Skip to content
DESKEN DA Download

Document everything you write. Before it goes out.

A professional
workstation for
communications.

AI works at lightning speed, but it can lack precision.
Desken brings drafting, source checks and human sign-off together in one seamless workflow. Get the power of the industry's best AI models, combined with the documentation trail your professional standards demand.

Download for Mac – free in beta See how it works
The Desken canvas with a press release project in five zones: material, write, verify, approve and deliver.
A real project in Desken, captured by the app itself. Havnestad Municipality is made up. The mistakes in the draft are the everyday kind.

2 August 2026

The EU AI Act's transparency rules apply now.

Desken automatically creates a documentation trail as you go, so you have insight into how the text came about and can account for the process. Desken doesn't make your texts compliant by itself.

Five dedicated zones. One professional canvas.

Desken is built as a professional pipeline from raw material to finished text. You build scalable workflows yourselves, with your own sources, standing rules and a strong documentation trail.

  1. 01MaterialReports, interviews and data
  2. 02WriteAI drafts with your rules
  3. 03VerifyEvery claim, several models
  4. 04ApproveA person signs off
  5. 05DeliverDocumentation included
The canvas with the whole flow run in five zones: material, write, verify, approve and deliver. The connections are blue from the sources, yellow from the AI draft, red from the check and green after the sign-off.
  1. 01 · MaterialThe report, the interview and the spokesperson are on the canvas. The connections are blue: raw source.
  2. 02 · WriteAI writes the press release and the briefing note under the workspace's rules. The line turns yellow: unchecked AI text.
  3. 03 · VerifyThree models cross-check every claim, and the risk check reads the text as five critical readers. They catch the pitfalls – the line turns red.
  4. 04 · ApproveA person makes a call on every finding and approves the text. Now the line is green.
  5. 05 · DeliverThe text is rendered for export. The channel pack builds only on the approved text, and the transparency report shows how it got there.

01 – Material

Material

The foundation for your text.

Put reports, interviews and data on the canvas. The models are instructed to use only this material, so the draft builds on your own factual basis. Whether they stuck to it is exactly what Verify checks.

An interview with the councillor, transcribed with timestamps and shown in Desken's inspector – next to the report and the spokesperson on the canvas.
  1. 1 / 3 · The canvasEverything the text may build on is on the canvas – your factual basis, all in one place.
  2. 2 / 3 · The sourcesThe report has been read and indexed, the interview transcribed, and the spokesperson is there with name and title.
  3. 3 / 3 · At 01:07The councillor says it himself: “I wouldn't dare say that.” That will matter when the text is verified.

Documents and web pages

PDFs, images and web pages become nodes. Quotes carry their file and page number all the way to the finished text.

Audio and video

Interviews are transcribed with timestamps. If a quote ends up in the press release, you can find it in the recording. The audio is sent to OpenAI or Groq using your own key, so transcription needs one of them to be allowed and doesn't work in Offline mode. Files can be up to 25 MB.

Open data

Pull figures straight from open APIs such as Statistics Denmark and the Danish business register – or through your own access to news services like Ritzau and Infomedia.

Spokespeople

Name, title and organisation live in a source library that follows you from project to project. Nobody has to type in the councillor's title for the third time.

02 – Write

Write

Your rules are in place before the model writes a word.

The Communications workspace contains fixed layouts and editorial rules. Fill in your boilerplate and key messages once – and they're built in automatically every time the model writes. The ready-made workspace comes in English and in Danish.

The draft editor with the press release. On the right, the inspector shows which layout and house material the workspace has put in.
  1. 1 / 3 · The draftAI writes the draft in the fixed layout. On the right: what the workspace has put into the node.
  2. 2 / 3 · From the workspaceThe layout, the editorial guidelines, the boilerplate and the key messages – in place before the model wrote a word.
  3. 3 / 3 · The press contactThe material didn't say who the press contact is. So it says [PRESS CONTACT MISSING] – not a name the model made up.

