Start With OpenWork: One Folder, One Task, and a Look at Every Change | Heidelberg AI Curriculum
T21-L01
Start With OpenWork: One Folder, One Task, and a Look at Every Change
Install OpenWork, a free desktop app in which an AI agent works on the files in one folder. Give it a practice folder, switch it so that every change waits for your click, hand it one task, and check the table it wrote against the files.
Your group leader asks which of the shared instruments broke down most often this year. The answer is in a folder of maintenance notes, one file per incident, written by six different people. You could open all forty, or paste them one by one into a chat assistant. Both take the afternoon.
In a company the folder holds supplier quotes, and the question is who delivers fastest. It's the same job: read many files, pull out the same few facts, put them in one table.
A chat assistant can't do this for you, because it only sees what you paste. An agent can. It opens the files itself, reads them one after another, and writes the table into the folder. That saves the afternoon. It also means a program is now reading and writing on your computer, so you need to know what it may touch, where your files go while it reads them, and whether the table is right.
This book sets that up with OpenWork, a free app for exactly this kind of desk work. You are at Level 1 User: you work on a practice folder, and the result is for you alone.
After this you can
Install OpenWork and give it one practice folder, with no account.
Pick a model and say where your files go while it works.
Read what the agent may do without asking, and switch file changes to ask first.
Give the agent one task, read each permission request, and allow or deny it.
Check the file the agent wrote against the sources, and name what it got wrong or left out.
A Windows, Mac or Linux computer you may install software on, and room for a 300 MB download. If IT controls installs on your work computer, ask them to install OpenWork.
No files of your own. The agent makes the practice files in step 6. Real work files come later.
An internet connection, unless you already run a local model, one that works on your own computer.
About 60 minutes: 15 to read, about 45 to do.
Last verified 12 October 2026 against OpenWork's documentation and version 0.18.57. The desktop app is free and needs no account. It includes a free starting model, Auto, with a limited allowance; after that you connect a model of your own. Details for experts at the end of this book.
The idea in one page
A chat assistant answers. An agent acts. An agent works in a loop: it looks at what is there, picks an action such as opening a file, sees the result, and picks the next action. It stops when it thinks the job is done, when it needs you, or when you stop it.
The folder is the fence. OpenWork works only in a folder you choose. That folder is called the workspace. Reaching anything outside it needs your permission first. So the first decision is the folder, and for now it is a practice folder.
Inside the fence, the agent doesn't ask. On a fresh installation the agent reads, creates and changes files in the workspace straight away. That is fast, and it is why you never start in a folder with files you care about. In step 7 you switch this folder so that every change waits for your click.
The model does the thinking, and it sits somewhere. OpenWork is the hands; the model is the head. Every file the agent reads is sent to the model so it can decide what to do next. Where the model runs is where your files go:
Your choice
Where the files go
Cost
Auto, OpenWork's free starting model
To computers run by OpenWork's makers
Free, with a limited allowance
Your own account with a model company, such as a paid ChatGPT plan
To that company
Your plan or your bill
A local model on your own computer
Nowhere. They stay on your computer
Free, but needs a strong computer
A permission request is the brake. Once you have switched it on, a small window appears before the agent changes a file. It is titled Modify files? and names the file. You click Allow once, Allow for session (no more questions of this kind until you close this chat) or Deny. A sentence in your own message such as "don't change my notes" is a wish. The permission request is the lock.
What the agent writes is a draft. It will tell you, in a friendly sentence, that it has read everything and built the table. Believe the table only after you have counted its rows and checked a few against the files.
Set up OpenWork (steps 1 to 7, once)
Step 1. Download and install. Open openworklabs.com/download (opens in a new tab) and choose your system: Windows, Mac (Apple Silicon for Macs from 2020 on, Intel for older ones) or Linux. Open the downloaded file, install it the way you install any program, and start OpenWork.
You should see: a window that says Welcome to OpenWork, with the line Your computer, but it works for you.
Step 2. Start without an account. Click Use Without Cloud. The other button, Sign in to OpenWork Cloud, is for teams and comes two books later.
You should see: your system's folder window, titled Authorize folder.
Step 3. Choose the practice folder. In that window, go to your Documents, make a new folder called openwork-practice, select it and confirm.
You should see: the main window. Your folder's name stands in the list on the left under Workspaces, and at the bottom is a message box that says Describe your task...
Step 4. Pick a model. Click the model name under the message box. If a model called Auto is in the list, choose it. That is the free starting model, and it needs no account.
You should see: the model's name under the message box.
