Create your first agent team with Orkeon Studio

You hear about AI agents everywhere, but in practice the tools often ask you to write YAML, open a terminal or understand an architecture. And most agentic tools live on the command line, under Linux or macOS. This one runs on Windows ❤️. Orkeon Studio tackles the problem from the other end: you describe a piece of work the way you would explain it to a colleague, and an assistant composes the team of agents that will do it. In this article, we install Studio, give it an AI model, then create a small team that reads your monitoring notes and writes a summary — all with the mouse, not a single line of code, not a single command to type.

One word before you start. Orkeon Studio is free and open source: the source code is public on GitHub under the MIT licence, there is nothing to buy, no account to create, no paid tier holding back the useful features. It is also a young application — the version used in this tutorial is a release candidate — and it will improve fast: releases follow one another closely, and each one folds in the fixes and the ideas its first users report. What you learn here will not be wasted on the way: the screens may move a little, the way of working stays. And the only thing that can cost you money is the one you choose — the AI model, with a free path for those who want to pay nothing at all.

Install Orkeon Studio

Studio runs on Windows only. That is the only technical prerequisite: a Windows 10 or 11 machine is enough, and you need nothing else.

1. Open the Orkeon releases page: https://github.com/Orkeon/orkeon/releases, click Assets and pick the most recent file, orkeon-<version>-win-x64.msi (for example orkeon-1.0.0-rc.6-win-x64.msi, the first .msi in the list). Or, more simply, click the address below: the current version will download. https://github.com/Orkeon/orkeon/releases/download/v1.0.0-rc.6/orkeon-1.0.0-rc.6-win-x64.msi — the download will start and you will find the file on your PC, in the Downloads folder. Each file comes with its sha256: digest on the same line, to check the download; for a first try, you can skip that check (screenshot 1 below).

2. Run the downloaded file. Windows shows a blue box, "Windows protected your PC": that is normal, the installer is not signed (a code-signing certificate costs a yearly fee). Check that you downloaded it from the official releases page, then click More info, then Run anyway (screenshot 2 below).

3. Follow the installation wizard: licence, progress, done. Studio installs into your user profile, without asking for a folder. If you reinstall or update later, close Orkeon Studio first.

4. Open the Start menu and launch Orkeon Studio (screenshot 3 below). The window opens in the language of your Windows, and you can switch it at any time from the title bar — the EN selector at the top right, which is what every screenshot in this tutorial uses, so the labels below match exactly.

Worth knowing: the installation also drops the orkeon command-line tool on your machine. You will not need it in this tutorial — Studio uses it behind the scenes.

Give your agents an AI model

An agent is a small program that works for you: it reads files, thinks, writes results. To think, it needs an AI model — the "brain", supplied by a service like OpenAI, Anthropic or DeepSeek, or by a model installed on your own machine. This is the first setting to make; without it, Studio's assistant cannot compose anything.

1. In the left navigation, ENVIRONMENT group, click Settings. The AI model tab opens: the Which AI model your agents use section lists your settings — a setting is an AI model and its address, which you assign to each team.

2. Click + New setting. A dialog opens: "A setting has a name and can serve several teams."

3. In Setting name, type a name that says what it is for — for example New setting-Deepseek. Under Where the model comes from, two groups of cards appear: ON YOUR MACHINE · FREE, NO KEY (Ollama, Docker Model Runner) and IN THE CLOUD · API KEY REQUIRED (OpenAI, Anthropic, DeepSeek, Mistral, Gemini…) — models that run online, on the provider's servers, hence the API key. Pick a card from that group — DeepSeek, our example from here on, is the simplest path to start. The Base URL and Model fields fill themselves in (screenshot 1 below).

4. An API key is a technical password that identifies your account with the provider. In the API key DeepSeek panel — its title follows the provider you chose — paste your key, then click Remember the key (screenshot 2 below). Studio keeps it in your Windows session — never in a file, never in a shared folder. No key yet? The box 💡 No account, tokens or key yet? below tells you how to create one. Back in the dialog, Test connection and Read the balance confirm the key works.

💡 No account, tokens or key yet? The online providers all work more or less the same way: you create an account, top up tokens, then create the API key. Here is the detail with DeepSeek, one of Studio's online cards, known for its low prices — the same steps apply to the others:

1. Create an account on platform.deepseek.com with your e-mail address (or your phone number).
2. Top up your account: in the Top up menu, choose an amount — from 2 $ — and pay online. Depending on the region, bank cards are accepted or not: if yours is declined, the platform offers other payment methods.
3. Create your key: in the API keys menu, click Create new API key. The key is shown once, as sk-…: copy it immediately into a folder outside Orkeon Studio's perimeter — for example C:\Users\<your name>\Desktop\api-key, though any folder Studio never opens will do. Keep that copy safe — it is the only one you will get: the provider will not show the key again, and Studio keeps none you can read.

5. Click Save (or Ctrl+S). Your setting appears in the list — the cards offer Edit, Duplicate, Delete and Use as default, and one of them carries the default badge: that is the one all your teams will use, including the assistant that is about to compose the next one.