Fixed layouts

Press release, briefing note and Q&A, customer story and newsletter. Each with the structure a communications team already uses.

House material

Boilerplate, key messages, banned words and tone of voice are built into every text as standing guidelines.

Gaps stay gaps

Where the material runs out, visible gaps are left: [NO EVIDENCE]. The model consistently marks where data is missing, rather than guessing.

Never write a figure, an effect, a ranking or a result that isn't in the attached material.

First rule in the Communications workspace
The Communications workspace with its standing rules and style.
The rules live in the workspace – not in the individual prompt, where they get forgotten.

03 – Verify

Verify

Disagreement is a signal.

Three different top models cross-check every claim independently of each other. Discrepancies are highlighted visually, so you can focus exactly where the models' assessments differ.

The consensus matrix: three models assess eight claims. The two rows where the models disagree are at the top.
  1. 1 / 4 · Sonnet 5The first model is done. Three claims are contradicted by the material, and two have no evidence.
  2. 2 / 4 · GPT-5.5The second model is done – and sees one claim differently. That one moves to the top.
  3. 3 / 4 · Mistral Large 3The third model is done. Now they disagree on two claims. Those are the ones a person should look at first.
  4. 4 / 4 · Only the disagreements“More than ever”: two models find no evidence, one says documented. Disagreement is highlighted as an important signal to you.

From the example · press release, draft 1

Havnestad halves the waiting time for building permits

Havnestad cuts 25 days off the waiting time for building permits

Five places in one press release where the text went beyond reality. The system catches the imprecise interpretations.

  • Contradicted by the materialhalves the waiting timeFrom 64 to 39 days is a drop of 39 per cent. Not 50. Report, p. 4
  • Contradicted by the materialcut 42 per cent offThe report says 39 per cent. p. 4
  • Contradicted by the materialmore than halved the number of applications sent backFrom 31 to 18 per cent is a drop of 42 per cent. p. 7
  • No evidenceamong the fastest municipalities in the countryThe evaluation doesn't compare with other municipalities – and the councillor himself says he wouldn't dare claim it. p. 9 · interview 01:07
  • Disagreementmore than everTwo models: no evidence. One: documented. The person signing off rejected that verdict. p. 5

Documentation check

Every claim is marked documented, no evidence or contradicted by the material – with a reference to where it says so.

Several models

Check with Claude, GPT, Mistral or a local model – side by side. You make the final decision, weighing up the models' assessments.

External check

Claims can also be checked against the web: one search per claim and a verbatim quote from the source.

Risk check

Five critical readers at once: the journalist, the competitor, the expert, the citizen and the employee. What's the worst headline this text could turn into?

One model's documentation check in the inspector, with a verdict for every claim.
Each check is a node with its own model – and its own report.
The risk check: the text read by five critical readers, with the biggest risk and concrete fixes.
The risk check points at the headline. That's what everyone reads first.

04 – Approve

Approve

A person signs off on the result, with a name and an exact timestamp.

The sign-off acts as a fixed checkpoint. Every finding from every check is gathered in one place, so you can deal with every point before the text moves on.

The sign-off: the findings from every check on the left, marked handled or rejected, and the corrected text on the right. The diff view in the sign-off: the difference between the AI draft and the approved text, word by word.
  1. 1 / 4 · The findings wait24 findings from three checks are waiting for an answer. Here, every single one gets an answer before the text is approved.
  2. 2 / 4 · Answer by answerEach finding is marked handled – or rejected, when it's the model that's wrong.
  3. 3 / 4 · ApprovedAll 24 have an answer. Then the text is approved – with a name and a time.
  4. 4 / 4 · What the human changedThe diff shows the difference between the AI draft and the approved text, word by word: “halves” out, “cuts 25 days off” in.

A fixed checkpoint

The sign-off is a built-in step in the professional workflow. If you export without one, you get a warning before the file is created.

Every finding gets an answer

The findings from every check are gathered, and each one is marked handled or rejected. Even when it's the model that's wrong.

