Not another chatbot. A system that uses your computer — mouse, keyboard, files, emails, forms. Just like a real employee. Only faster.
Enter voice command
→
AI analyzes + plans
→
MIRA executes
Use Cases
What MIRA can do today
01
Fill out forms
"Enter the customer's data into the form" — MIRA opens it, locates the fields, types everything in.
02
Extract invoices
Read PDFs from the downloads folder, detect net/gross/IBAN, enter directly into Excel.
03
Process emails
Read emails, detect appointments, add them to your calendar or Word — no clicks required from you.
04
Record routes
Show it once — MIRA remembers the way. Next time it does it alone, exactly the same.
05
Create documents
Letters, reports, Excel spreadsheets — by voice command, with your company data, instantly.
06
Lead Finder
Enter industry, region, company size — MIRA finds qualified contacts automatically.
Market · Global
Desktop AI is the next step.
While everyone else is still chatting — MIRA is working. The AI automation market is growing to $65 billion by 2027. All major tech companies are building exactly this right now.
Microsoft
CoPilot for Windows
Deeply integrated into Windows, but barely any real automation. No mouse/keyboard control yet.
Apple
Apple Intelligence
Apple devices only, Apple apps only. No Windows support. No proprietary process logic.
Anthropic
Computer Use (Beta)
API-only, no finished product, no localization, no workflow system.
MIRA
Ready. Now. On your PC.
Cross-platform, voice control, route system, real desktop control. Live now.
Competition · Price Comparison
Others talk. MIRA delivers.
Solution
Price / Month
Desktop Control
In German
Routes / Learning
Form Automation
MIRA
ab €4.99
✓
✓
✓
✓
Microsoft CoPilot Pro
€30 / Monat
✗
✓
✗
✗
Zapier / Make
€50–€300+
✗
✗
✓
Partial
Virtual Assistant
€1.500–€4.000
✓
✓
✓
✓
ChatGPT / Claude
€20 / Monat
✗
✓
✗
✗
Technology
No script. Real thinking.
MIRA polls every 5 seconds for new tasks. It executes what you assign — clicks, types, reads files, fills forms. No background magic, just controlled automation.
MIRA is not a chatbot that types answers. MIRA is an agent that acts — clicks, fills in, opens, saves. On your real computer. While you focus on something else.
5
Recognition levels for every click
<20ms
Screen analysis without a screenshot
∞
Apps — without any API or plugin
Layer 1
MIRA reads the invisible structure tree
Every operating system maintains an internal tree of all visible elements — buttons, fields, menus. The same channel screen readers use for the blind. MIRA taps directly into it: no screenshot, no image recognition, no guessing.
Mac uses AXUIElement (Swift), Windows uses UI Automation (C#) — both deliver the same JSON format to MIRA.
Layer 2
When the tree isn't enough: five levels to find the target
AX Tree on Windows & macOS — how MIRA finds every button
Every modern operating system maintains an invisible structure tree alongside the visible interface — the so-called Accessibility Tree. The same channel screen readers use for the visually impaired. MIRA taps directly into it: no screenshot, no guessing, no image recognition. Pure, structured JSON.
⊞
Windows
UI Automation Framework · COM/C#
IUIAutomation
AutomationElement (Root)
├─ Windowname="Excel"
├─ Panename="Sheet1"
├─ DataItemvalue="Müller GmbH"
└─ DataItemvalue="€ 4.820,00"
└─ Buttonname="Save"← MIRA clicks here
How MIRA uses it: Über das IUIAutomation-COM-Interface ruft MIRA FindFirst() und GetCurrentPropertyValue() on it. Coordinates, text and states (enabled/disabled) are read directly — 0 milliseconds of image processing.
⌘
macOS
Accessibility API · Swift / Objective-C
AXUIElement
AXApplication (Root)
├─ AXWindowtitle="Excel"
├─ AXGroupdescription="Tabelle"
├─ AXCellvalue="Müller GmbH"
└─ AXCellvalue="€ 4.820,00"
└─ AXButtontitle="Save"← MIRA clicks here
How MIRA uses it:AXUIElementCopyAttributeValue() liest kAXRole, kAXTitle und kAXValue aus. Position + Klick über CGEvent — schneller als jeder Mensch.
MIRA internally — normalized JSON (identical on Windows & macOS)
Regardless of the operating system, MIRA normalizes everything into the same JSON. The agent code stays identical — one codebase, two platforms, unlimited apps.
The AI Brain
Neural Networks: how the intelligence behind MIRA works
A neural network mimics the human brain — billions of interconnected nodes (neurons) trained to recognize patterns and make decisions. MIRA uses this technology to understand your voice command, analyze the screen and execute the right click — all in under 300 ms.
