Isolated storage
Your documents, mail index and databases sit in a layer with no route to the internet
Zokai runs wherever you decide: your own hardware, your data centre, the infrastructure provider you trust, or dedicated servers we operate for you in Germany. The AI works through a defined set of capabilities you control, never by clicking through screens. Calculations run in code you can inspect. Every change, run and approval is recorded in a ledger you can verify yourself.
Early Access. Zokai Station is available to selected organisations through commercial pilots. Scope, hosting and support are agreed individually for each pilot.
This page describes how Zokai is designed. The detailed architecture is available under NDA.
Your documents, mail index and databases sit in a layer with no route to the internet
API keys and account tokens are stored as encrypted local secrets, never in configuration files, and never sent to us
Zokai doesn't phone home. Outbound traffic only goes to services you connect
No shared databases or infrastructure between customers, on your premises or hosted
Every component runs with minimal rights and fixed resource limits
Your hardware, your data centre, any provider you pick, or managed by us in Germany. No US-parent provider in our software stack; the hosting jurisdiction is your decision
You choose the mode for each session. Each mode has its own fixed set of capabilities, so switching mode changes what the AI can reach, not just what it's told.
The AI can use
Your documents, spreadsheets, notebooks, mail, calendar, tasks and workflows, with a local model
What can leave your infrastructure
Nothing
The AI can use
Everything in Air-gapped, plus web research, YouTube transcripts and bookmarks
What can leave your infrastructure
Prompts to the cloud model you select. Delegated tasks are masked, and your files aren't uploaded wholesale
The AI can use
Everything in Private hybrid, plus GitHub and autonomous developer agents
What can leave your infrastructure
Cloud reasoning, for work you decide is suitable
Search all mail by meaning, draft replies, find free slots, book meetings. Gmail and Google Calendar today; Microsoft 365 and IMAP in development
Draft and edit proposals and contracts, cite PDFs, show who wrote what
Read and write Excel and CSV, insert real formulas, audit ledgers
Run exact Python and SQL analyses on your data
One question across mail, PDFs, spreadsheets, notes and files
Multi-source research, YouTube transcripts, bookmark collections (Private hybrid and Frontier)
Scheduled and event-triggered workflows with approval steps
Repository inspection and developer agents (Frontier)
Built on the open Model Context Protocol (MCP).
"When the monthly sales report arrives, analyse it, draft the board narrative, build the deck, and wait for my sign-off before anything is sent."
The numbers are computed by code, the same way on every run.
One step turns results into board-ready text, and the output is checked against a required format before the workflow continues.
The workflow pauses until a person approves.
Break when APIs change, charge per task, move data to third-party clouds
Fragile, expensive, can act without approval
Step-by-step, repeatable, approval gates, run on infrastructure you choose
One record for everything the people and the AI do.
Every create, edit, rename and delete, with time and source. Full content history is kept, so any earlier version can be restored
Every cell change with its previous value and time. Restoring creates a new entry instead of rewriting history
Every change to a workflow, every run, and every approval
Prompts and responses, kept as an append-only record
The assistant synthesizes executive priorities and books time blocks into the calendar on a single confirmation.
A single query retrieves project budgets, agreements, client emails, and scheduled meetings across connected sources.
The assistant restructures a client proposal; the document history shows every change and where it came from.
Search across multi-format workspace files with built-in previewers.
The assistant audits spreadsheet data: generating formulas, policy disallowance checks, and audit aggregates directly in the sheet.
End-to-end workflow execution: data analysis, structured narrative, authority gate sign-off, and presentation generation.
During Early Access, every deployment option is offered as a commercial pilot. Hosting, scope and support are agreed for each pilot.
Maximum control; nothing has to leave the building
Runs on the infrastructure and in the jurisdiction you already trust
Dedicated single-tenant servers, operated for you
Zokai is containerised and runs on standard Linux servers. Air-gapped mode runs current open models (around 30B parameters) locally. A typical setup is a server with an NVIDIA GPU with at least 24 GB of memory; 48 GB is recommended for teams and long documents. Apple Silicon workstations with 32 GB or more are enough for individuals and small teams. Hardware is sized as part of each pilot.
Architecture, data flow per mode, record design, component list, test evidence, security questionnaire.
We switch modes in front of you, trace an AI edit into the record, and run the verification live.
The free, open-source (AGPL-3.0) Community Edition is a developer workstation with a browser IDE, AI coding agent, semantic search and integrations. It doesn't include the workspaces, automation or audit trail.