How to use Desken

1 · What Desken is — and isn't

Desken is a workstation for communications work where claims have to stand up to scrutiny. You gather your material, let an AI model write a draft from that material, have every claim checked against it, and approve the text as a human before it goes out.

Desken is not an AI service. We run no server and have no model of our own. You connect the AI providers you want to use with your own keys, and the calls go straight from your machine to the provider. That also means you pay the provider for what you use — and that you decide who gets to see your material.

2 · Download, install and update

The app is available at desken.dk/hent. Desken is in closed beta, so downloading requires an access code from whoever invited you.

Mac

• Apple Silicon (M1 or newer — most Macs from 2020 onwards): choose “For Mac”.
• Intel (Macs from before 2020): choose “For Intel Mac”. Not sure? Apple menu → About This Mac. If it says “Intel”, that is the one. Intel Macs cannot run the Apple Silicon build.

The Mac build is signed and notarised with Apple, so it opens without warnings. Drag Desken into Applications and start it from there.

Windows

The Windows build is still in development and is not signed yet. Windows therefore warns about an “unknown publisher”: click More info → Run anyway. With Smart App Control switched on in Windows 11, the file is blocked outright. During the beta we recommend the Mac build.

Updates

The app keeps itself up to date. It looks for a new version when it starts and every six hours after that, downloads it in the background and then asks whether you want to restart now or later. The update is installed when the app closes. Your projects are not affected.

3 · Your user

The first time you open Desken, you create a user with a name and an email address. There is no password and no PIN in the desktop app — your computer's own login is the access control.

The name and email are sent to the publisher and used for one thing: the list of who is using the beta, so you can be told about new versions. Nothing from your projects goes with them. You are not creating an account with us, and there is nothing to log in to. The details are on the privacy page.

The name is used in one place inside the program as well: it appears on the approvals you make, so it can be seen who signed off on a text.

4 · API keys

Without at least one API key, the AI nodes cannot run. The fields are under ⚙ Settings → API KEY.

Desken can use:

• Google Gemini (USA) — free to get started with, can read images and search the web. A good first key. Web search and the external check specifically require a Gemini key.
• Claude (Anthropic, USA) — strong on text and analysis.
• OpenAI (USA) — including Azure OpenAI if your organisation has its own. That puts the data processing agreement under the Microsoft agreement you already have.
• Mistral (EU) · Groq (USA) · Z.ai / GLM (China) · ElevenLabs (USA, for speech).
• Ollama — models that run locally on your own machine. These also work in Offline mode.

Press 🔌 Test connection after entering a key, so you know it works before you build anything on top of it.

Where the keys are kept

Keys are encrypted with your computer's own keychain — Keychain on Mac, DPAPI on Windows — and stay on the machine. They are never sent to us. If the app's signature changes (which happens with a larger update), keys encrypted with the old one can no longer be read and have to be entered again. That is the keychain refusing, not the app.

5 · New task: the guide

Every new project starts in the New task guide. It has three steps:

1. Choose a type. Press release, briefing note and Q&A, customer story, social media pack — or “Other”, where you describe the task yourself.
2. Describe the task. What is being made, for whom, from what angle? Mention length, tone and what has to be included.
3. Add your material. PDFs, images, audio, video or text files. Each file becomes a node.

The guide also has a tick box, Also check claims against the web, which is ticked by default. Desken then adds an external check alongside the check against your material (see section 11). If the text is under embargo, untick it. The box cannot be ticked without a Gemini key, or when data protection is set to Offline.

You then choose the name and location of the project folder, and an AI model builds the workflow on the canvas from your description. With a local model this can take a couple of minutes.

If you would rather start from the fixed workflow, choose Use the standard workflow instead. And ⬜ TEMPLATE opens a gallery of ready-made workflows — the full flow and smaller templates for things like a PDF, an interview or social media. Your own canvas can be saved there as a template too.

