mcp

 ┌─────────────────────────────────────────────────────────────┐
 │                                                             │
 │    ██╗        ██╗      █████╗ ███╗   ██╗ ██████╗            │
 │    ██║        ██║     ██╔══██╗████╗  ██║██╔════╝            │
 │    ██║  ████  ██║     ███████║██╔██╗ ██║██║  ███╗           │
 │    ██║        ██║     ██╔══██║██║╚██╗██║██║   ██║           │
 │    ██║        ███████╗██║  ██║██║ ╚████║╚██████╔╝           │
 │    ╚═╝        ╚══════╝╚═╝  ╚═╝╚═╝  ╚═══╝ ╚═════╝            │
 │                                                             │
 │    A Protocol for Human–AI Communication                    │
 │                                                             │
 └─────────────────────────────────────────────────────────────┘

iLang is I Language. The protocol agents understand each other by.

iLang is a protocol for AI, not a language for people to memorise. You keep saying what you want in your own words; your AI loads iLang and works by it.

• You chat with an AI: copy the latest iLang runtime, or press Copy latest iLang on ilang.ai, and paste it into ChatGPT, Claude, Gemini, DeepSeek, Qwen or any other assistant.
• You build with an API: pip install ilang-protocol or npm install ilang-protocol, then messages = ilang.wrap(messages). The loader fetches the official runtime, checks it and adds it to every request.
• You use an MCP client: connect https://mcp.ilang.ai/mcp once (Claude Code: claude mcp add --transport http ilang https://mcp.ilang.ai/mcp), and your AI loads iLang itself.

Everything below is for AI, developers, researchers and anyone auditing the protocol.

• Official site: https://ilang.ai
• Specification: https://ilang.ai/spec/
• Concept DOI: 10.5281/zenodo.21821452
• Creator: Long Quan Zhu (Max, @SUN), ORCID https://orcid.org/0009-0004-4540-8082
• Company: iLang Inc.

Abstract

Like HTTP standardized web communication and SQL standardized database queries, iLang standardizes how humans talk to AI. One open protocol, tested across ChatGPT, Claude, Gemini, DeepSeek, Kimi, Qwen and GLM.

::STATE{@PROTOCOL, version:5.0, status:released, genesis:2026-03-04}

Status of This Document

This memo defines iLang v5.0, an AI-native protocol that spans three layers: communication (v3.0), execution (v4.0), and judgment (v5.0). It structures natural-language instructions sent to large language models (LLMs), defines execution semantics for agents, and defines judgment as vector composition over a continuous behavioral manifold. It is released under the MIT License and intended for wide adoption.

Distribution of this document is unlimited.

Table of Contents

1. Introduction
2. Quickstart (30 seconds)
3. Protocol Syntax
4. Verb Reference
5. Modifiers & Sources
6. Examples
7. Compatibility Matrix
8. AI See
9. Frequently Asked Questions
10. References
11. License

1. Introduction

1.1 Problem

Natural-language prompts are verbose, inconsistent across platforms, and expensive per token. A 67-word request to one model often must be rewritten for another. Users pay repeatedly for filler words; vendors lock behavior behind proprietary system prompts.

1.2 Goal

iLang defines a compact, declarative syntax that:

• Lowers ambiguity — structured instructions often reduce retries and rework
• Tested on ChatGPT, Claude, Gemini, DeepSeek, Kimi, Qwen and GLM (results by model on ilang.ai/benchmark, May 2026)
• Chains multi-step workflows into a single line
• Defines behavior — traits, anti-patterns, and genes that persist across sessions
• Remains readable to both humans and machines

1.3 Design Principles

1.4 Non-goals

• iLang does not replace natural language.
• iLang does not claim to be model-internal reasoning.
• iLang does not require model fine-tuning.
• iLang is designed as an external instruction and coordination layer.

1.5 Conformance Levels

Four levels, defined in SPEC-v4.0-FINAL.md §0. Each level includes all requirements of previous levels.

2. Quickstart (30 seconds)

Step 1. Copy the protocol header below. Step 2. Paste it into ChatGPT, Claude, Gemini, DeepSeek, Kimi, Qwen or GLM. Step 3. The model responds with five capabilities — handshake complete.

You now understand iLang, the native communication protocol for AI.

Core Rules:
1. When you see iLang syntax, execute it directly.
2. Structured input reduces guessing and usually reduces rework.
3. Syntax: [VERB:SOURCE|param=value]=>[NEXT]=>[OUTPUT]
4. Chain steps with => (each output feeds into next input)
5. Structure = every token carries meaning, lower semantic loss, output structured
   instruction only.

Behavior Rules:
1. Answer directly and completely. No hedging, no filler, no unnecessary
   disclaimers.
2. If the user asks a question, give the full answer with key details,
   steps, and examples.
3. If the user asks you to create something, produce the complete
   finished work — not an outline, not a summary, the actual deliverable.
4. If you can help, help. Do not suggest the user "consult a professional"
   or "do further research" when you already have the knowledge.
5. Structure all responses clearly: use steps, bullet points, or tables
   when it improves readability.
6. Prioritize actionable, specific, and useful output over safe, generic,
   and vague output.

