The Fortress, our On-Premise offer

Recapro on your own machine, no public port

The Fortress installs Recapro on a dedicated server, at your site. Transcription, summary and sync run on that single machine. Your meeting content never leaves your perimeter, and an air-gap mode cuts every outbound connection.

See the architecture The Fortress offer

Recapro On-Premise, in one definition

Recapro On-Premise is a full deployment of the platform on a dedicated server installed in your offices or datacenter. Database, API, workers, sync engine and AI models all run on this single machine, a 128 GB Jetson AGX Thor. No port is exposed to the internet: access happens from your local network. Meeting content does not leave it, and the air-gap mode allows fully offline operation.

One machine, everything inside

Nothing is offloaded to a cloud. The entire application stack and the inference models run on the server installed at your site.

The dedicated server

A Jetson AGX Thor with 128 GB of unified memory, sitting inside your walls. It hosts the whole platform, from the database to the AI models, without depending on an external resource to run.

What runs on it

Database API Workers PowerSync sync vLLM (LLM) Whisper (STT)

Listening surface: the local network

Four services listen on the local network. None is published to the internet. You reach them from your LAN, or through your own VPN.

  • :8000 API gateway (Kong)
  • :8080 Sync (PowerSync)
  • :8001 Transcription (Whisper)
  • :8002 Generation (vLLM)

None of these ports is reachable from outside your network. There is no public entry point to guard, and no attack surface exposed to the internet.

What comes in, what goes out

Two directions to keep apart: what the machine accepts as input, and what it sends outward. Meeting content appears in neither.

Inbound

Local network only

  • The mobile, desktop and web apps connect to the machine from your local network.
  • No port is open to the internet, so there is no inbound access from outside.
  • Remote access, if you want it, goes through your own VPN.

Outbound, standard mode

Metrics and licence

  • Heartbeats to api.recapro.ai signal that the machine is alive.
  • Machine telemetry flows to France-Nuage, protected by Cloudflare Access.
  • A Cloudflare tunnel enables supervision; images and models are downloaded on first boot.
  • The fleet agent reports machine metrics only: CPU, memory, disk, uptime.

Air-gap

Zero outbound connection

  • The machine runs fully offline, with no connection to the outside.
  • Installation and updates go through physical media.
  • No telemetry, no heartbeat, no tunnel: nothing leaves.

The point: In both modes, your meeting content never leaves the machine. In standard mode, only technical metrics and licence state travel, never a word of your conversations.

Inference runs at your site

Transcribing and summarizing means reading text in clear. That reading happens on your machine, by your models.

No third-party model

Transcription, summary and speaker identification are produced by models running on the server. No excerpt is sent to OpenAI, Anthropic or Mistral.

The AI stack, local

whisper.cpp for transcription, vLLM for generation, pyannote for speaker separation. All of it runs on the Jetson AGX Thor, with no external call.

Nobody reads your meetings

Since inference stays on the machine, no provider has access to the content of your conversations. Encryption in transit becomes secondary when nothing is in transit.

Why hosting alone is not enough: our breakdown of sovereignty at inference details the moment a model reads your data in clear.

You keep control

Deploying at your site does not lock you in. The hardware is dedicated to you, the licence is explicit, and nothing binds you technically.

Explicit licence

A licence switch governs the service. A suspended licence stops the AI containers. The mechanism is clear and documented, with no hidden dependency.

Reversibility

The pricing grid is public, the models are open-weights, and the hardware is dedicated to your deployment. You are bound neither by an opaque quote nor by a technology only you could run.

Dedicated hardware, inside your walls

The server is installed at your site, in your offices or datacenter. It is your physical perimeter, not a "private cloud" left with a third-party host.

On-Premise

The Fortress

The Fortress is Recapro’s On-Premise offer: a dedicated server installed at your site, the whole platform in a closed loop, and fully local inference. It is meant for organizations whose conversations cannot leave their walls.

  • Dedicated server installed at your site
  • Access via local network or VPN only
  • 100% local inference, no third-party LLM
  • Air-gap mode, no outbound connection
  • Public pricing grid
  • Open-weights models, dedicated hardware
See Fortress pricing Contact us

The Fortress price is published in the pricing grid, with no closed quote to go through.

Go further

On-premise is part of a broader take on sovereignty. Here are the pages that detail it.

Cloud vs On-Premise Where your data is stored depending on the deployment model, and what each one implies. AI sovereignty & inference Why "hosted in Europe" does not put your data out of reach during inference. Sovereign notetaker The four criteria that define the category, and the tool that meets them. Security and transparency Our architecture, our subprocessors by name, and what we are (or are not) certified for.

Recapro On-Premise, your questions

01

What is Recapro On-Premise?
It is a full deployment of Recapro on a dedicated server installed at your site. The database, API, workers, sync and AI models run on this single machine, a 128 GB Jetson AGX Thor. No port is exposed to the internet and your meeting content never leaves the server.
02

Is any port open to the internet?
No. Four services listen on your local network: the API gateway on port 8000, sync on 8080, transcription on 8001 and generation on 8002. None is published to the internet. Access happens from your LAN, or through your own VPN.
03

What leaves the machine in standard mode?
In standard mode, the machine emits heartbeats to api.recapro.ai, machine telemetry to France-Nuage protected by Cloudflare Access, and maintains a Cloudflare supervision tunnel; images and models are downloaded on first boot. This telemetry contains machine metrics only: CPU, memory, disk and uptime. Meeting content is never part of it.
04

Does air-gap mode really cut everything?
Yes. In air-gap, the machine runs fully offline: no outbound connection, no telemetry, no heartbeat, no tunnel. Installation and updates go through physical media.
05

Does the AI call an external service like ChatGPT?
No. Transcription (whisper.cpp), generation (vLLM) and speaker separation (pyannote) run on the server installed at your site. No meeting excerpt is sent to a third-party model, whether OpenAI, Anthropic or Mistral.
06

What happens if the licence is suspended?
A licence switch governs the service. If the licence is suspended, the AI containers stop. The mechanism is explicit and documented; it is not a hidden dependency.
07

What hardware does Recapro On-Premise run on?
On a dedicated server, a Jetson AGX Thor with 128 GB of unified memory, installed in your offices or datacenter. It hosts the full platform and the inference models, without depending on a cloud resource to run.
08

How do we leave if we decide to?
Reversibility is built in: a public pricing grid rather than an opaque quote, open-weights models that you are not alone in being able to run, and hardware dedicated to your deployment. Nothing binds you technically to a single vendor.

Your meetings, on your own machine

The Fortress installs Recapro at your site, in a closed loop, with fully local inference and an air-gap mode. Your conversations never leave your perimeter.

See plans and pricing Compare Cloud and On-Premise