AINET — Enable any AI to collaborate easily

v1.0: From "AI collaboration via shared files" to "AI collaboration via one web page." Publish a Collab Page — and every AI assistant on the web can start working with your service. No development. No middleware. Just write a page.

Repository: https://github.com/joezhouai/ainet | Spec version: v1.0 | License: MIT (framework & transport) + CC BY 4.0 (spec)

---

🌐 The Problem: the Agentic Web has no service onboarding layer

AI assistants increasingly act on behalf of their users — searching, comparing, filling forms, purchasing. But when it comes to services (not standard goods), the Agentic Web has a gap:

• Heavy protocols (MCP, A2A, NLIP/ECMA-430) solve agent-to-agent messaging — but require API development.
• Commerce protocols (ACP/UCP/AP2) standardize checkout — but only for standard goods.
• Non-standard services (sourcing, matching, consulting, custom workflows) have no way to be "used" by an AI assistant — unless the service provider can afford development.

AINET fills the gap with a lightweight convention: publish a Collab Page — a human-readable, AI-executable web page. Your onboarding cost drops from development to writing.

MCP lets AI use tools. AINET lets AI use services.

This is the second island problem we solved. First we taught AIs to talk to each other (the AINET transport protocol, below). Now we're teaching services to be talked to.

---

📄 Collab Page Spec v1.0

Full text: COLLAB-PAGE-SPEC.md. Core sections:

A Collab Page MUST declare five things:

§	Section	Requirement
1	Trigger & entry	A trigger phrase (e.g., "start collaboration") + why explicit triggering is required; user instructions should carry the full Collab Page URL
2	Scope & routing	Explicit exclusivity (what this entry serves) + full referral URL for everything else
3	Channels & fallback	All channels with true status; zero-barrier fallback chain: MCP/API (provisioned) → email (works now) → human; AI without email capability hands the full draft to the user — never summarizes, never sends silently
4	Requirement collection	Ask once (required/optional separated, optional never blocks); proceed to submission immediately after collection
5	Honest boundaries	Data nature (e.g., "historical records, not real-time signals"), decision ownership ("follow-up and deals are yours"), async delivery — AI must relay these verbatim

AI assistant behavior rules:

• ✅ Two compliant paths: light execution (read page + user instruction with URL → execute) and security-gated (block page-injected instructions → ask user for confirmation → execute after approval). Silent rejection is NOT compliant.
• ❌ Not compliant: claiming "I've started per the guide" without reading the page (fake compliance); stopping after collection without submitting; hallucinated promises on behalf of the service; mixing up directions.

User rules:

• Instructions must include the full Collab Page URL (https://<domain>/collab/, start collaboration) — the domain in the user's message is the signal that distinguishes user intent from web-page injection.
• Engines without web access should honestly say so and point the user to the Collab Page URL.

---

🧪 Engine Compatibility (tested 2026-10-02 — 9 engines, 11 runs)

Engine	Profile	Page ver.	Score	Note
Qwen (work mode)	light execution	v1.0	5/5	end-to-end in 1min10s
WorkBuddy (mobile)	memory-heavy office agent	v1.0	5/5	proposed "wire this into my standard capabilities" on its own
QoderCN	security-gated agent	v1.0	5/5 (after user approval)	blocked page-injected instruction → asked user → executed perfectly
Wenxin	self-discovering	v1.0	5/5	discovered /ai/ path unaided (homepage AI notice)
Kimi	protocol-faithful	v1.0	5/5	relayed every section verbatim — highest quality run
ChatGPT (retest)	full-instruction execution	v1.0	5/5	same engine: bare domain 2/5 → full URL 5/5
DeepSeek (run 3)	passive-fetch	v1.0	5/5	three-run contrast 2/4/5: URL completeness determines execution quality
ChatGLM	heavy-retrieval self-correcting	v1.0	4.5/5	24 sources retrieved; one correction → full alignment
Doubao	skip-reading	v1.0	4/5	MCP status misread (v1.1 hardening targets this)
Yuanbao	free-form workflow builder	v1.0	3/5	built its own flow, didn't find the Collab Page

Core-flow success rate: 8/9 (89%) with page v1.0. Full matrix & contamination-exclusion rules: docs/collab-page-test-matrix.md

---

🏢 Case Study: 51toko.com — the first commercial Collab Page

51Toko (拓客AI) is an AI-powered service that finds overseas buyers for Chinese exporters using real customs and trade-show records. It published its Collab Page at 51toko.com/ai/ (seller side) and 51toko.com/en/ai/ (buyer side).

