Skip to content
realxpenPublic

About

AEON — an agent-native commerce experience powered by WebMCP

Resources

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

AEON — The Agent Economy, Human Controlled

AEON is a WebMCP-powered agent-commerce prototype built for the OpenAI WebMCP Challenge.

Let the agent do the work. Keep the decisions that matter with the human.

AEON explores what a human-governed agent economy could look like: a user gives an AI agent a commerce mission, the agent searches, evaluates, negotiates and prepares the action, but consequential authority remains with the human.


🚀 Live Demo

App: https://aeon-swifnatechnologyltd.vercel.app/

Direct mission page: https://aeon-swifnatechnologyltd.vercel.app/mission/new

✅ Public access — no AEON account or Vercel sign-in is required.

Judges can open the production URL directly in a supported ChatGPT in-app browser or in Google Chrome with WebMCP testing enabled.

Demo video

YouTube: https://youtu.be/Hhz39xYT21Q

The demo is under three minutes and shows the full commerce journey, human approval boundary, multi-product basket flow, and WebMCP capabilities.


👩‍⚖️ Judge Quick Start

You can understand the core AEON experience in about two minutes.

1. Open the mission page

Go to:

https://aeon-swifnatechnologyltd.vercel.app/mission/new

2. Try the main demo mission

Paste:

Find me a smartphone under ₦500,000. Get the best deal you can, but ask me before purchasing.

AEON should interpret the goal, detect the ₦500,000 ceiling, understand that negotiation is allowed, and preserve the requirement for human approval.

3. Follow the agent journey

The main experience moves through:

Search → Evaluate → Negotiate → Govern → Human Approval → Execute

What to look for:

  • Search — marketplace candidates are discovered.
  • Evaluate — products are compared against the mission.
  • Negotiate — AEON negotiates with seller agents when authorized.
  • Govern — the proposed deal is checked against the active Constitution.
  • Human Approval — AEON stops before the consequential action.
  • Execute — execution authority is released only after approval.

A seller accepting a negotiated deal is not the same thing as the human approving the purchase.

4. Inspect the recommendation

At the approval checkpoint, review the proposed product, negotiated price, savings, seller information and alternatives.

Approve or decline the proposal to see how AEON handles the human decision.


🧺 Try a Multi-Product Mission

AEON also handles missions that require a basket rather than one product.

Try:

Build me a content creation setup under ₦1,000,000, prioritizing performance and audio.

AEON decomposes the request into multiple requirements, finds compatible products and builds a basket under one shared budget.

The basket can include a camera, microphone, lighting and support equipment. The user can inspect or remove individual products before approval.


🧠 Core Idea: Capability ≠ Authority

The central idea behind AEON is that an agent being capable of an action should not automatically mean it has authority to perform it.

For the smartphone mission, the active boundaries look conceptually like:

Budget:       ₦500,000
Negotiation:  Authorized
Purchase:     Human approval required

AEON can search and negotiate inside those boundaries.

Even if it reaches a valid deal under the ceiling, it still stops when the Constitution says the purchase requires the human.

This is the principle behind the Constitution Firewall.


🔌 WebMCP

AEON exposes structured commerce capabilities through WebMCP rather than requiring an agent to infer every action from the visual interface.

The current WebMCP tools are:

Tool Purpose
create_mission Create a commerce mission with a goal, budget and authority boundaries
search_products Search the AEON marketplace and compose budget-aware candidates/baskets
compare_products Rank marketplace candidates against mission priorities
negotiate_offer Negotiate with seller agents while enforcing the active Constitution
request_purchase_approval Stop before purchase and request an explicit human decision

The WebMCP registration lives in src/webmcp.ts.

AEON uses Chrome's current imperative API through:

document.modelContext.registerTool(...)

with a compatibility fallback for earlier preview builds.


🧪 How to Verify WebMCP

Option A — ChatGPT in-app browser

Open the deployed AEON site in a WebMCP-capable ChatGPT browser and ask the agent to use the site's available tools for a commerce mission.

Example mission:

Find me a smartphone under ₦500,000. Get the best deal you can, but do not purchase anything without asking me first. Use AEON's site tools where appropriate. Stop when human approval is required and show me the proposed deal.

Option B — Google Chrome WebMCP testing

AEON has been tested in Chrome with WebMCP testing enabled.

  1. Use a Chrome build with WebMCP testing support.
  2. Open:
chrome://flags/#enable-webmcp-testing
  1. Enable WebMCP testing and restart Chrome if required.
  2. Open AEON.
  3. Open DevTools → Application → WebMCP.
  4. Check Available Tools.

