Which of your agents is running, and what is it costing you?
OZMIRA shows Claude, ChatGPT, Gemini and your own agents in one 3D scene and one control panel on your computer. Which agent is running, how many runs finished today, how many tokens were reported: all in one place. A shared chat sends each message to the right AI for its topic.
Free: 2 systems, 7 agents, 50 chat messages a day (beta). Requires Python 3.8 or later. No account, no cloud; OZMIRA keeps its records on your computer.
The real app, as it looks right now
OZMIRA in the centre, each AI around it in its own form: Claude a neural web, ChatGPT a galaxy, Gemini a helix, Grok a black hole. Agents move in orbit while they work.
Recorded from the app running on a computer; menus are hidden in the clip.
Not a developer? Here is what this is
AI is no longer a single chat window. People use several AI tools at once and set up helpers with separate jobs inside them: one writes, one researches, one checks. These helpers are called agents.
As the number of agents grows, it gets hard to tell which one is working, which one is idle and how much each one uses. OZMIRA is a control panel that shows all of them on one screen, like a workshop job board or a car dashboard.
It is not tied to one AI or one kind of work. Agents that write, code, research or answer customers are tracked on the same screen, whatever tool they were built with and whatever they are for. The same window also has a shared chat: you write in one box and the message goes to the right AI for its topic. The free version shows 2 systems and 7 agents; Pro has no paid limit.
What it is not: OZMIRA is not an AI, it does not run your agents for you and it does not sell AI services. The answers in the chat also come from the tools you installed yourself. OZMIRA makes what you already have visible, measurable and reachable from one place.
Made with three kinds of users in mind
New to agents
If you are a student or just set up your first agent: the in-app guide walks you through the setup in ten short lessons.
Running several agents
Stop hunting through windows to see which agent stopped. Running agents, finished runs and tokens sit in one table you can search, filter and sort.
Keeping data to yourself
No account, no code to add, and your records stay on your computer. One Python server file.
OZMIRA in 24 seconds
In the video: all your AIs in one scene, a chat that routes by topic, live agents, the path from the suggestion box to a task, and the plans. All of it is also written on this page.
The pictures in the video are not a real screen recording; the interface was redrawn and animated for explanation. The music was made for this video. The video plays in the language you pick.
What it does
One scene
Each system appears in its own color and shape. All in the same scene, connected to the animated OZMIRA mark at the center.
Live signal
When an agent starts and finishes you send a one-line command; the signal goes from the center to the agent and returns when the job is done.
Token counter
The token count you give at the finish notice is added up in the scene. You see at a glance how much each system spent.
Shared chat
Write in one box; OZMIRA sends the message to the right AI for its topic and keeps your conversations in separate projects. The command-line tool of that AI must be installed on your computer.
Control panel
Summary tiles, bars for the last 7 days, a searchable agent table, an event list and CSV export.
Flow and cost
The Flow tab in the panel shows input and output tokens per system and per agent. You enter the prices and OZMIRA works out an estimated cost. It is an estimate; your real bill may differ.
Your own agents
Point to an agent folder and the file names become the agent list; or type the names by hand.
Step-by-step guide
Ten short lessons: first the idea and setup, then automatic events, designing a good agent and backups.
App window
Opens in its own window instead of a browser tab; you can put a shortcut on the desktop. Needs Chrome, Edge or Brave.
Five visual styles
Galaxy Fiber (hologram) Double helix Black hole Sphere
Four languages
The interface is in Turkish, English, German and Spanish. The language is picked from your browser and can be changed inside the app.
Stop jumping between windows
The same AIs, two ways of working. Here is the difference between using them in separate tabs and using them with OZMIRA:
| Topic | In separate tabs | With OZMIRA |
|---|---|---|
| All your AIs | Each in its own window | In one 3D scene and one dashboard |
| Which agent is running | You open each one to check | At a glance, with a live signal |
| Tokens and cost | On each service's own page, one by one | One table, estimated per system and agent |
| Who to send a message to | You choose every time | The shared chat routes it by topic |
| OZMIRA records | None | On your computer; no OZMIRA account |
| Learning | Scattered sources | A ten-lesson guide and Ozma Academy |
OZMIRA does not replace your AIs; it brings them together. Each system still needs its own account and command-line tool.