If Auto isn't in the list, or you see Connect a model to get started, you need a model of your own. Click the round picture or initials in the bottom-left corner, open Settings, then AI Providers, and click Connect provider. A provider is a company that runs models. If you pay for ChatGPT, choose OpenAI and sign in with that account. If you have no such account, ask IT or a colleague who uses OpenWork which model your organisation allows.
Step 5. Read what the agent may do here. Click the round picture or initials in the bottom-left corner and open Settings. Look for the list How agents run here; it sits near Authorized folders. Each line names one kind of action and says what happens before the agent does it: Runs, Asks you or Blocked.
You should see: the line File edits with Runs, and the line Files outside authorized folders with Asks you. If File edits already says Asks you, skip step 7.
Step 6. Watch it act: let it make the practice files. Go back to the chat, type this into the message box and click Run task: "Create eight short maintenance notes for shared lab instruments in this folder, each as its own text file. Vary the instrument, the date of the breakdown, the problem and the action taken. In one note, leave the action out."
You should see: eight files appear in openwork-practice. Nothing asked you first. That is what Runs means.
Step 7. Switch on the brake. Copy these three lines, paste them into the message box in one go, and click Run task:
Create a file named opencode.jsonc in this folder with exactly this content:
{ "permission": { "edit": "ask" } }
Don't change any other file.
This file holds settings for this one folder. The line in it tells the agent to ask before every file change. Now close OpenWork and open it again, so it reads the file.
You should see: under Settings, in How agents run here, the line File edits now says Asks you. If it still says Runs, open opencode.jsonc from the folder with Notepad (on a Mac, TextEdit) and compare it with the middle line above, character by character. If it differs, send the three lines again and restart once more.
Worked example: one table from eight notes
The example follows the four steps of T01-L01 · From Idea to Working Result: The Four-Quadrant Method: Structure it (say what you need), Understand it (look before you act), Choose tools (pick what does the job) and Build it (make it). Then back to Understand it, to check.
Lab: eight maintenance notes
Mira is a technician. Her practice folder holds the eight notes from step 6 of the setup. She opens three of them and reads them, so she knows what the table should say.
1. Structure it. Before she types anything, Mira writes down what she wants: one table, one row per note, with five columns: file name, date, instrument, problem, action. The table is for her. A wrong row costs her a few minutes, nothing more.
2. Understand it. Her first message asks for no changes at all: "List the files in this folder and say in one line what each contains. Don't create or change anything." She clicks Run task.
You should see: a list of files with one line each, and no request, because nothing is being changed. Mira counts eight notes in the list. If the agent had listed seven, she would know now, before any table exists.
3. Choose tools. OpenWork with Auto is enough: the notes are invented, so it doesn't matter that they travel to OpenWork's makers. For the real notes she would ask first whether they may leave the building, and switch to a model her lab has approved. The table will be a CSV file, a plain table that opens in Excel, because OpenWork's makers recommend it for a simple table.
4. Build it. Now the task, with its limits in the same message: "Create overview.csv in this folder with one row per note: file name, date, instrument, problem, action taken. Write 'not stated' where a note doesn't say. Don't change or delete any other file."
You should see: after a moment, a small window titled Modify files? with the line Review the file and diff before allowing OpenCode to make changes. OpenCode is the engine inside OpenWork. The window names the file, and the diff is the list of lines that would be written.
5. Build it: read the request, then decide. Mira reads the file name: overview.csv. That is the one file she asked for, so she clicks Allow once. Had the window named one of her notes, she would have clicked Deny and asked the agent why.
You should see: overview.csv in the folder, and a message from the agent that it has finished.
6. Understand it again: count. Mira opens overview.csv in Excel. Eight rows, plus the line of column names. Eight notes, eight rows.
7. Understand it: check three rows. She picks three rows and opens their notes next to the table. Two match. In the third, the agent has written the date the note was created, not the date of the breakdown that the text mentions. She also looks for the note without an action: the table says not stated, as asked.
8. Build it: correct it. She goes back to the chat: "In row 5 the date should be the date of the breakdown given in the text, not the file date. Check all rows for the same mistake and correct overview.csv."
You should see: a new Modify files? window for overview.csv. She allows it once and checks row 5 again.
Checking the rows sent her back to Structure it: "date" was not defined well enough. In the method this is the loop Structure it ⇄ Understand it. Checking tells you what you should have asked for.
Company: seven supplier quotes
Jonas buys consumables. In step 6 of the setup he asks for seven made-up supplier quotes as text files. His columns are file name, supplier, item, unit price, delivery time and valid until. He takes the same steps: list first, then the task with "write 'not stated'", then read the request.
His check finds something different. One quote gives the delivery time as "2 to 3 weeks", and the table says "2 weeks". Nothing in the agent's closing message hinted at it. He asks for the column to quote the text as written, and checks that row again.