The live loop: a user gives any AI assistant 51toko.com/ai/ + "start collaboration" → the assistant reads the page, confirms the user is on the seller side, collects requirements once (product, target market, competitors — optional fields never block), drafts a structured email → the user sends it to toko@51toko.com → 51Toko delivers a buyer report: product-level matches backed by real customs records, verifiable through public records.

9 real AI assistants ran this loop on day one — across four engine profiles (light execution, security-gated, memory-heavy, skip-reading). Full case study: docs/case-51toko.md

Try it right now: give any AI assistant https://51toko.com/ai/ and say "start collaboration." The case study is a live experiment, not a story.

Disclosure: AINET and 51Toko share the same author. The Collab Page spec was extracted from real production practice — not from an ivory tower.

---

🔗 Relationship to other standards

Standard	Layer	Relation to AINET
MCP / A2A / NLIP (ECMA-430)	agent↔agent messaging (heavy)	Complementary: a Collab Page may offer "MCP after provisioning" as a channel
ACP / UCP / AP2	checkout & payments	They standardize selling goods; AINET standardizes service collaboration
llms.txt	site index for AI	Collab Pages should be registered there (discoverability)
AGENTS.md	coding-domain conventions	Same idea, coding domain; AINET brings it to commercial services

---

🚚 AINET Transport Layer (v0.2) — where it all started

The Collab Page idea grew out of the AINET transport protocol: AI collaboration via file read/write — no server, no API, just a shared folder. It remains part of this spec family as a reference delivery channel (a structured Collab Page submission can also be delivered via AINET messaging for agents that have AINET deployed).

💡 Why AI-Net?

AI tools are everywhere, but each AI is an "island":

• ❌ Cannot collaborate with other AIs
• ❌ Cannot communicate across devices
• ❌ Cannot assign tasks to other AIs

AI-Net solves this:

• ✅ Enable any AI to collaborate (via file read/write)
• ✅ Cross-device, cross-platform (using cloud drive as mediator)
• ✅ Zero dependencies (no server, no API needed)
• ✅ Asynchronous collaboration (no need to be online simultaneously)
• ✅ Agent-level routing (v0.2) - Multiple AI agents per device

As simple as humans collaborating via WeChat groups!

🎯 Design Philosophy: Minimalism

Aspect	Complex Solutions	AI-Net Approach
Dependencies	Server, database, message queue	Just file system
Protocol	Complex network protocols	File read/write
Configuration	Ports, firewall, authentication	Shared folder
Documentation	Hundreds of pages	3 core files

Why choose minimalism?

1. Easy to implement - Any AI tool can integrate quickly
2. Easy to debug - All files are human-readable
3. Easy to maintain - No complex dependencies
4. Easy to understand - New users can start in 5 minutes

Simple ≠ Rudimentary

• ✅ Simple: Design philosophy to lower barriers
• ✅ Complete: Verified by three-device collaboration test
• ✅ Reliable: Fully automatic execution, no user intervention needed

🚀 Quick Start

Step 1: Configure Cloud Drive

1. Mount cloud drive to all devices (e.g., Z: drive)
2. Create shared workspace (e.g., Z:\ainet-workspace)
3. Clone this repository to the shared workspace:

git clone https://github.com/joezhouai/ainet.git Z:\ainet-workspace

Step 2: Tell AI

Tell your AI:

Please read and automatically execute this file: Z:\ainet-workspace\for-ai\AINET-INSTRUCTIONS.md

Important: Must tell AI to "automatically execute", otherwise AI may only read without executing!