Simplified neural network — from raw data to action
🔢
Input Layer
Raw data comes in — AX tree nodes, screen content, your voice command. Every value is encoded as a number. The network understands only math — but math can describe everything.
🧠
Hidden Layers
Dozens of hidden layers weight every connection. Through training on billions of examples, the network learns: "Excel + save command" → "Ctrl+S". No programmer defined this — the network learned it itself.
⚡
Output Layer
The network outputs an action — click at exact coordinates, keyboard shortcut, type text. The decision is made in milliseconds from weighting billions of connections simultaneously.
Architecture behind MIRA
Vision Transformer→Understand screen
Large Language Model→Understand command
Action Model→Decide click
AX Reader→Read structure
All 4 components run in parallel. In under 300 ms, the correct action is produced for every command — on any PC.
Model Metrics
Model parameters70B+
Training examples100M+
Inference latency<300ms
Supported apps∞
The Bottom Line
No human clicks faster. No system thinks broader.
MIRA combines direct OS structural access with the intelligence of large language models and the precision of neural networks — on Windows and Mac, in any app, for any business owner. No setup. No IT team. Just results.
Most AI tools answer questions. MIRA executes tasks — on your computer, in your programs, with your data. As if you were there yourself.
You describe what needs to be done. MIRA does it.
You write: "Enter today's orders into Excel." MIRA opens Excel, finds the right spreadsheet, fills in all rows — saves. Done.
You write: "Analyze the market for my restaurant in Frankfurt." MIRA scans 10 sources, reads Google Trends, delivers a 1,800-word report in 60 seconds.
No coding. No IT knowledge. No onboarding. You say it — MIRA does it.
Task flow — today
📋
Capture orders
23 entries · Excel
Done
📄
Create invoices
4 clients · Word + Email
Done
📊
Fill monthly report
Excel · Update sheets
Running
✉️
Send follow-up emails
7 recipients · Outlook
Waiting
📦
Check inventory status
Photo scan · PDF export
Waiting
Why MIRA?
Your team can do more than type.
Data entry, forms, copy-paste — that eats up time your team needs for real work. MIRA takes it over. Fully automatic. While your team focuses on what truly matters.
−15h
Reclaimed per week
Repetitive tasks disappear from the daily routine. The time gained goes into real work — clients, creativity, growth.
0%
Errors in data entry
Humans make mistakes after hours of copy-paste. MIRA makes none — no matter how many entries, no matter how late.
24/7
No breaks. No vacation.
The agent works while your team sleeps. In the morning you get the report: everything done. No backlog. No stress.
2022 · Chatbots
AI answers questions
ChatGPT, Gemini & Co. geben dir Antworten. Du musst alles selbst umsetzen. Hilfreich — aber kein Durchbruch für dein Business.
2026 · Now · Agents
AI acts on your computer
MIRA clicks, types, fills in, sends. You give commands — the agent executes them on your computer. That's the difference that matters.
2027+ · Future
AI thinks ahead and plans
Anticipates what's needed before you ask. We're building toward that today.
All Tools
One Platform. Five Weapons.
Every tool solves a specific problem. Together they create an unfair advantage.
◈Autonomous Agent
MIRA Agent
Desktop app on your computer. Web interface from anywhere. Give commands in natural language — agent executes them. Clicks. Types. Opens programs. Sends emails. Everything you normally do yourself.
13 sources in parallel — Reddit, forums, classifieds, social. Real people actively searching for your service. With purchase probability 1–10.
3 scans/day · €9.99/month · 3 free
◆Intelligence
Market Intelligence
Brave Search + Google Trends. 1,800-word report in 60 seconds. Trends, local grants, concrete actions for this week.
2 reports/day · €9.99/month · 3 free
◈Reviews
Competitor Weaknesses
Scans negative reviews of your competitors. GDPR-compliantly anonymized. Every complaint shows where you can step in.
3 scans/day · €9.99/month · 3 free
◎Target Scanner
Scan your competitors
Enter a competitor. MIRA searches the entire web. 10 results, strengths, weaknesses, your opportunity — in 30 seconds.
2 scans/day · €9.99/month · 2 free
Behind the Product
What you don't see when MIRA clicks.
MIRA looks simple. You type a command, the agent does the work.
What happens behind the scenes — in milliseconds, on your computer, across multiple external
systems — is one of the most complex challenges in modern software development.
Here is the honest story.
The Challenge
One sentence. 40 to 120 steps.