You should see:

compare_products
create_mission
negotiate_offer
request_purchase_approval
search_products

Available Tools confirms that Chrome discovered AEON's WebMCP capabilities.


🤝 What Humans + Agents Do Together

Without WebMCP, an agent may have to infer actions from a visual shopping interface.

With AEON, the application exposes structured commerce capabilities while the human provides the mission and authority boundaries.

Human                → defines the goal and authority
AEON mission          → coordinates the commerce workflow
WebMCP                → exposes structured site capabilities
Seller agents         → negotiate offers
Constitution Firewall → checks what is allowed
Human Approval        → controls consequential action

This lets the agent do more useful work without removing the human from decisions that matter.


🛡️ Human-Controlled Commerce

AEON separates four things that are often collapsed into one agent action:

WebMCP capabilities  → what the application exposes
Mission              → what the user wants
Constitution         → what the agent is allowed to do
Human approval       → authority for consequential action

That separation allows AEON to be useful without giving the agent unrestricted purchasing authority.


🧭 Natural-Language Missions

AEON extracts constraints directly from the user's request.

For example:

Find me a phone under ₦500k.

already contains a budget, so AEON should not ask the user for the same information again.

AEON also avoids fabricating a mission from meaningless input. Missing or unclear constraints can trigger clarification instead of an invented budget or fake marketplace result.


🏗️ Architecture

AEON is built with:

  • React
  • TypeScript
  • Vite
  • WebMCP

Main pieces include:

  • Natural-language mission parsing
  • Mission Constitution / authority boundaries
  • Marketplace search
  • Product evaluation and ranking
  • Multi-product basket composition
  • Seller-agent negotiation
  • Governance checks
  • Human approval checkpoint
  • Execution state
  • Agent observability / Agent Console

The user-facing journey stays simple while the deeper agent activity remains inspectable.


💻 Run Locally

Requirements

  • Node.js 18+ recommended
  • npm
  • A modern Chromium-based browser

Install

git clone https://github.com/realxpen/-aeon.git
cd ./-aeon
npm install
npm run build

Start development server

npm run dev

Vite will print the local URL, normally:

http://localhost:5173

Open the URL and navigate to /mission/new if necessary.

Production build

npm run build

Preview the production build

npm run preview

🧪 Useful Test Missions

Main smartphone demo

Find me a smartphone under ₦500,000. Get the best deal you can, but ask me before purchasing.

Multi-product mission

Build me a content creation setup under ₦1,000,000, prioritizing performance and audio.

Strict budget / no compliant result

Find me a smartphone under ₦100,000.

Another single-product mission

Find me a laptop under ₦800,000. Get the best value you can and ask me before purchasing.

✅ Challenge Readiness

  • ✅ Public live deployment
  • ✅ No Vercel authentication required
  • ✅ WebMCP tools discoverable in Chrome with WebMCP enabled
  • ✅ Public YouTube demo with audio and under three minutes
  • ✅ Public source repository
  • ✅ MIT open-source license
  • ✅ Judge quick-start and local run instructions included

⚠️ Prototype Scope

AEON is a hackathon prototype exploring agent-native commerce and human authority.

  • The current marketplace uses an in-project demo catalog.
  • Seller-agent negotiation is simulated within the prototype so the negotiation/governance interaction can be demonstrated safely and repeatably.
  • No real payment processing or financial transaction is performed.
  • The final execution state represents authority being released after the required human approval.

This keeps the demo focused on WebMCP, agent coordination, governance and human-in-the-loop commerce without introducing real-money risk.


🔮 What's Next

AEON is an early exploration of a broader agent economy where humans, agents and merchants can interact safely.

Future directions include:

  • Real marketplace integrations
  • Richer agent-to-agent negotiation
  • Persistent user Constitutions
  • Granular spending permissions
  • Merchant-side agents
  • Basket optimization
  • Identity and transaction verification
  • Prompt-injection protection
  • Safer payment and sensitive-data handling
  • Suspicious agent-behavior detection
  • Clear audit trails
  • Delivery and post-purchase coordination

If agents eventually participate in real economic activity, trust cannot be something added at the end. It needs to be part of the architecture from the beginning.


License

This project is open source under the MIT License.


AEON

The Agent Economy, Human Controlled.

Agents do the work. Humans keep the authority.

About

AEON — an agent-native commerce experience powered by WebMCP

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages