AECS

AECS-1 (AI Email Consumption Specification) is an open standard for deterministically normalizing raw RFC 5322/MIME email into AI-ready JSON — the schema, the threading algorithm, the six content levels, and the rules for cleaning them without losing the sender's words. It is published under CC0 1.0 (public domain): implement it in any language, without asking permission.

@mvrx/aecs is the framework-agnostic, zero-infrastructure-dependency MIT reference implementation of that spec: parsing, content cleaning, threading, and message normalization, with no Cloudflare or storage assumptions baked in.

• AECS-1 specification — the normative document
• AECS-SDK-1 specification — the target SDK surface for implementations
• JSON Schema — machine-checkable NormalizedEmail shape
• Conformance suite — fixtures + an independent checker (verify.py)
• Docs site — browsable version of the above
• Storage guides — tables, keys and indexes for Cloudflare D1/R2, SQLite, PostgreSQL, MySQL, MongoDB and DynamoDB (examples/storage/)
• Changelog — what changed in each release

Install

npm install @mvrx/aecs

Usage

import { parse } from "@mvrx/aecs";

const email = await parse(rawRfc5322Message);

console.log(email.messageId);       // normalized Message-ID
console.log(email.threadId);        // deterministic thread identifier
console.log(email.content.clean);   // quote-stripped, signature-stripped body
console.log(email.content.forAI);   // wrapped, AI-ready representation

parse() accepts a raw message as a string, ArrayBuffer, Uint8Array, or a ReadableStream<Uint8Array>.

Content levels and cleanup

Field	What it holds
content.raw / content.text	The full body, quoted history included
content.clean	Quoted history and signatures removed. The sender's own lines are always kept, including answers written below or between quotes
content.forAI	clean, normalised and bounded for an LLM

Cleanup follows AECS-1 §4.3.1. If it would leave nothing, clean falls back to text and processing.cleanFallback is true. forAI is still a lossy view: if your app acts on the body, keep content.text as a fallback.

Typed decisions (Jev, Cloudflare Clef)

Ask decision models typed questions about an email and get typed answers with probabilities back, not free text. See AECS-SDK-1 §6.3.

import { clefProvider, decideEmail, jevProvider } from "@mvrx/aecs/decisions";

const { answers } = await decideEmail(email, clefProvider(env.AI), {
  route: { type: "choice", instructions: "Which team?", criteria: { billing: null, support: null } },
  urgent: { type: "noul", instructions: "Is this urgent?" },
});
answers.route.choice;   // "billing" | "support"
answers.urgent.noul;    // 0..1

License

MIT. See LICENSE.