6 · The canvas

• Nodes are the small boxes. Each does one thing. Drag them in from the palette on the left, or press ⌘K. Every node is shown in the palette from the start, and Press release, Briefing note and Q&A and Customer story sit under Output in every workspace.
• Customise at the top of the palette lists every node with a tick box. Whatever you remove doesn't disappear — it goes under More nodes at the bottom, and one click brings it back. Every node can also be found with ⌘K.
• Connections are dragged from a node's right-hand dot (output) to another node's left-hand dot (input). Click a connection and press Delete to remove it.
• Zones are the coloured areas the nodes sit in — material, processing, production, checking, approval, output. They are for overview; you can rename and move them.
• Navigation: drag the background to pan, scroll to zoom, and use ⊡ Fit everything in view to get the whole flow on screen.

What the connection colours mean

The colour follows the status of the data, so you can see trustworthiness flow through the workflow:

Running

▶ RUN NODE in the right-hand panel runs the selected node on its own. ▶ RUN FLOW at the top (⌘↩) runs every connected node in the right order, shows progress and turns into a red ■ STOP if you want to interrupt it. If a node is out of date — because the material, the prompt or the model has changed since it last ran — ↻ OUT OF DATE runs just those and everything downstream.

What does this one do?

When you select a node, the right-hand panel shows What does this one do?: what the node is for, and what it is missing before it can run — for instance that a check has no material to hold the text up against, that a web search has no Gemini key, or that an export has no approval in front of it. The panel opens by itself when something is wrong, so you see it before you pay for a run. It uses no AI and fixes nothing on its own. If the workspace has renamed the node, its name in the node catalogue is shown too, so you can look it up in the manual.

7 · Ask about the canvas

The 💬 ASK ABOUT THE CANVAS button in the bottom right-hand corner opens a chat about your workflow. It is also under ⋯ MORE and on ⌘K. Ask about whatever is giving you trouble: why a node can't run, what is missing before export, or which node to use for a task.

• The model is chosen at the top of the chat: the default model from Settings, or any other text model you have a key for. A question about the workflow can be handled by a small, fast model — those are labelled as such.
• Data protection applies here too. In Offline mode only a local model answers, and a provider you haven't approved gets the same refusal as a node would.
• Node names in the answer are clickable and take you to the node on the canvas. If a node has a problem, it appears as a suggested question while the chat is empty. You can also ask from the node itself: 💬 Ask the chat about it at the bottom of “What does this one do?”.
• The chat can do the harmless things itself: show and open a node, run one node, run on from it or run the whole flow, save the project and export as Word, PowerPoint, Markdown or PDF. It only does so when you ask, and the thread says with ✓ what happened.
• The chat can suggest a change — a new node, a connection, an instruction or a new name — but it doesn't carry it out. The suggestion is shown as a list, and nothing changes until you press Apply to canvas. A single ⌘Z undoes the lot.
• The chat cannot delete anything, and it cannot create a route around the approval. If a suggestion touches nodes that have already run or been approved, it says so, so you know what will go out of date.
• Every change you accept is recorded on the node and appears in the transparency report under Changes from the canvas chat. The conversation itself is not saved in the project — an answer from a chat is not documentation.

If the canvas is empty, the chat points you to the New task guide, which can build the workflow for you. The chat can be made larger with the handle in its top left-hand corner; double-click it for the default size.

With 🎙 next to ASK you can speak instead of typing. It needs a Groq or OpenAI key for the transcription. The answer is read aloud in a natural voice from ElevenLabs if you have an ElevenLabs key (the voice is chosen in Settings), otherwise from OpenAI, and otherwise by your computer's own voice. You can ask questions, ask it to run, open or export something, and say “yes” or “no” to a suggestion.

8 · Getting material in

• PDF import — drag a PDF onto the canvas. Open the viewer and select text, and the selection becomes a quote node with a page number that follows it all the way to the finished text.
• Image import — images and scanned documents. Text in the image is read with OCR, which is built into the app and works without an internet connection.
• Audio/video — interviews are transcribed with timestamps, so a quote in the text can be found again in the recording.
• URL import — bring a web page in as material.
• Web search — search the web from the canvas.
• API import — figures straight from Statistics Denmark, the Danish business register, the Danish mapping agency and RSS, or through your own subscriptions.
• MCP connector — fetch material from a remote MCP server (Model Context Protocol) over HTTP. The node calls one specific tool or fetches one specific resource and brings the answer in as text; it is not an open conversation with the server. SHOW AVAILABLE TOOLS/RESOURCES lists what the server offers. The arguments you write on the node are sent to the server. If you write {{input}} in a text value — for example {"field": "{{input}}"}, where “field” is the name the server uses (see them with SHOW AVAILABLE TOOLS/RESOURCES) — it is replaced with the text from the nodes connected into the MCP node; an AI node can then write the search terms and the MCP node searches for them. With no input, the node stops with an error.
• Research search — find scientific articles in Scite from the text you connect. Desken writes an English search (or you write it yourself), and you get a list with title, authors, year, DOI, abstract, how much the article is supported or contradicted by others, and a warning if it has been retracted. Requires your own Scite account: press Sign in to Scite on the node. Only the search itself is sent to Scite.
• Manual text — notes, the angle, an interview transcript. Double-click opens a large editor.
• Source directory — name, title and organisation for spokespeople, so nobody has to type the councillor's title for the third time. The directory follows you from project to project.

9 · Workspaces and house material

A workspace is the frame every node in the project works within. Desken ships one ready-made: 📣 Communications. It names the nodes the way a communications department does (Press release, Documentation check, External check, Approval, Spokesperson, Channel pack) and puts fixed rules into every production and output node.

The rules in the communications workspace include:

• Never write a figure, an effect or a result that isn't in the material.
• If a claim has no evidence, write [NO EVIDENCE: what needs to be obtained] rather than phrasing it anyway.
• Quotes are used verbatim. Never invent a quote, and never rewrite an existing one.
• Only name people who appear in the material, and only with the title given there.

House material

The workspace has four empty fields for you to fill in: About us – boilerplate, Key messages, Banned words and phrases and Tone of voice – excerpt. They are added to every text as standing guidance. They ship empty on purpose: a workspace must never put invented house material into a sender's mouth. An empty field affects nothing.

The workspace also holds ready-made layouts — press release, briefing note and Q&A, customer story, newsletter — with the structure a communications department already uses, and ready-made prompts, among them Risk check: five critical readers at once (the journalist, the competitor, the specialist, the citizen and the employee).

A workspace is a file. You can edit it, save your own and share it with a colleague.

10 · Writing the draft

The draft is written by the model you chose on the node — from the material the node is connected to, and the rules the workspace puts in. When you pick a model per node, there is no hidden shared model behind it: if the node says Claude, Claude is what ran.

Where the material doesn't reach, a visible gap is left — [NO EVIDENCE] — rather than a guess. Those gaps are meant to stay there until you have obtained the evidence.

11 · Checking the claims

The Documentation check (called Fact-checker in the journalism workspace) reads the text against the material it is connected to and marks every claim:

The External check goes to the web instead: one search per claim and a verbatim quote from the source. It requires a Gemini key, and it sends a short search string per claim — not the text and not the material. That is why web search is approved separately under Data protection: a department with a data processing agreement covering its text model should be able to say yes to search without saying yes to everything.

Several models, side by side

Put several checks on the same text, each with its own model. Where they agree, you are on firm ground. Where they disagree is where you should look. The ⚖ Consensus matrix (under ⋯ MORE) shows every check side by side, claim by claim. Rows where the models disagree are highlighted. Expand a row to see each model's exact verdict.

12 · Approval

Approval (Human review in the standard workspace) is the fixed control point. Double-click the node: every point raised by the checks appears as a checklist on the left, and the final text is on the right, where you edit it. Each point is marked ✓ Handled or ✗ Rejected — including when it is the model that got it wrong.

• RUN FLOW never touches your approved text. AI does not overwrite what a human has written.
• The approval is tied to the material. If the material changes afterwards, the approval is marked out of date.
• A diff shows what the human changed — the difference between the AI draft and the approved text, word by word.
• If you export something that hasn't been through a check and an approval, Desken warns you before the file is created.

