AAI (Agent App Interface)

An open protocol that makes any application accessible to AI Agents.

---

Background: The Agent-Software Paradox

Early 2026 revealed a paradox: AI-driven market panic ($285B wiped in a day) vs. NVIDIA CEO Jensen Huang's rebuttal—"AI will use tools, not replace them."

The truth lies in between: AI needs software tools, but only those it can access. Applications invisible to Agents face obsolescence—not because AI replaces software, but because Agents will choose accessible alternatives over undiscoverable ones.

The Future of Applications

Tomorrow's applications will serve two audiences:

Interface	Audience	Purpose
GUI	Humans	Visual interaction, discovery, enjoyment
AAI	Agents	Programmatic access, automation, orchestration

Both interfaces access the same core logic. The GUI remains for humans who prefer visual interaction or need to explore. AAI enables Agents to operate apps on behalf of users—reliably, quickly, and at scale.

This is not about replacing apps. It's about making every app accessible to the AI assistants that users increasingly rely on.

---

The Problem

AI Agents (Claude, GPT, OpenClaw, etc.) can already operate well-known applications like Apple Mail or Microsoft Outlook -- because LLMs have seen their AppleScript/COM interfaces in training data.

But what about your app?

• Desktop apps: Even if your app supports AppleScript, COM, or DBus, LLMs have never seen your documentation. They don't know your command names, your parameters, or your data model. Your app is invisible to Agents.
• Web Apps: Your service has a REST API, but Agents don't know your endpoints, auth flow, or request format. Without a standardized descriptor, your API is just another undiscovered URL.
• Apps with no automation at all: Completely unreachable.

The Solution

AAI gives every application -- desktop or Web App -- a standardized, machine-readable descriptor (aai.json) that tells Agents exactly what your app can do and how to call it.

Without AAI:
  Agent knows Apple Mail        ✅  (in LLM training data)
  Agent knows Microsoft Word    ✅  (in LLM training data)
  Agent knows your desktop app  ❌  (never seen it before)
  Agent knows your Web App API   ❌  (never seen it before)

With AAI:
  Agent discovers your app via aai.json  ✅
  Agent reads tool definitions           ✅
  Agent calls it directly (native binding, stdio, or API) ✅

One aai.json file turns your app from invisible to fully Agent-accessible.

Who Benefits

App Developers -- Make Your App Agent-Ready

Your App's Situation	Without AAI	With AAI
Well-known desktop app (Mail, Outlook)	Agents already know it	Agents discover it formally
Web App with REST API (Notion, your Web App)	Agent doesn't know your endpoints	Agent discovers API tools via aai.json
No automation / no API at all	Completely unreachable	Add interface + aai.json, Agent-ready

The key insight: Having an API or automation support is not enough. Without a standardized descriptor, Agents have no way to discover your app's capabilities. AAI is the bridge between "having an interface" and "being Agent-accessible."

Agent Developers -- Zero Integration Work

Connect to any AAI-enabled app through standard MCP. No per-app custom code, no scraping documentation, no hardcoding commands. Works for both desktop apps and Web Apps.

Users -- Agent-Driven Productivity

With AAI, users delegate daily work to AI Agents instead of operating applications manually. Compared to traditional UI automation:

• Higher reliability: Structured IPC/API calls are deterministic, unlike screenshot-based automation
• Faster execution: Direct communication bypasses visual recognition and mouse simulation overhead
• Seamless integration: Agents orchestrate multiple apps naturally, creating end-to-end workflows

How It Works

1. App provides aai.json describing its tools:
   • Desktop apps → bundled inside the app at YourApp.app/Contents/Resources/aai.json
   • Web Apps → hosted at https://yourdomain.com/.well-known/aai.json
2. Agent connects to AAI Gateway via standard MCP (stdio)
3. Gateway scans /Applications/ on startup and builds a list of AAI-compatible desktop apps
4. Agent discovers available apps and tools on demand; web app descriptors are fetched automatically when needed
5. Gateway executes the call:
   • Desktop apps → JSON over native bindings (Apple Events / DBus / COM), OS-managed authorization
   • Local adapters → JSON over stdio
   • Web Apps → JSON over HTTPS, Gateway-managed OAuth 2.1 authorization

Both humans (via GUI) and Agents (via AAI) access the same core application logic. Neither interferes with the other.

Contact

gybeml@gmail.com

License

Apache 2.0