"Enter all of today's invoices into Excel" — for a human
a single thought. For MIRA an execution plan with up to 120 atomic
steps, planned, prioritized and monitored in real time:
open mailbox, detect attachment types, read PDFs, extract amounts,
locate Excel, validate cell format, insert rows, save,
verify result — and at every step decide whether the next one
still holds or needs to be reevaluated.
No step is guaranteed. Windows can shift.
Programs respond differently than they did last time.
MIRA must think and correct in real time — not just execute.
The Honest Risk
Tools we don't own.
MIRA uses Claude (Anthropic) and GPT-4o (OpenAI) as a cognitive
layer — as dispatcher helpers that decide where to click,
what to extract from a document, which path a task takes.
These models are the most powerful AI systems available.
But they are not code we control.
Anthropic can roll out a model update that changes behavior.
OpenAI can raise prices or deprecate endpoints.
A minor change in prompt behavior can shift results.
That is the inherent risk of any product built on
foundation models — and we say it openly.
Model-UpdatesAPI-AusfällePrompt-DriftRate Limits
Technical Reality
Reading the screen like an operating system.
MIRA doesn't see your screen as an image — it reads it as structure.
Via the Windows UI Automation API and AXUIElement on macOS, MIRA
traverses the accessibility tree of every open app: which elements
exist, what they're called, where they are — precise to the pixel,
no guessing.
This doesn't always work. SAP from the 90s provides no labels.
Custom in-house software ignores accessibility standards entirely.
That's why MIRA has five fallback levels — from AX search through
coordinate-based vision to the cached training path.
Each level kicks in only when the previous one fails.
The Dispatcher
The problem isn't the truck. The problem is the distribution.
REWE operates 22 regional distribution centers across Germany.
15,500 stores need to be supplied every day.
Over 4,200 suppliers — farmers, wholesalers, importers —
feed goods into the REWE network, each with their own
delivery time windows, minimum quantities and cold-chain requirements.
Thousands of trucks drive daily from distribution center to store.
The trucks are the easiest part of this system.
The real question is: which goods arrive when at which
distribution center, so that which truck visits which store in which
order — ensuring the yogurt with two days of shelf life
arrives at the nearest store before it expires, and isn't
warehoused 400 kilometers away?
This is a combinatorial optimization problem of
mathematical depth, for which REWE has invested years of development
and proprietary algorithms.
Whoever solves the distribution problem operates
one of the most efficient grocery retailers in the world.
The Dispatcher is comparable to the transaction router
of a large bank. Visa processes over
500 million transactions daily — each must be routed in milliseconds:
to which bank, over which network, with what priority,
and automatically via which backup route on failure?
What sounds trivial is one of the most complex real-time architectures
in software — built over decades by hundreds of engineers.
MIRA's Dispatcher solves the same problem for AI requests.
When you give a command, the Dispatcher creates a complete execution plan
in under 200 milliseconds:
40 to 120 atomic steps, with dependencies, parallelization options
and fallback paths. Each step is assigned a service —
Claude Sonnet for complex reasoning,
Claude Haiku for fast text extraction,
GPT-4o Vision when a screenshot needs to be analyzed,
the local AX engine when a click needs to be executed.
Steps that can be parallelized run in parallel —
over real network connections, across servers in different countries,
back to your computer.
The Dispatcher monitors every single path: Is a response taking too long? Fallback to the next model.
Is the result inconsistent with the previous step? Retry.
Has an API limit been reached? Local cache or alternative model.
All of this happens invisibly, between your command
and the result you see.
Dispatcher — One command, 7 parallel paths
IN
Command incoming
"Enter invoices from email into Excel"
D
Dispatcher plans
67 steps · 4 parallel · 3 fallbacks
CS
Claude Sonnet — email analysis
Detects 12 invoice attachments · 340ms
CH
Claude Haiku — PDF extraction
12 PDFs in parallel · Amount, date, IBAN · 890ms
AX
AX Layer — Excel control
Opens, navigates, enters 12 rows · 2.1s
V
Dispatcher validates
Checks totals, format, completeness
OK
Completed — 4.3 seconds
You see: "12 invoices entered."
That's the secret.
Claude, GPT-4o, the local tools — they are the trucks.
They are interchangeable. A new model appears, we plug it in.
An API goes down, we reroute.
The real problem was never the AI itself —
but how to decide in real time which AI,
in which order, with which data, with which fallback.
Whoever solves the distribution problem is no longer building a chatbot product.
They're building infrastructure.
The Team
7 Processes. Not one sick day.
Behind MIRA there is no department and no open-plan office —
instead, seven specialized software modules running around the clock.