Version history

Every time you press SAVE in the text editor or in Approval, a version is saved on the node itself. The HISTORY tab lists the versions with the time, the word count, who saved it and whether the text was approved. Select a version to see what has changed since.

RESTORE puts an older text back in the editor. It isn't saved until you press SAVE — and then it becomes a new version, so nothing is lost. The history lives in the project file and moves with the folder. Each node keeps its latest 30 versions; the history does not follow a node when it is duplicated or saved as a template.

13 · Delivery

• Word (.docx), Markdown (.md) and PDF (via print). The file lands in the project's export folder.
• Channel pack — the approved text cut for LinkedIn, Facebook and X, so you don't start over for each channel.
• 📋 Transparency report — which models saw which material with which instruction. It is built from what is already saved in the project: no AI calls, no interpretation. It shows the workspace, the run history, the full system prompt for every node — verbatim — and finally the approval, with what the human changed. Changes you accepted from the canvas chat have a section of their own.
• ⏱ Run log — what ran when, for how long and with which model.

14 · Turn the flow into an app

A finished flow can be shown as a small tool: a couple of fields, a Run button and a result — with no canvas. It is for the colleague who needs to use the workflow but not build it.

1. Press MAKE APP in the toolbar. The guide suggests fields and a result based on the flow: the empty text nodes and web searches as fields, and the flow's final step as the result.
2. Adjust the fields' names and help text. There can be no more than two fields, and a field can be made optional.
3. Press Make the app. From then on the button reads APP, and Edit the app in the app's top bar opens the guide again.

In the app, you fill in the fields and press ▶ Run. The result can be copied or saved as Markdown, and the check's verdicts are one click away. If anything is missing before the flow can run — a key, a search query, an MCP address — the app says so before you press Run. Edit on the canvas switches back. If you want the project to open straight into the app, tick Open the project directly in the app.

• The approval cannot be skipped. If there is an approval along the way, the result is only shown once the text has been approved — until then the app shows Open the review.
• A web search field says where the text goes: whatever you type is sent to a web search at Google Gemini. The search uses only the field's own text — never content from a PDF or an interview further up the flow.
• Fields and the result can also be set one at a time under SHOW IN THE APP in the node's panel.

15 · Data protection: three modes

Under ⚙ Settings → DATA PROTECTION you decide where your material may be sent when AI nodes run. The current mode is always shown in the top bar, so you can see it without opening settings.

URL import, API import and the MCP connector bring data in and are not affected by the mode. If a node tries to use a provider that isn't permitted, it stops with a message explaining why — it does not run half-way. The same goes for the canvas chat.

Statistics and crash reports

Under ⚙ Settings → BETA & FEEDBACK there are two options, both switched off until you switch them on yourself:

• Share anonymous usage statistics — which features are used, for example that a check has run. Never your content, prompts, file names, addresses or keys.
• Send crash reports — if Desken crashes, a technical report is sent to Sentry with the error type, the Desken version and the operating system. Your projects and files are not sent. The choice takes effect the next time Desken starts.

16 · Without an internet connection

Behind a firewall, on a train or in Offline mode, the following works as normal:

• The canvas, all projects, all material, all editing, approval and export.
• OCR of scanned PDFs and images, including Danish. The language data and the engine are built into the app.
• Semantic search in your own material (Ask your sources). The search model is built in as well.
• Local Ollama models.

The following needs a connection:

• All cloud models — Gemini, Claude, OpenAI, Mistral, Groq, Z.ai — and Ollama models tagged cloud.
• Web search and the external check.
• Transcription of audio and video. It exists only in the cloud and therefore does not work offline.
• URL and API import, the MCP connector and the app's check for updates.

17 · Projects, backup and moving

A project is a folder on your own disk, in a location you choose. The folder holds the project file together with kilder/ (sources) and eksport/ (exports). The paths are relative, so the whole folder can be moved, archived or handed to a colleague.