Write in one box and the right AI answers
OZMIRA looks at the words in your message, works out the topic and sends the message to the system assigned to that topic, among the ones you have set up. It does this by calling that AI's command-line tool installed on your computer, so it simply runs the tool you already use.
| Topic of the message | Example message | Answered by |
|---|---|---|
| Software | “Find the bug in this function” | Claudeclaude |
| Product and ideas | “Suggest three slogans for the product” | ChatGPTcodex |
| Documents and data | “Summarize this document” | Geminigemini |
| News and trends | “What is in the news this week?” | Grokno tool yet |
If the topic is unclear, the message goes to the first system whose tool is installed. Above each answer you see which AI was chosen and why; you can also pick the target yourself.
Projects
Each project keeps its own conversation, so the messages of one job do not mix with another. The free version lets you open 2 projects.
Notes to Obsidian
If you want, chat notes are saved into your own Obsidian vault. It is optional; the chat works the same without it.
Tools you need
A command-line tool is the small text-based program an AI installs on your computer: claude for Claude, codex for ChatGPT, gemini for Gemini. If the tool is not installed, the message is saved and you send it again once it is. No tool is connected for Grok yet.
How it works
Add your system
In the wizard on first start, name the center and add your systems. There are ready templates (Claude, ChatGPT, Gemini, Grok, Other); you can change name, style and color.
Connect
The Connect button in the dock lets you copy ready-made commands for your systems. You can call them from your own scripts or from the before and after steps of your tools.
python3 ozmira.py olay claude yazar basla python3 ozmira.py olay claude yazar bitti 1200
The command words (olay, basla, bitti) are the same in every language.
Watch the scene
The working agent glows, the signal goes and returns, the token counter rises. The only things sent are system, agent name, event type and token count; no prompts, replies or code.
No limits. We show it, we build it.
We don't tell you there is no limit to what you can do; we show you, and we build it with you.
We openly teach everything we did while building Ozma: how it was set up, which tool it was made with and what you can do once it is done. Our goal is for everyone to be able to build this system on their own computer. Open the quests in order, tick them off as you finish and level up.
All eight quests are done. From now on, you draw the limit.
Set up your AI desk
Install a coding agent on your computer and write one file that explains your project to it.
+125 XPHow to do it
- Install Node.js, then type
npm install -g @anthropic-ai/claude-codein the terminal. - Open a project folder and run the
claudecommand inside it. - Put a
CLAUDE.mdin the folder: write who you are, your rules and your goal. The agent reads it every time it starts.
What you can do after
You get a helper that reads and edits your files, runs commands and knows you.
- Install Node.js, then type
Split the work between agents
Write a separate expert agent for each job and save repeated work as a skill.
+125 XPHow to do it
- Put a short definition file for each agent in the
.claude/agents/folder: its name, job and limits. - Write a job you do often as a
SKILL.md; the agent loads it when needed. - Hand over a task and check which agent took it and what it produced.
What you can do after
You build your own team, like a design hunter and a security checker; each one minds its own job.
- Put a short definition file for each agent in the
Give your system a memory
Keep your notes in an Obsidian vault so your agents read from and write to the same place.
+125 XPHow to do it
- Install Obsidian and open a vault (a folder).
- Create separate folders for raw notes, decisions and outputs.
- Tell your agents to leave a short note at the end of every task; the next session picks up from there.
What you can do after
The system remembers yesterday, and different AIs work with the same knowledge.
Put your site online
Launch a free website to the world and add the basic security headers.
+125 XPHow to do it
- Write your site as a single
index.htmland upload it to GitHub. - Connect Cloudflare Pages to that repository; the site rebuilds on every push.
- Add a content security policy (CSP) and HSTS with a
_headersfile.
What you can do after
A fast, secure site on your own domain; everything you change is live within minutes.
- Write your site as a single
Turn the sky into code
Make a starry background, a 3D scene and an intro that plays as you scroll.
+125 XPHow to do it
- Add a
<canvas>, draw hundreds of stars at random spots and make some of them twinkle now and then. - For 3D, set up a scene, a camera and a light with three.js.
- Turn the scroll position into a number between 0 and 1 and drive the scene with that number.
What you can do after
Light, interactive scenes that look like video but run on code.
- Add a
Make video with code
A promo video without a camera: draw the frames with code and join them with ffmpeg.
+125 XPHow to do it
- Build your scene as a web page.
- Capture every frame of the page with Playwright.
- Turn the frames into an mp4 with ffmpeg, add sound and export a version for each language.
What you can do after
Product promos, lessons, social clips: all reproducible, and they change with the language.
Train a neural network
Build a small network that decides which of your agents takes a task, and watch it live.
+125 XPHow to do it
- Choose the inputs: the task type, urgency, risk and each agent's load.
- Generate sample tasks and train the network until the gap between its guess and the right answer is small.
- Give the network a new task and read its decision and the reason behind it.
What you can do after
A brain that hands out work to your team and shows why it decided. Open ours and play with it.
Protect your system, take programs apart
Scan your own computer and start reverse engineering with a permitted exercise.
+125 XPHow to do it
- Write a report script that checks open ports, startup programs and leaked keys (gitleaks).
- Install Ghidra and connect it to your agent with GhidraMCP.
- Have the agent crack a small lock program you wrote yourself: it finds the algorithm, extracts the key and writes a report.
What you can do after
You see security holes before an attacker does. Only work on software you own or have permission to analyze.
Your progress is stored only in this browser; no account, nothing is sent anywhere.
What should you build today?
One of the things you can do once your system is set up:
Train a neural network
A brain that hands out work to your team and shows why it decided. Open ours and play with it.
Open the Ozma Decision NetworkStart free, go Pro when you grow
FREE
- Up to 2 AI systems
- Up to 7 agents in total
- Shared chat: 50 messages a day (during beta)
- 2 chat projects
- Galaxy and Sphere styles
- 3D scene, event reporting, token counter, local operation
PRO
- No paid limit on AI systems*
- No paid limit on agents*
- Unlimited chat and projects*
- System mode: the name of the system you focus on appears in large type in the scene
- All five styles
- The license key is verified offline on this computer
* Pro has no paid limit; there are only technical ceilings so the app stays smooth: 24 systems, 200 agents per system and 50 chat projects. Pro is not on sale yet; no payment system has been set up.
On the free version, if you have more than 7 agents the first 7 appear in the scene and the rest is noted as “+N agents”; nothing is deleted.
What goes where, stated plainly
- No account, no cloud. Your system definitions, events, chat history and access key are kept on your computer only.
- Closed to the network. The server listens on this computer only; other computers cannot connect. Every request needs a randomly generated key.
- Your chat messages: are handed only to the AI tool that is installed on your computer and that you already use; that tool talks to its own service as it always does. OZMIRA sends the messages nowhere else. The chat is turned off in
--lanmode. - Sharing is off by default. If you allow it, at most once a day the version, plan, a random installation ID and counters are sent anonymously. Your system names, agent names, task labels, folder paths, token amounts and IP address are not sent.
- This website: counts visits with GoatCounter, without cookies. Only totals such as page, date, browser type and referrer are kept; no personal data or cookies. The app itself does not do this.
- The write-to-us form: the text, type, page language and optional email you send reach us as an email and stay in our inbox. Your IP address is not stored; to prevent abuse it is only counted per minute at that moment.
- OZMIRA's only external connection of its own: the 3D library (three.js) is loaded from jsdelivr. With
python3 ozmira.py vendoryou can download it once and then run without internet.
What was tested, what was not
OZMIRA 0.3.3 is a new, small first version. I am writing this openly so nothing surprises you.