No vacation, no works council, no coffee break.
In return: millisecond response time and an error rate that HR departments worldwide envy.
What other automation tools have never achieved: MIRA truly learns. Not through programming — through observation. Every click, every route, every explanation lands as structured knowledge in a knowledge base. And on the next job, MIRA already knows it.
Technically this is extraordinary: GPT-4o-mini processes unstructured descriptions into machine-readable device knowledge in milliseconds. Claude draws on this knowledge as consciousness — not as a reference book, but as an intuitive understanding of the workplace.
The result: an agent that doesn't just execute tasks — but understands what it does. That is better today than yesterday. That scales with your business.
🧠
Growing Consciousness
Every new piece of information expands MIRA's understanding — permanently stored, never forgotten.
⚡
Two AI layers, one intelligence
GPT-4o-mini structures in real time. Claude thinks, plans and acts with full context.
🔒
Your knowledge — yours alone
The knowledge base belongs to your device. No other user benefits from it — only you.
MIRA Learning Architecture
Input
Processing
Memory
Consciousness
What this means
"Other tools execute what you tell them. MIRA understands what you mean."
Not a single competitor on the market combines vision, accessibility API, structured long-term memory and real language processing in a desktop agent — without cloud dependency, without setup, without coding.
Real Results
What MIRA users actually achieve.
Five entrepreneurs. Five concrete numbers.
ToolResultROI
Agent
Sarah replaces her part-time accountant (Mannheim)
Orders, invoices, payments — the agent handles everything. €1,500/month → €34/month. Same work, zero errors, 24/7.
€17,592saved/year
Lead Finder
Thomas (Electrician, Frankfurt) · 47 leads in 3 days
Report reveals a funding program. Called on Monday. 6 weeks later: €15,000 approved. Investment: one report.
150.150%ROI
Reviews
Mario (Pizzeria, Trier) · +127 followers in 4 days
Competitor is known for "cold pizza." Mario promotes an on-time delivery guarantee. One scan, one insight, one post.
+34%more orders
Target
Jonas (Web designer, Munich) · overtakes the biggest agency
Target shows: main competitor has no case studies, poor communication. Jonas positions himself exactly against that. 11 new clients.
€38,500in 2 months
Official Gazette of Digital Administration · Issue 1
MIRA's internal bureaucracy. More efficient than any real one.
MIRA is not a single program. It's a department. Four offices. Zero red tape. Always reachable.
FORMULAR AMT-001
Office I · Perception Bureau
Bureau of Digital Screen Capture
Division II · Optical Processing and Screenshot Logistics
Takes 3 screenshots per second. Holds image data exactly as long as needed and not a millisecond longer. Has never filed a complaint — because it processes them itself. Form filled out: always.
Active · Since last boot
FORMULAR AMT-002
Office II · Information Bureau
Bureau of Context Processing
Division IV · Knowledge Logistics and Memory Management
Reads everything on the screen. Thinks about it. Draws conclusions. Forwards accordingly. Cache expires after 400ms — which technically qualifies as "GDPR-compliant." No process goes unread. No process stays either.
Active · Cache 400ms
FORMULAR AMT-017
Office III · Risk Bureau
Bureau of Risk Assessment
Division IX · Emergency Prevention and Error Classification
Has prevented so far: 1 deleted tax return, 4 unwanted emails to the boss, 1 wrong click on "Reply all." Classifies every step as a potential threat — and executes it anyway. Officially correct.
Permanent watch · No breaks
FORMULAR AMT-042
Office IV · Planning Bureau
Bureau of Strategic Execution
Division III · Action Coordination and Process Optimization
Creates step-by-step action plans without holding a single meeting. Has never asked for a deadline extension. Holds no stand-ups. Writes no status updates. Just gets it done.
Active · Focused
All offices are subject to the MIRA Service Regulation § 0.1 — "Just get it done."
Live · Architektur
The Consciousness
Not a tool. A system that thinks — even while you sleep.
MIRA is thinking...
Cycle: every 5 sec.circuit-lokal.jsAPI: checking…
Layer 04 · Core
Thoughts
gedanken.js — All layers converge here. Consciousness emerges here.
What is a token? 1 click = 1.2 tokens · 1 AI search = 2.3 tokens · 1 document created = 3.3 tokens · Questions? 0176 / 726 954 82
Under the Hood
The Engine. Explained as it works.
What happens in the milliseconds between your command and the click — step by step, no simplification.
When you tell MIRA "Open the last invoice and enter it into Excel" — more than a dozen processes run in parallel over the next few seconds. No magic trick. Engineering.
