompter with a 26-objection playbook, a company brief and lead context pulled from the CRM. Visible only to the team, configured per workspace. A single entry point to the language model for all AI features, so changing the model is just a configuration change.

Week
8

Conversation analysis and production

Post-talk analysis with speaker recognition, scorecard, signals, and post-transcription search. Image and sound hygiene: noise reduction, virtual background, automatic lowering of the resolution when the USB camera drops to several frames. Implementation on one VPS in Warsaw under systemd, without Docker.

Self-hosted messaging for business: when your own server makes sense and when it doesn't

An article for business owners and sales leaders considering moving chat, voice calls, and client meetings onto their own infrastructure. We explain when a self-hosted business messenger is a sensible choice, what needs to be designed into it, and which risks to plan for from day one.

What a self-hosted messenger is and how it differs from Slack or Meet

A self-hosted communication platform is an app for chat, calls, and video conferencing running on a server the company manages (its own or a rented VPS). Messages, recordings, and transcripts stay in its own infrastructure, not in a SaaS vendor’s cloud.

The difference isn't just about where data is stored. With your own deployment, the company decides on recording retention, who has access to recordings, and which tools (e.g., AI models) may process them. With off-the-shelf services, the provider makes those decisions in its terms of service.

A self-hosted business messenger doesn't have to mean writing everything from scratch. A sensible implementation consists of mature open-source components (media server, database, queues) and a custom application layer that addresses the team's specific needs.

An alternative to Slack and Google Meet: when it's worth considering

For most companies, off-the-shelf tools are enough. A custom solution starts to make sense in three situations, which often occur together:

  • client conversations are data you don't want to hand over (recordings, transcripts, sales notes),
  • the team uses several tools in parallel and pays for each one per user,
  • you want AI to work with your playbook and CRM data during the call, not just summarize it afterward.

In the implementation described above, the team worked with four tools: TeamSpeak for voice, Slack and Discord for chat, Google Meet for client meetings, and separate recording software. Four logins, four sets of notifications, and decisions got lost between them. JSONMeet replaced them with a single app.

IssueOff-the-shelf SaaS toolsSelf-hosted messenger
Recordings and transcriptsin the provider's cloud, retention on its termson the company's server, with retention set by you
Billingper-user subscription in every toolserver and maintenance cost, independent of headcount
AI during the calllimited to the vendor's featuresany model, your own playbook, data from the CRM
Maintenanceon the vendor's sidehandled by the company or a technical partner
Launchimmediatelydesign and implementation (here: 8 weeks)

Recording calls on your own server: requirements worth writing down

Before the first line of code is written, it’s worth writing down the business requirements. In the project described, they were: one app, a self-hosted server, clients join from a link without an account or installation, screen sharing good enough to show an interface in 4K, and recording with a clear notice to all participants.

This leads to specific features. A guest lands in a waiting room and doesn't connect to the media server until the host lets them in. Recording happens server-side to an MP4 file, and a “Recording” banner is visible throughout the call. The recordings list has a retention policy with automatic deletion, configurable from 7 days to unlimited.

This last point matters for GDPR. If the recordings are stored with you, you don't need another data processing agreement every time you want to transcribe or analyze them with a different tool.

Architecture: WebRTC, media server, and client firewalls

The core of every messaging app with video is an SFU media server, which receives streams from participants and forwards them on. In JSONMeet, this is a self-hosted open-source LiveKit: SFU for audio and video, Egress for recording, and Ingress for receiving the stream from the presenter’s app.

The most common risk with self-hosting WebRTC is a corporate firewall on the client's side. It needs to be planned for from the start. Here, the server has built-in TURN over TLS on port 5349 and a fallback TCP transport, and for the most restrictive networks there's a second IP address with TURN on port 443, which looks like ordinary HTTPS traffic to the firewall.

The second decision is video quality. The camera goes out in VP9 with three quality layers and screen sharing in H.264 simulcast, so a recipient with a weaker connection gets a lighter layer without degrading the picture for others. For studio presentations, a separate app encodes the video in hardware and sends it via WHIP with no server-side re-encoding. Presets go up to 4K30 and 1440p60.

AI prompter for sales reps: AI during the call, not after it

A meeting summary after the fact doesn’t help much a sales rep who hears an objection and has a few seconds to respond. That’s why, in the implementation described, the call transcript is fed live to a language model together with the company’s playbook: a matrix of 26 objections, price negotiation rules, and a company brief.

Before the call, the sales rep selects a lead from the CRM, and the prompter receives its context. During the call, it shows the detected objection, a suggested response, questions to ask, and a readiness-to-close score on a scale of 0 to 100. The client doesn't see this panel. After the call, analysis, a scorecard, and transcript search are available.