• Autosave runs every 30 seconds. ⌘S saves immediately.
• 📁 PROJECTS keeps track of your cases.
• ↓ Project file under ⋯ MORE makes a single file for backup or for moving to another computer. 📂 Import project file… reads it back.

Large files are not copied needlessly: a big video is added in an instant and does not take up space twice when the disk allows it.

• Unsaved work: if Desken finds work at start-up that never made it into a project folder, it tells you — with the project's name and number of nodes — and offers to open it. The open project folder is not overwritten by it.
• Projects from a newer version: a project saved by a newer Desken than yours is not opened. You are told to update, and the file is left untouched. Older projects are upgraded automatically, without anything being removed.

18 · Keyboard shortcuts

Press ? in the app for the full list. On Windows, ⌘ is Ctrl.

19 · When something goes wrong

“AI runs require a key” — or the node does nothing

There is no key for the provider the node uses. Enter it under Settings and press Test connection.

The web search or MCP node fails, and the rest of the flow is skipped

The node is missing something before it can run: a Gemini key, a search query, an MCP address or a chosen tool. Look under What does this one do? on the node — or, in the app view, at the list of what is missing before the flow can run.

The check “documents” everything

If the material is connected directly to the check, it is checked against itself. Remove the connection and connect only the text; the check finds the material by itself further up the flow. See section 11.

The node stops with “is not on your list of approved providers”

Data protection is set to Approved providers only or Offline. Approve the provider, change the mode, or pick a model that is allowed to run.

The keys are gone after an update

If the app's signature changes, your computer's keychain can no longer read what was saved with the old one. Enter the keys again. This does not happen with ordinary updates.

Transcription fails

Check three things: that the file is under 25 MB, that OpenAI or Groq is permitted under Data protection, and that you are not in Offline mode.

“This project was saved by a newer version of Desken”

The project was saved with a newer Desken than yours. Update the app — the file has not been touched in the meantime.

The canvas chat doesn't answer

The chat uses the model shown at the top of it. If there is no key for that model, add one under Settings or pick another model. In Offline mode only a local model answers.

The consensus matrix is empty

It reads the answers from the check nodes. If you have edited a check's prompt so that it no longer answers in the form the matrix reads, there is nothing to show. Put the prompt back, or run the check again.

A node is marked “out of date”

The material, the prompt or the model has changed since the node last ran. This is not an error — it is a reminder that the result no longer matches what is above it. ↻ OUT OF DATE runs just those nodes and everything downstream.

Windows says “unknown publisher”

The Windows build isn't signed yet. More info → Run anyway. With Smart App Control switched on in Windows 11 it cannot be run at all; use the Mac build during the beta.

Something else

Send a report with ⋯ MORE → 💬 Send beta feedback. It opens an email with the version and operating system filled in — write what you tried, what happened and what you expected.

20 · Support and contact

Write to hej@grumpyproductions.dk. It is the same address for bugs, questions, requests and for being taken off the list again.

What to include

It saves a round trip if the email contains:

• Version and operating system. The version is under ⚙ Settings → BETA & FEEDBACK. If you use 💬 Send beta feedback, both are filled in for you.
• What you tried, what happened, and what you expected.
• Which node and which model, if it concerns a run.
• A screenshot, if something looks wrong.

How quickly we answer

Desken is in closed beta and is built by one person. We therefore promise no particular response time yet — and we would rather say so plainly than publish a number that won't survive a week of teaching. Urgent problems — the app won't start, or a project won't open — go first. A response time will be published when sales open, along with a data processing agreement.

Other things people ask

• What happens to our data? See the privacy page. It is written so an IT department can use it for an assessment.
• What does it cost? The beta is free and has no end date. When sales open, Desken becomes a subscription, and you will be told in good time. The end date of the beta will then be shown under ⚙ Settings → LICENCE. Even after that, your projects, editing and export keep working — only AI runs require a licence.
• Can we have it on Windows? Yes, but the build is still in development and unsigned. Write to us if you need it — it helps us prioritise.
• Can we run it on our own Azure? Yes. Set up Azure OpenAI under Settings, and the data processing agreement sits under your own Microsoft agreement.