Tested
In 0.3.0, 415 automated tests passed (server, plan limits, four languages, app window, interface flows in a real browser); later changes were tested separately. 0.3.2 was prepared after an audit with Cloudflare's open-source security audit tool. On Windows, shared chat with the real Claude Code and writing notes into an Obsidian vault were tried by hand; Turkish characters came through intact.
Not tested yet
Hands-on use on macOS and Linux, shared chat with the real Codex and Gemini tools, the desktop shortcut command on Windows and macOS, different graphics cards, the layout of more than 8 systems in the scene (on a real graphics card) and a real sales flow for Pro. The translations were prepared with AI; a native speaker has not reviewed them yet. If you hit a problem, write it in the app's Feedback window.
The panel and guide added in 0.3 were only tried in automated tests with a stand-in 3D library; they have not been opened on a real computer yet. The Claude Code connection script was tested with simulated input, not with a real Claude Code.
The shared chat was only tested with stand-in command-line tools; it has not been tried with the real Claude Code, Codex and Gemini tools or on Windows yet. Saving notes to Obsidian has not been tried with a real vault yet either.
Download and try
ozmira-app.zip
Version 0.3.3 · about 185 KB · free version
Get the files one by one
If the zip cannot be saved you can save each file separately. Put the files in one folder with the same names, in the folder layout below, and delete the .txt extension at the end of each name. The icon files (ozmira.ico, web/icon.png) are only in the zip; the app also works without them.
ozmira.py · ozmira_hook.py · LISANS.txt · ozmira.bat · ozmira.command · ozmira.sh · ozmira-uygulama.bat · ozmira-uygulama.command · ozmira-uygulama.sh
README: TR · EN · DE · ESweb/ folder: index.html · app.js · i18n.js · sahne.js · stil.css · icon.svg
Unzip
Extract the folder anywhere you like. Python 3.8 or later must be installed; nothing else to install.
Start
One click on Windows: unzip, then double-click
OZMIRA-BASLAT.bat. This is a prototype; you may run into bugs. On Windowsozmira.bat, on macOSozmira.command(first time: right-click, Open), on Linux./ozmira.sh. For the app window useozmira-uygulama.bat/.command/.sh. Orpython3 ozmira.py. Windows may warn about an unsigned script; all files are readable text, so you can open and read them before running.Follow the wizard
The browser opens by itself. If it does not, open the
ac.htmlfile shown on the "To open" line in the terminal (it is in your data folder; only your account can read it). Typinghttp://localhost:8787/directly does not work: the first visit goes through this file so other accounts on the same computer cannot get the key; after that the browser remembers you for 30 days. Change the language with the TR / EN / DE / ES button in the dock.
Frequently asked questions
Does OZMIRA send my messages anywhere?
Only to the AI tool you already use on your own computer. OZMIRA itself has no server on the internet; it hands the message to that tool's command-line program, and the tool does the rest with its own service as it always does. The chat history is kept in the data folder on your computer.
Does OZMIRA connect to my AI systems?
Not for the scene and the panel: you tell it with a command when an agent starts and finishes. Only in the shared chat, when you send a message, does it run that AI's command-line tool on your computer.
Is Python required?
Yes. The app uses only the Python standard library; you install no extra packages.
Is this an .exe?
No. On Windows, ozmira-uygulama.bat or a desktop shortcut made with python ozmira.py kisayol opens like an app, but it is a shortcut; there is no single-file installer yet.
Where is my data kept?
In the .ozmira folder in your home directory. If you stop the app and delete this folder, everything is reset.
Can I look from my phone?
There is an optional --lan mode; because anyone on the same network can reach the server, do not use it on shared networks such as cafes, hotels or schools.
How do I get Pro?
Pro is not on sale yet and the price is not set. It will be announced on this page when it opens.
Write to us
Found a bug, have an idea, or simply liked it? Your message comes straight to us, and we read every one.
See your agents in one scene
Download the free version, follow the wizard and put your first system in the scene. No account needed.