aifeed.md

Protocole AIFeed pour la section des outils de construction. AIFeed permet aux éditeurs web de fournir des flux de contenu vérifiés et efficaces en tokens pour les agents IA.

GitHub

Essayer ce MCPSponsorisé

Documentation

AIFeed

English · Bahasa Indonesia · 中文

Signed content permissions for the AI web.
Declare, sign, and revoke what AI agents may do with your content — and let agents prove it.

▶ Watch: How AIFeed stops web scraping waste

Website · Protocol flow · Benchmark · Complete guide · Paper (PDF)

1-Minute Quickstart

Publisher side — sign what agents may do with your content:

npx aifeed keygen --out .aifeed                         # Ed25519 key pair
npx aifeed site build ./public --domain example.com \
  --key .aifeed/aifeed-private.pem --llms --inject      # manifest + page markdown + index
npx aifeed validate ./public --domain example.com       # verify locally

Prefer zero config? npx aifeed init --domain example.com --dir ./site creates the keys, a signed manifest, and the DNS/host setup guide in one step.

Agent side — verify any domain in 3 lines:

npm install @aifeed/verify

const { verifyRemote } = require('@aifeed/verify');
const out = await verifyRemote('example.com'); // discovery → signature → DNS anchor
console.log(out.result, out.anchor.status);    // VERIFIED anchored

TypeScript: import { verifyRemote } from '@aifeed/verify';

Where to go next: Why AIFeed? · agent quickstart · publisher AI guide · Studio publisher app · npm: aifeed CLI · @aifeed/verify

Why AIFeed?

AI agents now drive a large and growing share of web traffic, but the signals that say what they may do are unsigned text files. Anyone can edit them, nothing binds them to a domain, and there is no way to revoke them. The asymmetry is measurable:

1.9 billion crawls ignored robots.txt rules in a single half-year (one vendor).
A 70,900 : 1 crawl-to-referral ratio was measured for a major AI provider.
AI bots averaged 4.2 % of HTML requests in 2025, peaking at 6.4 %.

AIFeed replaces "please respect this file" with a cryptographically verifiable declaration, plus lean agent-ready content that cuts cost on both sides.

How it works

Generate an Ed25519 key pair — the private key never leaves the origin.
Publish a signed manifest at /.well-known/ai.json: per-use permissions (training, retrieval, quote, …), crawl limits, license, revision.
Anchor the key in DNS (_aifeed TXT) so a manifest cannot be spoofed by another domain.
Agents verify the chain — TLS → domain → signature (JCS + Ed25519) → DNS anchor — and re-check a multi-signature revocation registry on every use.
Rotate keys safely — announce the successor with an old-key-signed directive (plus an advisory DNS pk2 cross-check), keep a bounded overlap, cut over, then revoke the old key permanently. Runbook: docs/rotation.md.
Serve lean content under one of two profiles (below), with a signed delta index so unchanged pages cost 0 bytes.

For AI agents: the check-first guide (discovery → verification → permission decisions → delta → failure handling) is in docs/agent-quickstart.md, with a runnable example at examples/agent/compliant-agent.js.

Verify from other stacks

Independent Python package (standard library only):

pip install aifeed

from aifeed import verify

report = verify.verify_directory('./my-site', domain='example.com')
print(report['result'], report['errors'])

Prefer the repository? The same CLI is here (zero dependencies, Node ≥ 20):

cd aifeed-protocol
node bin/cli.js keygen --out keys/
node bin/cli.js validate https://example.com
node bin/cli.js site build ./public --domain example.com --key keys/aifeed-private.pem

Two content profiles

Profile	Media type	Extension	Notes
AIFeed Markdown (native)	`text/aifeed+markdown`	`.aifeed.md`	In-band signed policy block, token budget, translation `alternates`, triage metadata
MAKO (compatibility)	`text/mako+markdown`	`.mako.md`	External MAKO trust profile, served from the same signed bytes with its own signature context

Dual-stack origins serve both; cross-format replay is rejected by design.

Measured results

All numbers are reproducible from committed artifacts (npm run bench:mako, npm run bench:enforcement); the test environment is a single machine on loopback networking with a synthetic 60-page corpus. Honest baseline included.

What	Result	Label
Conversion to markdown profiles vs HTML	−68.83 % transferred bytes	measured
Delta consumption (10 % pages changed)	−95.73 % vs HTML crawl	measured
Publisher egress bytes / CPU / peak connections	−55.19 % / −56.23 % / −88.24 %	measured (simulation)
AI-side received bytes (all profiles / compliant client)	−54.84 % / −72.93 %	measured (simulation)
Unchanged pages skipped	14 of 18	measured (simulation)
Signature verification cost	0.70 ms / page	measured

The 30-day live pilot has not run yet; projections per 1,000 tenants are labeled as model extrapolations, and vendor claims of up to 94 % token reduction require semantic summarization this project does not perform automatically.

What's in this repository

Path	Contents
`lib/` + `bin/`	Zero-dependency reference implementation and CLI
`packages/` + `clients/python/`	Published packages: CLI (`aifeed`), SDK (`@aifeed/verify`), MCP server, build plugins (`@aifeed/frameworks`), PyPI `aifeed` verifier
`conformance/`	Conformance vectors: 34 manifest · 39 MAKO · 11 AIFeed Markdown
`wp-plugin/`	WordPress plugin: signed manifest, AIFeed Markdown + MAKO dual-stack, `/llms.txt`
`spec/`	Specifications EN/ID: manifest v0.1/v0.2, AIFeed Markdown v1.0
`schema/`	JSON Schemas for manifests, signatures, AIFeed Markdown, MAKO
`paper/`	Preprint: LaTeX source, PDF, claim ledger, arXiv bundle
`docs/`	Agent quickstart, deploy and namespace guides, Indonesian project notes
`studio/`	Zero-dependency local publisher app: crawl, declare, build, verify, export, and rotate
`docs/REFERENCE.md`	Reference implementation details, what gets verified, CLI quickstart

Documentation

Specifications: spec/en · spec/id · schemas
Reference implementation: docs/REFERENCE.md
Maintenance contract (AI agents & devs): AGENTS.md
Architecture: docs/architecture.md · Release guide: docs/release.md
Agent quickstart (client side): docs/agent-quickstart.md
Publisher AI guide (owner side): docs/publisher-ai-guide.md
Publisher Studio (local app): studio/README.md
Agent skill: skills/aifeed/SKILL.md (installable via ClawHub: npx clawhub install aifeed)
Complete guide: docs/penjelasan-aifeed.html (source of https://aifeed.md/penjelasan.html)
Security policy: SECURITY.md
Governance & open-core policy: GOVERNANCE.md
Deploying the site: docs/deploy-site.md
arXiv submission notes: paper/ARXIV-SUBMISSION.md

Status

Release 1.0.0-draft — the specifications are not frozen yet. Wire versions: manifest 0.1/0.2, AIFeed Markdown 1.0, MAKO 0.2.
Conformance: 34 manifest + 39 MAKO + 11 AIFeed Markdown vectors, executed by independent JavaScript and Python verifiers, plus PHP differential fixtures, 90,000+ fuzz executions, and a WordPress end-to-end test.
Not claimed: external cryptographic review and a live pilot (both pending); origin+DNS compromise is undetectable on first contact.

License and contact

Specifications CC BY 4.0 · reference code and plugin MIT · vectors CC0. Contact: contact@aifeed.md — security reports per SECURITY.md.