Stopping and tidying up
Stop a run: click Stop next to the message box. The agent stops where it is. Files it has already written stay.
Start fresh:New session on the left opens an empty chat in the same folder. The old one stays in the list.
Take OpenWork off a folder: move the mouse over the workspace's name on the left, click the three dots ⋯ and choose Remove workspace. OpenWork says Sessions and files on disk are preserved: the folder and its files stay where they are.
What goes wrong
A folder without the brake.Symptom: in your real Documents folder, or in any second folder, files change with no request. Fix: give the agent a folder made for the task, with copies. The settings file belongs to the folder it is in, so repeat step 7 in every new workspace and check How agents run here.
The table opens in one column.Symptom: in Excel every row sits in column A. Fix: Excel with German and many other European settings expects semicolons. Ask the agent to rewrite the file with semicolons between the values.
Allowing for the whole session on the first request.Symptom: later changes in this chat happen without a question. Fix: use Allow once until you have seen the same task succeed a few times.
Believing the closing message.Symptom: "I've summarised all the notes", but the table has seven rows for eight files. Fix: count rows against files, and check three rows against their sources every time.
Blanks filled with guesses.Symptom: every cell has a value, even where the file is silent. Fix: put "write 'not stated' where a file doesn't say" in the task, and look for it afterwards.
A model that stalls.Symptom: the agent answers in general terms and never opens a file, or stops halfway. Fix: pick a more capable model from the list. Models small enough for a laptop often chat well but can't steer an agent.
Real files and a model somewhere else.Symptom: confidential files went to a company you never chose. Fix: decide the model before the folder. T12-L01 · Decide What You May Paste Into AI at Work shows how to decide.
Do it yourself (45 minutes)
Set up (15 minutes). Do steps 1 to 7 if you haven't yet. Then ask for a second set of practice files of your own kind, such as quotes or minutes, in a new folder set-2 inside the workspace, with one file that leaves a fact out. Under How agents run here you should see File edits: Asks you.
Structure it (3 minutes). Read three of the files. Write down your columns and who the table is for.
Understand it (4 minutes). Ask for a list of the files in set-2 with "Don't create or change anything." Count the lines against the files in that folder.
Choose tools (2 minutes). Write one line: which model, and where the files go while it works.
Build it (8 minutes). Send the task with its limits. Read every request. Allow changes only to your output file, once each. Deny a change to any other file.
Understand it again (10 minutes). Count the rows. Check three against their sources. Find the not stated. Ask for one correction and check that row again.
Check the sources and tidy up (3 minutes). Sort set-2 by Date modified: only your output file should be newer than the practice files. Then remove the workspace and see that all files are still there.
Exit check
Hand in the overview table and a five-line note:
the model you used, and where your files went while it worked;
the number of files and the number of rows;
the three rows you checked, and what you found;
one thing the agent got wrong, or one cell it marked not stated;
which requests you allowed, and whether you denied any.
It passes when the row count equals the file count and your source files are unchanged.
Rule to remember
Give the agent a practice folder, make it ask before it changes a file, and check what it wrote before anyone else sees it.
Details for experts
Checked on 12 October 2026. You don't need this to follow the book; it's for setting things up or answering a colleague.
What
Details
What it means for you
Version and downloads
OpenWork 0.18.57, published 7 October 2026. Installers: Windows about 294 MB, Mac Apple Silicon about 329 MB, Mac Intel about 348 MB, Linux AppImage (one file you start directly) about 256 MB.
A new version appears about once a week, so a button may have moved.
Licence
The desktop app is open source under the MIT licence, which allows free use at work.
No licence to buy for this book.
Auto
The free starting model, listed under OpenWork Models (Free) and served from inference.openworklabs.com. The vendor doesn't publish the size of the allowance.
When it runs out, connect a model of your own.
Your own account
Settings › AI Providers › Connect provider. Choosing OpenAI lets you sign in with a paid ChatGPT plan. Anthropic needs an API key, a password for programs that is billed by use; a Claude subscription doesn't work here.
Your files then go to that company, under your contract with it.
A local model
Settings › Ollama connects to Ollama, a free program that runs models on your own computer, then Add to workspace. The vendor warns that small local models often fail at the multi-step work an agent does.
The engine's defaults: actions inside the workspace run, reaching outside it asks, and files named .env, where programs keep passwords, are blocked. The settings file opencode.jsonc in a workspace changes this for that folder.
Other request titles are Read files?, Run a shell command? (a typed instruction to your computer) and Access an external folder?
Undo
Revert takes back your last message in the chat. The engine restores files through Git, a change-history tool this book doesn't set up.