An important detail: transcription and the model run on in-house GPU servers connected to the application server over a private network. Conversation content never goes to an external AI provider. A single entry point to the model for all AI features means that switching models is just a configuration change.

Maintaining a self-hosted messenger without a DevOps team

A common question is: who will maintain it? The answer depends on how simply you design production. JSONMeet runs on a single VPS in Warsaw (8 vCPU, 16 GB RAM) with services under systemd, without Docker or an orchestrator. Deployment is a single script: archive from git, build, database migrations, service restart.

In testing, this server handled a room with about 30 participants, 10 of whom were sharing video, at under 50% CPU load. Ping from Poland is about 10 ms. The Electron desktop app pulls interface updates from the server, so there's no need to update it manually on each computer.

For companies that don't want to run a server themselves, a sensible option is an instance maintained by a technology partner. The data still stays in dedicated infrastructure, and responsibility for updates and security is set out in the contract.

Before deciding, it's worth calculating three things: how many people and tools the change will affect, how valuable call recordings are to you, and whether AI should work on your data. If the answers point toward control over your data, a self-hosted messenger for your company is worth considering.

Three questions we hear when we show JSONMeet.

Why own a messenger when Google Meet and Slack are ready and cheap?

For most companies, ready-made tools are enough. Own makes sense when customer conversations are data you don't want to give away (recordings, transcripts, sales notes), when you pay per user for four tools at once, or when you want an AI that runs on your playbook and your CRM during the conversation, not after it. JSONMeet was created for these three reasons at the same time.

WebRTC behind the client's company firewall will not work.

This is a real risk and we have planned it from day one. The media server has a built-in TLS TURN on port 5349 and emergency transport over TCP. In the scenario of the most restrictive client network, a second IP address is provided from the TURN on port 443, which looks like regular HTTPS traffic to the firewall. The guest enters from the browser, without installation.

Who will maintain it? We don't have a DevOps team.

Production includes one VPS and six services under the systemd: media server, API, reverse proxy, Redis and containers for recording and receiving the stream. The implementation is one script: archive from git, build, migrations, restart. There is no orchestrator, no Docker Compose. The desktop application updates the interface itself, from the server. For external customers, we offer JSONMeet as an instance maintained by us.

Does the AI souffler work during the conversation or only after it?

In progress. The transcript of the conversation flies live to the language model running on our GPUs, with a playbook of 26 objections and a lead context downloaded from the CRM. Sufler detects objections in the customer's statement, prompts ready answers and shows the readiness to close indicator on a scale of 0–100. The customer cannot see the airfoil panel. Analysis after the fact (scorecard, search after transcripts) is also available, but it does not replace the live souffler.

Are the call recordings left only on our server?

Yes, that was the whole game. The recording for MP4 goes to the volume connected to your VPS. Retention is configured per plan (from 7 days to no limit). The customer hears the message about the recording and sees the „Recording” banner throughout the conversation. Neither recordings nor transcripts leave your infrastructure — there is no external cloud provider.

How much does JSONMeet self-hosting cost compared to Slack + Meet + TeamSpeak fees?

It depends on the team, but the return on investment in JSONMeet typically starts with 8–15 people when the per-user costs in the finished tools start to add up. The specific bill depends on which modules you need — convert with the calculator below. We also offer an instance managed by us if you do not want your own DevOps.

Does JSONMeet replace CRM and sales tools?

No. JSONMeet is a communication layer that integrates with your CRM (Pipedrive, HubSpot, Salesforce, own API). The lead context from the CRM feeds the souvenir before the conversation, and the result of the conversation (transcription, scorecard, note) returns to the CRM after its completion. The sales call is in JSONMeet, the lead data is in CRM — and the two systems talk to each other after rest.

Calculate the cost of implementation

Select the scope of JSONMeet implementation in the company. The one-time price includes installation on your server, configuration of modules and team training. Your monthly subscription covers hosting, security patches, and updates.

Step 1JSONMeet modulesselect the set of features you actually use
Quantity:1
Quantity:1
Quantity:1
Quantity:1
Step 2Integrationsso that the souffler gets the context of the lead, and the recordings go where they belong
Quantity:1
Quantity:1
Quantity:1
Quantity:1
Step 3Implementation and trainingsetup of your server and training of the sales team
Razem jednorazowoPLN 0

Want client calls on your own server, with AI on your side?

We'll show you JSONMeet live: you'll join as a guest through the waiting room and see the prompter from the salesperson's side. 30 minutes, no sales pitch.

Resources · knowledge and evidence

Check before we talk

Services associated with this deployment. Take a look at the offer, see how we approach similar problems in other customers.

Next step

Did you like this article?

If after reading this you feel your company also has processes worth rebuilding, book a free call. Together we will check where sales are really leaking and what makes sense to implement first.

Schedule a call