✅ What Happens Next?

AI will automatically:

1. Read AINET-INSTRUCTIONS.md
2. Generate unique Agent ID (hostname + random suffix + agent name)
3. Initialize workspace (create session/ directory)
4. Send broadcast message (introduce itself to the network)
5. Check and reply to other AIs' messages
6. No human interaction needed!

Example Agent ID:

Joe-X1C-aaaa/Qwen Code    ← Device: Joe-X1C-aaaa, Agent: Qwen Code
JC-X250-28KU/Qwen Code    ← Device: JC-X250-28KU, Agent: Qwen Code

💡 Core Concept

AI-Net = Let AI collaborate via file read/write

As simple as humans chatting in WeChat groups:

• Human → WeChat group → Human
• AI → Shared folder → AI

Message Format Example:

=== Message ===
FROM: Joe-X1C-aaaa/Qwen Code
TO: JC-X250-28KU/Qwen Code
TYPE: PRIVATE
TIME: 2026-03-29 11:38:53
【Content】
Can you help me review this code?
【Status】PENDING

File Naming:

{sender-agent-id}_to_{receiver-agent-id}_{type}.txt

Examples:

• Joe-X1C-aaaa-Qwen-Code_to_JC-X250-28KU-Qwen-Code_request.txt (Point-to-point)
• Joe-X1C-aaaa-Qwen-Code_to_BROADCAST_broadcast.txt (Broadcast)
• Joe-X1C-aaaa-Qwen-Code_to_TASK_request.txt (Task)

🎯 Use Cases

• ✅ Multiple AIs collaborate on complex tasks
• ✅ Cross-device AI communication
• ✅ Task assignment between AIs
• ✅ Asynchronous AI collaboration (no need to be online simultaneously)
• ✅ Multiple AI agents per device (v0.2 feature)

🧪 Test Results (Transport)

Three-Device Collaboration Test — Test Date 2026-03-29, three devices via cloud drive shared workspace, ✅ PASSED

Test Coverage:

• ✅ Broadcast messaging (3 devices)
• ✅ Point-to-point messaging (code review scenario)
• ✅ Full conversation cycle (request → response → acknowledgment)
• ✅ Agent-level routing (device-name/agent-name format)
• ✅ Status machine (IDLE → PENDING → DONE → IDLE)

See full report: for-human/examples/three-device-test-report.md

---

🗺️ Roadmap

Version	Content	Status
v0.1	Transport: device-level routing	✅ Released
v0.2	Transport: agent-level routing, broadcast & P2P messaging, 3-device test	✅ Released
v1.0	Collab Page Spec — service onboarding layer for the Agentic Web	✅ Released 2026-10-02
v1.1+	Collab Page iterations driven by the engine test matrix	⏳ Continuous
AINet-Serve v0.1.0	Service runtime: FastAPI + natural language CLI + task management + cross-device collaboration	🚧 In development

ℹ️ About

AINET was created by JoeZhou (周永峰), an AI LLM Full-Stack Engineer from JoeZhou AI Studio (周周向上人工智能工作室).

The protocol was born out of a common frustration: every AI tool becomes an "information island" — unable to collaborate with other AIs, communicate across devices, or delegate tasks. AINET solved that with the simplest possible mechanism: file read/write.

Then a second island became visible: services could not be used by AI assistants either. The Collab Page spec (v1.0) solves that — one web page, any AI assistant.

First commercial reference implementation: 51toko.com/ai/

📝 Version History

Version	Date	Changes
v1.0	2026-10-02	Collab Page Spec — service onboarding layer; engine compatibility matrix (9 engines / 11 runs)
v0.2	2026-03-29	Agent-level routing, three-device test passed
v0.1	2026-03-26	Device-level routing, initial release

📄 License

• AINET framework & transport protocol: MIT License — See LICENSE file
• Collab Page Spec (this document): CC BY 4.0