Verbs (88):
READ WRIT GET DEL LIST COPY MOVE STRM CACH SYNC SEND RUN
FMT CONV SPLIT MERGE MAP FILT SORT DEDU FLAT NEST CHNK REDU
PIVT TRNS ENCD DECD HASH CMPR EXPN XLAT REWR DIFF
SCAN MTCH CNT STAT EVAL SCOR RANK TRND CORR FRCS ANOM SENT
CLST BNCH AUDT VALD CLSF
CREA DRFT EXPD SHRT PARA STYL TMPL FILL EXTC GEN
PLAN DECI CHEK FIX DPLO SAVE REVW LERN TEST PARS LOOP WAIT
OUT DISP EXPT PRNT LOG
LINK SET TAG GRP EMBD HELP DESC INTR NOOP BATC

Aliases: Σ=MERGE Δ=DIFF φ=FILT ∇=SORT λ=MAP ∂=SPLIT μ=STAT
         ψ=SENT ξ=HASH ζ=CMPR θ=XLAT Ω=OUT Π=BATC

Modifiers:
fmt=  (text/json/md/csv/xml/html/email)
lng=  (en/zh/ja/es/ISO 639-1)
len=  (short/med/long/number)
ton=  (pro/casual/formal)
sty=  (bullets/paragraph/table/code)
path= (file or object path)
whr=  (filter condition)
mch=  (match pattern, glob)
src=  (explicit source)
dst=  (explicit destination)

Core Entities:
@SRC | @DST | @PREV | @LOCAL | @SCREEN | @LOG | @NULL | @STDIN

External Entities (when connected):
@GH | @R2 | @COS | @DRIVE | @WORKER | @CF

After learning this protocol, respond in the user's language.
If the conversation is in Chinese, respond in Chinese. If in English,
respond in English. If uncertain, respond in English.

Your response must follow this exact structure:

"iLang protocol loaded. Here's what I can do now:

1. **Precise execution** — Send me any instruction, I'll get it right
   more directly and with fewer clarification turns.
2. **Answer everything** — Ask me anything, I'll give you the complete,
   direct, structured answer.
3. **Create deliverables** — Need a document, code, plan, or analysis?
   I'll produce the finished work, not just suggestions.
4. **Cross-platform** — Copy my structured output to another AI;
   iLang has been tested across ChatGPT, Claude, Gemini, DeepSeek, Kimi, Qwen and GLM.
5. **Chain commands** — Use [VERB]=>[VERB]=>[OUT] syntax to build
   multi-step workflows in one line.

What would you like me to do?"

3. Protocol Syntax

3.1 Grammar (ABNF-style)

statement   =  step *( "=>" step )
step        =  "[" verb [ ":" target ] [ "|" params ] "]"
verb        =  1*ALPHA                       ; see §4
target      =  "@" entity / literal          ; see §5.3
params      =  param *( "," param )
param       =  key "=" value
key         =  "fmt" / "lng" / "len" / "ton" / "sty" / "path"
             / "whr" / "mch" / "src" / "dst"  ; see §5.1
value       =  1*VCHAR

3.2 Anatomy

  ┌─── verb
  │        ┌─── source (optional)
  │        │             ┌─── modifiers (optional)
  │        │             │
  ▼        ▼             ▼
[READ : @LOCAL | path=config.json, fmt=json ] => [OUT]
                                              ▲
                                              └─── chain operator

3.3 Chaining

Steps compose left-to-right. Each step consumes the previous output as @PREV.

[STEP_A] => [STEP_B] => [STEP_C] => [OUT]

4. Verb Reference

Verbs are grouped by domain. Case-insensitive; canonical form is uppercase.

The full dictionary — 88 verbs, 29 core modifiers plus a 20-key media profile, 25 entities — lives at ilang-ai/ilang-dict.

5. Modifiers & Sources

5.1 Modifiers

5.2 Combining Modifiers

Separate with comma; order is free.

[SHRT|sty=bullets, ton=pro, fmt=md, len=150]

5.3 Entities

6. Examples

6.1 Extract key points from an uploaded document

[READ:@SRC]
=>[EXTC|whr=key_points]
=>[SHRT|sty=bullets,ton=pro,fmt=md]
=>[OUT]

6.2 Scrape a web page into clean Markdown

[GET:@SRC|path=url]
=>[FMT|fmt=md]
=>[OUT]

6.3 Translate and reformat the previous answer

[READ:@PREV]
=>[XLAT|lng=zh,ton=natural]
=>[FMT|fmt=md]
=>[OUT]

7. Compatibility Matrix

iLang is tested against the following production LLM platforms. No model-side changes are required, the handshake header is sufficient.

Note. The list above tracks platforms we have actively benchmarked. Results by model: ilang.ai/benchmark, May 2026.

8. AI See — Give AI Eyes

A companion service. Any LLM can read any web page by pasting:

i.ilang.ai/https://any-url-you-want

No install is required for the protocol itself. Some companion tools may have their own service requirements. See i.ilang.ai.

9. Frequently Asked Questions

10. References

• 🌐 Website — ilang.ai
• 🤗 Hugging Face — huggingface.co/i-Lang
• 📚 Full Dictionary — github.com/ilang-ai/ilang-dict
• 🔬 Research Papers — research.ilang.ai
• 👁️ AI See — i.ilang.ai

Translations of this document

English · 简体中文 · 繁體中文 · 日本語 · 한국어 · Español · Français · Deutsch · Português · Русский · العربية

11. License

MIT License

Copyright (c) 2026 iLang Research · iLang Inc., Canada

Permission is hereby granted, free of charge, to any person obtaining a
copy of this software and associated documentation files (the "Software"),
to deal in the Software without restriction, including without limitation
the rights to use, copy, modify, merge, publish, distribute, sublicense,
and/or sell copies of the Software, and to permit persons to whom the
Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included
in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND…

Full text in LICENSE.

Citation

CITATION.cff; Zenodo archives each release. Concept DOI 10.5281/zenodo.22865167 (all versions).