If you would rather pay nothing and keep everything on your machine, the Ollama card works without a key or an account — but it asks you to install Ollama yourself, an extra step we avoid here.

Describe the work to hand over

1. In the left navigation, AGENT TEAMS group, click My teams, then Create a team. The assistant opens: "Nothing is written on your machine before adoption." Four steps — Describe, Compose, Try, Adopt — and nothing is written until the last one.

2. At the Describe step, in the What work do you want to hand to this team? field — "Write it the way you would explain it to a colleague. No technical words needed." — paste this mission:

   Every week, read the notes placed in the "input" folder. Based on these notes,
   write a monitoring summary in the file "monitoring-summary.md". The summary must
   present the three key takeaways, with a short paragraph for each point, and it
   must indicate the source for each point. The output file must be named
   "monitoring-summary.md".
   

3. Nothing to prepare on disk at this point: the folders the team will read and write are chosen at the next step, and when you keep them inside the team, Studio creates them itself at adoption. The input and output folders do not exist yet — that is expected, and nothing has to be declared.

4. If you want a tighter wording, click Rephrase (screenshot below): the assistant rewrites your text, and Back to my words undoes it if you do not like the result — "Request rewritten — Back to my words restores what you typed."

5. Then move to step 2, Compose. The assistant does not hand you the team straight away: it first asks where it will work.

Check the composition, then run a trial

1. First screen: Your team's folders — "Read from your request. Rename a folder if you like, say where each one lives, then confirm — the team will use exactly these." Our mission asks for two: input, marked read — the one where you will drop your notes — and output, marked written — the one where the summary will be written. For each one, the Location selector offers Inside the team and Folder on this machine (screenshot 1 below). input is marked Read and output Read and write: an agent only touches a folder through the right the team gives it. Keep both on Inside the team: the team simply gets an input and an output folder of its own, created when you adopt it — the line under the selector reads "inside the team: input", and there is nothing to declare in Settings. Folder on this machine is for folders you already keep elsewhere: those must be declared in advance in Settings → Authorized folders, one by one — that is the only case where a declaration is required. Then Confirm the folders seals it (screenshot 3 below).

   Before trying, drop the documents the team will have to process into the input folder: one or two monitoring notes are enough. Click Open the folder (screenshot 2 below) — the hint by the folders says it plainly: "The trial reads the folders kept inside the team from the working session — drop your documents there with « Open the folder »; they move into the team when you adopt it." If the folder is empty when the trial runs, the agents have nothing to read and the trial runs on air.

2. Next comes the proposal (screenshot 4 below): "Here is the team I propose". For our mission, one agent is enough — a monitoring analyst and writer, who reads every note in the input folder, decides which three points matter, and writes monitoring-summary.md from what the notes say. The assistant gives its reason on the same screen: the job is one short document built from a handful of notes, so a single clear role covers it, with no hand-off that could blur the link between a point and the note it came from. The agent is shown with its abilities under Can: — list, read, write — and you can add another one with Add an agent; for a first trial, keep the proposal as it is.

3. The composition ends with a summary — "Team composed — you can try it." The Try the team button launches the trial (screenshot 4 below) — and if you are in a hurry, Adopt without trying exists, but the trial is worth the minute it costs.

4. Rest assured: "A trial changes nothing definitive: the result is written into a session folder you can throw away."

5. When the trial ends, the screen shows 4 of 4 criteria met and The verdict, with its score out of 1.

   The verdict is a score, not a formality: every point of your mission is checked, one by one, and the score of 1.00 says they were all met — 4 of 4 criteria met in our case. When something is off, What the engine suggests changing lists the corrections that would raise it. Our monitoring team, well framed from the start, came back green across the board on its very first trial, with a score of 1.00 — written 1,00 in the capture below, where the number follows the Windows regional settings. Nothing to correct here, but a lower score is not a dead end: the panel on the right, "The trial is judged. What do we do next?", offers the same four buttons every time. Looks good to me keeps the result and takes you to the adoption step. Run the trial again replays it as it is, for a second sample. Fix and retry applies the engine's suggestions to a sharpened mission and runs the trial again — that is the button that turns a 0.85 into a 1.00. Abandon drops the composition and sends you back to Describe. A red verdict is not a failure: it is the engine telling you what to make more precise — and the trial writes only into a session folder you can throw away:

Count your trials, though: the session only grants three.

Why three trials and no more? Every trial makes the AI model really work — and therefore spend tokens. The limit of three per composition session simply caps the bill if a trial goes round in circles. When the ceiling is reached, the session is Stopped and the team waits for you in My teams; to carry on, the simplest is to go back to Describe, refine the mission and start a fresh composition.

