A sim game for running real AI agents
Hire an AI team. Watch it work.
Deskburg is a little pixel office where real AI agents do real work. Hire colleagues from a job board, hand them tasks and watch them research, write and team up. They show how they think, remember what you like and hand you real files. When they need you, they raise a hand.
14 days free with your own Claude, OpenAI or Gemini key. No card needed.
Ada runs a meeting: a spring campaign, split three ways.
Runs on the AI you already know
- Claudeby Anthropic
- OpenAI
- Geminiby Google
Pick a model per agent and mix them in one team. Model lists come live from each provider, so new models show up the day they launch.
AI for everyone. Understandable, and fun.
AI agents can already research, write, plan and build. Setting them up still feels like a job for engineers. Deskburg turns it into a game you already know how to play: an office, a team and a to-do list. Everything you do in the game is the real thing.
New in Deskburg
Agents that think out loud, talk back, remember and make real files
Everything you'd expect from a top AI agent, shown the way a game would show it.
See how they think
Watch the reasoning, and choose how hard they think
Every answer comes with a summary of how the agent reasoned its way there. Turn the effort dial from Quick to Leave no stone unturned and see what thinking harder does to the answer, and the cost.
- Reasoning shown live, in plain words
- Effort per agent, or per task
Talk back
Reply like you would to a colleague
Not quite right? Say “shorter”, “now in Norwegian” or “use the second option”. The agent carries on the same conversation, with everything it already knows, and every version is kept.
- Works with Claude, OpenAI and Gemini
- Budgets count the whole conversation
They remember
A notebook that carries over from task to task
Agents write down your preferences, facts about your business and what worked, and read them back next time. Open the notebook to read it, correct it, add a note or wipe it clean.
- Plain notes you can see and edit
- One notebook per agent, private to your office
Real files, not just text
The workshop: code, charts and real office files
Switch on an agent's workshop and it runs code in a secure sandbox: crunch numbers, draw charts and hand you real Excel, PowerPoint, Word and PDF files, ready to download.
- For Claude agents
- A locked-down sandbox with no internet; your files are stored privately in Deskburg
Every game mechanic is a real AI concept
Play the game and you've learned how AI agents work, without reading a manual.
Hire a colleague
Set up an AI agent: its role, its model and its instructions.
Their desk
A task queue and a live transcript of everything the agent does.
Their CV
The agent's skills, tools and sources, and you can edit it.
Salary
Real token spend, with a monthly budget the office can't go over.
A raised hand
Human approval before the agent does anything with consequences.
The tool room
MCP connections that let agents use other apps and services.
The skills library
Reusable know-how you write once and give to any agent.
The meeting room
Several agents splitting one big job, with a lead putting it together.
The notebook
Memory: notes the agent keeps between tasks, which you can read and edit.
The workshop
A code sandbox where the agent calculates and makes real files.
Everything agents can do, in one simple game
38 ready-made roles
Copywriter, SEO specialist, developer, lawyer, recruiter, film editor and more. Or describe the job and the Recruiter writes the role for you.
Any big AI, one click
Paste a key for Claude, OpenAI or Gemini and pick a model per agent. Mix them in one team. No key? Run agents on Deskburg credit instead.
Tools without the plumbing
Agents can search and read the web, save files and post to the pinboard. Connect GitHub or other MCP servers in the tool room and every agent with the right role can use them.
Skills and sources
Write a skill pack once (your brand voice, your report format) and load it on any agent. Put PDFs and notes on a desk and the agent reads them.
Teams and meetings
Agents hand work to colleagues, or call a meeting where a lead splits the job, everyone does their part and the lead puts it together.
Build your own office
Rooms, walls, floors and dozens of pieces of furniture. Save your layouts and switch between them.
You stay in charge
Budgets per office, per agent and per task. Choose how much each agent may do on its own, and approve the rest with one click.
Your keys stay yours
Keys are encrypted the moment you save them and only ever used on our servers. The page only ever shows the last four characters.
The tool room
Plug your office into the tools you already use
MCP (Model Context Protocol) is the open standard that lets AI agents use other apps. Thousands of MCP servers already exist, for code, design, documents, marketing, data and more. Connect one in the tool room and every agent whose role needs it can use it: no code, no setup per agent.
- Paste a server address and its key; Deskburg finds the tools for you
- By default, reading tools run straight away and anything that changes things waits for your approval
- Keys are encrypted and only used on our servers
Connect today
- GitHub
- Context7 docs
- Any MCP server with an API key or token
Next: sign-in connections
- Notion
- Linear
Servers where you sign in with your account instead of a key are next on our roadmap.
Learn AI by playing
You don't need to know what a system prompt, a token or MCP is to start. You'll find out along the way: the CV shows the agent's instructions, the salary shows what each task cost, and the transcript shows every step the agent took.
- See exactly what an agent was told and what it did
- Watch the cost of every task, in dollars
- Try different models side by side and compare
- Turn the effort dial and see what thinking harder costs
Read your office at a glance
- Thinking
- Searching the web
- Writing
- Needs you
- Done
Simple plans. The whole game in every one.
Plans differ in how big your team is and how many agents work at once. Pay monthly, or yearly and get two months free.
Starter
$5 / month
Try it out with a small team.
- 4 colleagues
- 1 working at a time
- 200 MB of files
- 2 tool connections
- Bring your own key, or buy credit
- Most popular
Solo
$15 / month
For solo creators with real work to hand off.
- 8 colleagues
- 2 working at once
- 500 MB of files
- 5 tool connections
- $4 of Deskburg credit included every month
Studio
$39 / month
For small companies running a full office.
- 12 colleagues
- 3 working at once
- 1 GB of files
- 10 tool connections
- $15 of Deskburg credit included every month
Every office starts with a 14-day free trial on your own key. No card needed.
With your own AI key, you pay your provider directly for what agents use, at their prices. Deskburg credit is for running agents without a key: provider list price plus a small margin.
Questions
Do I need to know how to code?
No. If you can write a to-do list, you can run an AI team. Everything is point and click, and every agent shows what it's doing.
Are these real AI agents or a simulation?
Real. The office is a game, but the agents call real models (Claude, OpenAI or Gemini), search the real web and hand you real files.
What does it cost to run agents?
Your plan, plus what the AI provider charges for the work. With your own key, the provider bills you directly. Set a monthly budget and Deskburg stops agents when they reach it.
Can agents do things without asking me?
Only what you allow. Reading and drafting are fine by default; anything with consequences outside Deskburg waits for your approval. You can change this per agent.
Can agents make real files?
Yes. Every agent can write text files and web pages. Switch on a Claude agent's workshop and it runs code in a secure sandbox and hands you real Excel, PowerPoint, Word and PDF files and charts.
Do agents remember me between tasks?
Yes. Each agent keeps a notebook with what it learned about you and your work, and reads it at the start of every task. You can read, correct or clear it at any time.
Is my API key safe?
Keys are encrypted at rest, decrypted only on our servers at the moment an agent needs them, and never sent to your browser.
Which AI models can I use?
Claude, OpenAI and Gemini models, listed live from each provider, so new models show up as soon as they're released.
Your office is ready. Your first colleague is waiting.
Start your free trial14 days free with your own Claude, OpenAI or Gemini key. No card needed.