Tied to the material

The sign-off carries a name and a time. If the material changes afterwards, it's marked as out of date.

What did the human change?

A diff shows the difference between the AI draft and the approved text – word by word.

05 – Deliver

Deliver

The report isn't just for show.

The approved text is rendered straight to Word, PDF or Markdown. Alongside it, you can download a transparency report that serves as your documentation of how the text came about.

The transparency report: the notice that it contains only saved data, the workspace, the run history and the full system prompt for the first step.
The end of the report: the human review with the name of the person who approved the text, and the time.
  1. 1 / 5 · Saved data onlyThe report is built from what's already saved in the project. No AI calls. No interpretation.
  2. 2 / 5 · The workspaceIt says which workspace the nodes were built in – and what the workspace put into them.
  3. 3 / 5 · The run historyWhen the flow ran, for how long and with which models.
  4. 4 / 5 · The system promptEvery node with the full system prompt it actually ran with – verbatim, including the rules from the workspace.
  5. 5 / 5 · The sign-offFinally: who approved the text, and when – and what the human changed.

Word, PDF and Markdown

The approved text lands straight in the project's export folder – ready to send.

Channel pack

The approved text, adapted for LinkedIn, Facebook and X. You don't start from scratch for every channel.

Transparency report

Get the full overview of which models saw which material, with which instructions. The documentation is built automatically, as an integrated part of your work.

A project is a folder

Sources, exports and the project file sit together on your own disk. Ready to archive, move and hand over.

The name on an approval is the user who was signed in to Desken when the text was approved. It is information in the project file – not a verified signature.

The transparency report on its own sign-off
The channel pack: the approved press release adapted for LinkedIn and Facebook.
The channel pack builds on the approved text – not on the draft.

Your data

Your files. Your hardware. Your choice.

We never see your texts.

Desken runs locally on your own machine and connects directly to your own AI agreements. We have no access to your sources or texts – and you decide which providers get to see them.

  1. Open

    All the providers you've set up with your own keys.

  2. Approved only

    Only the providers you've ticked – for example the ones you have a data processing agreement with.

  3. Offline

    No content leaves the machine. Only local models run, and web search, external checks and transcription are switched off.

Desken's settings: data protection set to Offline. No content leaves the machine, and only local Ollama models may run.

Works with 8 of 8 may be used

  • Azure OpenAI in your own region
  • Claude
  • OpenAI
  • Google Gemini
  • Mistral (EU)
  • Groq
  • Z.ai
  • Ollama – fully local

One key for the team

Use your organisation's own Azure OpenAI – so data processing is covered by the Microsoft agreement you already have.

The keys stay at home

API keys are encrypted with the computer's own keychain – Keychain on Mac – and stored only on your machine.

No middleman

Desken calls the AI providers directly from your workstation. We register your name and email when you sign up, and your texts never pass through us. The rest is spelled out on the privacy page.

Built for industries that demand uncompromising documentation.

Public sector communications

Press releases and texts for citizens, where every figure has to hold up against the organisation's own reports.

Regulated industries

Where marketing has to be approved, and where you need to show who signed off on the text.

Agencies

Customer stories and campaigns, where the client's figures and quotes have to hold up – even when the client is reading along.

Pricing

Desken

Free in beta

No expiry date and no credit card.

  • Every feature
  • You only pay your own AI provider – directly
  • Updates itself
  • Your projects are your files, and they stay with you
Download for Mac

Desken for teams

When sales open

The price for the final version is still being worked out. Invoice payment and Azure OpenAI as a provider are there from day one, and you'll get all the details well before the commercial launch.

  • Pay by invoice – including EAN invoicing for Danish public bodies
  • One Azure key for the whole team
  • A workspace with your own rules and house material
Write to us

Try it on
a real text.

Download the app, add your own API key and run a current text through the flow. It takes a few minutes to set up and gives you the full picture of how thorough the models and the pipeline are.

Download Desken for Mac

Desken is in closed beta, so the download needs an access code. Write to us if you'd like to join. The Windows version isn't ready yet.