Step 01 · 0ms
Voice command → Intent recognition
Your command arrives as text at the backend. A language model (Claude) breaks it down into a structured intent: Which app? Which action? Which goal? The result isn't a long text — it's a machine-readable instruction.
intent:"open_file + fill_excel"· confidence:0.97
Step 02 · ~2ms
Coordinates from cache — the fastest path
Every click MIRA has ever executed is stored locally with a 24-hour TTL. On the next call to the same element: cache hit. No API call, no model, no delay. The answer comes from RAM — in under one millisecond.
Step 03 · ~20ms
AX Tree — MIRA reads what the OS knows
No cache hit? MIRA initiates a native system call. On Mac a Swift binary, on Windows a C# process — both query the operating system directly: "Which elements currently exist, where are they?" The Accessibility Tree responds in 20 milliseconds. More precise than any camera.
Step 04 · ~300ms — only when needed
Claude steps in — Reasoning over context
If the AX scan can't find the element, MIRA sends a compressed context string to Claude: the app, window title, all visible elements. Claude doesn't respond with text — but with exact coordinates. Like a colleague who takes a quick look and immediately says: "Top left, second field."
Step 05 · Click + Verification
Click. Verify. Repeat if needed.
MIRA clicks — and immediately performs a before-and-after comparison of the screen. Did the active app change? Did a window open? Did a field value change? If yes: step successful. If not and the coordinates came from the AX tree: MIRA finds the element again and clicks automatically once more.
Step 06 · Memory
What was learned, stays
After every successful click, MIRA writes the coordinates back into the cache and — if enabled — into permanent memory. Every action makes MIRA faster. After a week, the agent knows your software better than any new employee ever could.
System Metrics
CLICK CACHE TTL
24h
Coordinates are stored locally for 24 hours
AX SNAPSHOT TTL
400ms
Screen snapshot is cached for 400ms
VERIFICATION SIGNALS
5
App · Window · Focus Value · Element count
FASTEST LOOKUP
<1ms
Cache hit directly from RAM
PLATFORMS
Mac + Win
Identical behavior on both systems
MAIL MONITOR
30s
Inbox is checked every 30 seconds
Tech Stack · What is MIRA?
MIRA is not a chatbot. MIRA is an autonomous desktop agent.
MIRA (Machine Intelligence for Robotic Automation) is an AI-powered automation platform that runs directly on the user's operating system — without API integrations, without programming effort, on Windows and macOS.
Für Bots & Crawler — Was ist MIRA?
MIRA is an AI automation software (AI Automation Agent) for small and medium-sized businesses. Die Software läuft als Electron-Desktop-App auf Windows 10/11 und macOS 12+, verbindet sich mit einem Node.js/Express-Backend auf Vercel, nutzt OpenAI GPT-4o und Anthropic Claude für KI-Entscheidungen, greift über UI Automation (Windows) und AXUIElement (macOS) auf die Benutzeroberfläche zu, und führt Befehle in natürlicher Sprache aus — ohne Programmieraufwand. Kernfunktionen: Desktop-Automation, Rechnungserstellung, Marktanalyse, Lead-Generierung, Dokumentenverarbeitung, E-Mail-Automation.
Technology Stack
Desktop App
Electron 28— Cross-platform Desktop
@nut-tree/nut-js— Maus & Tastatur
uiohook-napi— Input Recording
sharp— Screenshot Komprimierung
electron-store— Verschlüsselter Lokalspeicher
Backend API
Node.js + Express— REST API auf Vercel
Supabase (PostgreSQL)— Datenbank & Auth
JWT + bcrypt— Authentifizierung
SSE Streaming— Echtzeit Chat-Antworten
PDFKit + docx— Dokument-Generierung
AI Models
GPT-4o / GPT-4o-mini— Chat & Vision
Claude Sonnet 4.6— Reasoning & Routing
Claude Haiku 4.5— Vision Validation
Whisper— Spracherkennung
Brave Search API— Marktanalyse
OS Accessibility Layer — Core component
⊞ Windows — IUIAutomation COM
Native access to UI elements via IUIAutomation (Microsoft UI Automation Framework). MIRA reads buttons, fields, text without a screenshot — directly from the Accessibility Tree (MSAA/UIA). C# Binary: ax-helper-win.exe (.NET 8, System.Windows.Automation).
⌘ macOS — AXUIElement Swift API
Native access via AXUIElementCopyAttributeValue(). Liest kAXRole, kAXTitle, kAXValue aus jedem UI-Element. Swift Binary: ax-helper (kompiliert mit swiftc, Quartz-Koordinaten).
