Install from a single file
Ready containers, or a container-free option for policies that forbid them. No internet access needed.
For banks, healthcare, the public sector and anyone who cannot send call recordings outside. Calls, transcripts and AI answers stay with you.
Already a customer? Log in
Updated
Yes, Leksta has a separate on-premise edition for your own servers. Transcription, call analysis, the knowledge base and the language model run inside your network, and in air-gap mode the product never reaches the public internet. The edition has been validated on test machines, including Astra Linux SE 1.8, and is offered as a pilot.
The on-premise edition runs the same console: call log, transcripts, scorecard checks and Ask your calls. What changes is where the computing happens. Speech recognition and the language model run on your servers, so conversation text never goes to cloud AI services.

In air-gap mode the product checks its settings at startup and refuses to start if any address is public. Cloud AI, cloud PBX polling and outbound webhooks are switched off, and integrations may only target addresses in your private network.

Ready containers, or a container-free option for policies that forbid them. No internet access needed.
Install, recording and transcription of a real call validated on test machines, including closed software environment mode.
Databases, recordings and settings archived with a checksum. A damaged archive is rejected on restore.
A failed upgrade rolls back automatically and work continues on the previous version.
OIDC and Active Directory sign-in: the directory group sets the role. Validated on a test directory.
Dashboards, alerts and log search inside the perimeter. The support bundle excludes passwords, phone numbers and transcripts.
Services without extra privileges, configuration checked against CIS benchmarks, key rotation with one script.
Who changed what: PBX, knowledge base, access keys, roles and privacy settings.
Problem. Customer call recordings cannot go to external cloud services.
What the AI does. Call transcription and scoring run on the bank's servers, with integrations limited to the internal network.
Outcome. Quality control on every call without data leaving the perimeter.
Problem. Patient conversations contain sensitive information.
What the AI does. Local install, personal data masking and retention set by the clinic.
Outcome. Managers see booking call quality while patient data stays in the clinic.
Problem. A domestic OS and offline operation are required.
What the AI does. Install on Astra Linux SE 1.8 from an offline package, air-gap mode with no outbound traffic.
Outcome. AI analytics can be trialled in a pilot on an isolated network.
The edition is not yet in production use at customers, so we start with a pilot on your test environment.
OS, container policy, network restrictions, call volume and whether you need a local language model.
Speech recognition is the main cost. We size for your volume and mode: overnight batch or real-time transcription.
We install from a single file on your test server and connect the PBX over SIP inside the network.
Together with your security team we run a network capture and confirm no traffic leaves the network.
We compare transcription and analysis quality on your recordings and agree on the next stage.
Yes. In air-gap mode the product never reaches the public network and will not start if a public address appears in its settings. This was validated by network capture on an isolated test machine.
An air gap is a closed network that is physically or logically separated from the internet. Every component, AI included, runs inside it, and data never leaves the organisation's network.
Yes, it runs on Astra Linux SE 1.8. Install, upgrade, rollback and transcription of a real call were validated on clean test machines. We validate a specific customer image as part of the pilot.
No. We state only what has been validated: the product runs on Astra Linux SE 1.8 and its hardening follows CIS recommendations. We work through your security team's requirements in the pilot.
It depends on the number of agents and the transcription mode: overnight batch or real time. A local language model needs a server with a GPU and enough memory. We size it for your volume during the pilot.
Not entirely. Call classification matches the cloud, while goal detection is less accurate locally for now. That is why the pilot compares results on your own recordings.
Not yet: it has been validated on isolated test machines and the install archives are not signed yet. We offer a pilot on your test environment.
An upgrade arrives as a single offline file. A backup is taken first, and if the upgrade fails the system returns to the previous version automatically.
Tell us your security requirements and we will outline a pilot on your servers.