Skip to content
CiAConsultancy in ActionAI Operations for Real Businesses

Case study · 2026

CIA N.I.C.K.E.D. Office

Documents, spreadsheets, presentations, PDF and Markdown — with an assistant that writes into them, running on a model on your own computer.

A full office suite whose AI runs on your own machine

Industry
Capability build · Local-first AI
Engagement
Fork → de-cloud → rebrand → package → ship
Result
Documents, spreadsheets, presentations, PDF and Markdown — with an assistant that writes into them, running on a model on your own computer.

The challenge

Every office suite now has an assistant, and every one of those assistants is somebody else's server. That is a real problem for the businesses we work with: a solicitor drafting a client letter, an accountant with a spreadsheet of somebody's finances, a founder writing a board deck. The work is confidential, and "we don't train on your data" is a policy, not an architecture.

The obvious answer — run a local model — solves the wrong half. Local models are readily available; what is missing is an office suite that can actually use one. Existing suites either have no assistant at all, or route through a vendor account you cannot switch off.

Building a word processor, spreadsheet and presentation editor from nothing is a multi-year project. Rewriting the file formats alone would be. The realistic route was to start from something that already opened .docx, .xlsx and .pptx correctly — and then take the cloud out of it without breaking the parts that worked.

What we built

We forked an existing open-source suite under the Apache Licence 2.0 and replaced its provider layer with a single local endpoint. The vendor's cloud client, account system and search integration were deleted rather than disabled — a switched-off feature is a feature that can be switched back on by a config change.

Removing a cloud is harder than deleting the network calls, and this is the lesson worth carrying: the dangerous leftovers are the GUARDS and the PROMPTS. A guard that refuses an action and tells the model to use the removed cloud tool leaves no legal way to do the job. A system prompt that still spends five hundred words explaining a deleted feature is worse than one that never mentioned it, because a model resolves the contradiction by following the more specific instruction. Each of those failed silently and plausibly — the app looked fine and produced nothing.

Verification therefore had to be behavioural, not unit-level. Every claim is checked by driving the packaged application end to end and asserting on the SAVED file — a .pptx unzipped and read, not a canvas scraped — because the AI panel shares the DOM with the document and scraping the page passes on the assistant's own description of what it did.

The suite was then rebranded end to end across nineteen locales, given its own identity and bundle id so it installs alongside the original rather than over it, and packaged for macOS, Windows and Linux.

The outcome

Five editors in one application — documents, spreadsheets, presentations, PDF and Markdown — opening and saving the .docx, .xlsx and .pptx files people already have.

The assistant writes into the document rather than into a chat box: ask for a memo and the memo appears on the page; ask for a five-page deck and the pages assemble on the canvas.

Mind maps that most suites cannot draw at all: a topic broken into branches with connectors that stay attached when you drag a node, generated from a prompt or from the presentation already open.

Free for individuals, permanently. Organisations pay a subscription, required above one hundred people. The software itself remains Apache-2.0; the terms govern service and support, not the code.

The only address the application's own code contacts is the model endpoint you set — and out of the box, that is your own computer.

The safeguards, in plain English

What it actually does to keep your code and data safe — without the jargon.

1

No account to sign into

There is no login, no API key and no subscription check inside the software. The application has no idea who you are and no way to find out.

2

One destination, and you choose it

The model endpoint is a setting. By default it is Ollama on your own machine; change it and you change where your documents go. That is the only lever, and it is yours.

3

The cloud client was deleted, not disabled

The vendor's cloud provider, account system and search client were removed from the source rather than switched off behind a flag.

4

Updates are deliberate

There is no auto-update, which also means fixes arrive only when you choose to download a new version — a trade-off we state rather than bury.

By the numbers

What was delivered — verified facts from the build, not projected returns.

4,600+
Tests passing
Measured 2026-08-21: 4,618 automated tests across the suite
None
Cloud services required
No account, no API key, no telemetry, no auto-update
19
Locales
Compile-time exhaustive — a missing translation fails the build
3
Platforms
macOS, Windows and Linux installers

Built with

  • Electron
  • TypeScript
  • React
  • Ollama
  • OOXML (.docx/.xlsx/.pptx)
  • Vitest
  • Playwright
  • Apache-2.0 fork

Want something built like this?

We design and ship real, data-driven products — not demos. Tell us what you're trying to make and we'll talk through fit.