Adopt the team and run it for the first time

1. At the Adopt step: at the very top, Team name holds the long name the assistant generated from your mission — Weekly Monitoring Summary in our capture, which still shows it at that point. Replace it with something short, Weekly monitoring: that name is the one the list and the folder on disk carry, and the card's Rename button changes it at any time (screenshot 2 below shows the short name). Weekly monitoring becomes teams\weekly-monitoring\. Leave the Model setting as it is (the default setting you just created), and under When should it work? choose On demand ("I launch it myself"). The Adoption instruction section — "An instruction applies to every agent, travels with the team, and becomes a point to check at trial time." — can stay empty for this first team. Click Save to my teams (screenshot 1 below): it is only at this moment that Studio writes something to your disk.

2. The team appears in My agent teams, as a card (screenshot 2 below) — "Each team is an autonomous folder on this machine — copiable, shareable, deletable."

3. Click Launch. The Launch a team screen opens (screenshot 1 below): "A trial runs like a real run but stays out of the history; a real run is recorded there." The Team field already holds your team — Weekly monitoring — and the card under it recaps the mission, the folders it reads and writes, and the model setting it uses: "Team ready: 1 agent, 2 declared folders." Check that it is the right team, leave Run type on Real, then click the big blue Launch now button. Validate (dry-run), next to it, checks the launch without running anything.

   The Options step, on the left, keeps the defaults — a normal run, with the progress shown. You have nothing left to do but wait: the technical journal under the result (screenshot 2 below) shows the exact command Studio runs for you, line by line.

4. A dozen seconds later the run comes back Finished (screenshot 2 below) — succeeded, the duration, the tokens sent and received, the model, the output folder where the result landed, and the agent's tool calls in one line — and an Open the result button appears: it opens the output folder in Windows Explorer (screenshot 3 below), where your monitoring-summary.md awaits, with the three points and their sources.

The History screen (in the left navigation) keeps a trace of every launch. Drop new notes into the input folder and launch again: the team redoes the summary with the new content.

Where your team and its files live

Two locations, both on your side of the screen — nothing on online servers, nothing in the registry.

The team itself. When you click Save to my teams, Studio writes the team's definition — its mission, its agents, their rights — into a dedicated folder: C:\Users\<you>\Orkeon\teams\weekly-monitoring\ — the name you gave it, shortened into a folder name, so Weekly monitoring becomes weekly-monitoring (one folder per team, under your profile's Orkeon\teams folder). That folder is self-contained: copy it to another machine and it reappears in My agent teams; delete it and the team disappears from the list, leaving no litter anywhere else.

The input and output folders. Both live inside the team: C:\Users\<you>\Orkeon\teams\weekly-monitoring\input, where you drop the notes — the ones from the trial moved there at adoption — and …\output, where monitoring-summary.md lands. They exist as soon as the team does, and nothing needs declaring: folders that belong to the team are allowed by that alone. If you would rather read and write a folder you already keep elsewhere, choose Folder on this machine at the composition step and declare it first in Settings → Authorized folders — those are the only folders Studio refuses to open until they are declared, one by one.

What now? Learn at your own pace

Your first team will not be your best — and that is normal, even intended. While trying to build an elite team of agents, I ran into obstacles that kept me from reaching my goal on the first try. As Nelson Mandela said: "I never lose. I either win or learn." That apparent failure was a unique opportunity for analysis: the lessons drawn from the first red verdict — a framing too vague, outputs badly named — let me build a better-framed team, closer to what I expected from it, and more capable. Three concrete levers to get there faster:

1. Timestamp your outputs. Ask the assistant for a dated file name — monitoring-summary-2026-10-16.md rather than monitoring-summary.md. Each launch leaves a trace instead of overwriting the previous one: you keep the history, and your agents stop treading on each other.
2. Get help from a free LLM. Gemini, ChatGPT or another: describe your goal in your own words, then ask it to rephrase it as a structured mission — the team's role, what it reads, what it must produce, the pitfalls to avoid. Paste the result into Describe. A precise framing on the way in is a better team on the way out — and you learn to write better prompts along the way.
3. Replay the verdict. The trial is cheap and it does not flatter you: what it flags in red is exactly what needs fixing. One click on Fix and retry, and you move forward.

Above all, all this learning stays inside one single application: Orkeon Studio is one of the two interfaces of Orkeon — the simpler one, designed to get started. The other one, Orkeon Workshop, is for developers. What you learn here — framing a mission, structuring inputs and outputs, reading a verdict — carries over directly: the day you want to go further, the same tool is waiting for you, under a more expert interface. You move up without ever changing applications.

The natural next step is to give your agents a mailbox so the summary lands straight in your inbox — that is the subject of the second article in this series, which covers Gmail and Outlook/Hotmail.

Upcoming article: [Send an email with your Orkeon Studio agents]

Have a goat day 🐐

Keep reading

Publish a .NET package whose origin can be verified

dotnet add package is an act of trust in a stranger. For a package with a company behind it, the trust is in the compan…

My coding agent wrote outside the repo. Nothing stopped it.

You let a coding agent loose on a repository. It reads what it needs, then writes — a file two directories up, a ~/.co…

Augmented Outside-In Discovery with Claude Code

In a previous article available below, I walked through a structured approach to exploring new codebases. An Outside-I…

Subscription details

You are receiving this because you are a free subscriber to Goat Review.