AIFeed is a content permission protocol, toolset, and integration solution for the AI web. Its core goal is to enable publishers to declare, sign, and revoke, in a cryptographically verifiable way, what AI agents may do with their content, and to enable agents to prove that they have complied with these declarations.
Purpose AI agents currently drive a large and growing share of web traffic, but the signals describing what agents may do are usually just unsigned text files. Anyone can edit these files, they have no binding relationship to the domain, and they cannot be revoked. AIFeed replaces the "please respect this file" approach with cryptographically verifiable declarations, while also providing streamlined, agent-oriented content to reduce costs for both sides.
Main capabilities
- Keys and signatures: Generate Ed25519 key pairs, with the private key never leaving the origin server.
- Signed manifest: Publish a signed manifest at /.well-known/ai.json, including permissions divided by purpose (training, retrieval, citation, etc.), crawl restrictions, licenses, and revision versions.
- DNS anchoring: Anchor keys via _aifeed TXT records, so that the manifest cannot be forged by other domains.
- Verification chain: Agents verify TLS → domain → signature (JCS + Ed25519) → DNS anchoring, and recheck the multi-signature revocation registry on every use.
- Key rotation: Announce successor keys through instructions signed by the old key (plus a recommended DNS pk2 cross-check), maintain bounded overlap, and permanently revoke the old key after the switch is complete.
- Streamlined content configuration: Provide two content configurations (AIFeed Markdown and MAKO), together with a signed incremental index, so that unchanged pages consume 0 bytes.
- Cross-stack verification: Provide an independent Python package (standard library only) as well as a zero-dependency CLI for Node ≥ 20.
How to use Publisher side: When zero configuration is preferred, run npx aifeed init --domain example.com --dir ./site to create keys, a signed manifest, and DNS/host configuration guidance in one step. Agent side: In TypeScript, you can import { verifyRemote } from '@aifeed/verify'; to verify any domain in 3 lines. The agent's check-priority guide (discovery → verification → permission decision → incremental → failure handling) is located at docs/agent-quickstart.md, runnable examples are located at examples/agent/compliant-agent.js. The publisher AI guide is located at docs/publisher-ai-guide.md, and the local publisher application is located at studio/README.md. The agent skill is located at skills/aifeed/SKILL.md and can be installed via ClawHub: npx clawhub install aifeed.
Applicable scenarios It is suitable for website publishers who want to declare content usage permissions to AI agents in a verifiable way, and for AI agent developers who need to comply with these permissions and prove their own compliance. It is also suitable for engineering teams that need to verify signed manifests and content vectors across languages (JavaScript, Python, PHP).
Limitations and notes The current version is 1.0.0-draft, and the specification has not yet been frozen. Wire format versions: manifest 0.1 / 0.2, AIFeed Markdown 1.0, MAKO 0.2. Conformance testing includes 34 manifest, 39 MAKO, and 11 AIFeed Markdown vectors, executed by independent JavaScript and Python validators, plus PHP differential fixtures, 90,000+ fuzz executions, and one WordPress end-to-end test. What is explicitly not claimed: external cryptographic review and live pilots are both pending; when the origin and DNS are compromised at the same time, first contact cannot be detected. The 30-day live pilot has not yet been run; projections per 1,000 tenants are marked as model extrapolations; the vendor-claimed maximum 94% token reduction requires semantic summarization, and this project does not automatically perform that operation. All measurement results can be reproduced from committed artifacts (npm run bench:mako, npm run bench:enforcement); the test environment is a single-machine loopback network, uses a synthetic 60-page corpus, and includes honest baselines.
Repository contents Specification: spec/en, spec/id, schemas; reference implementation: docs/REFERENCE.md; maintenance contract: AGENTS.md; architecture: docs/architecture.md; release guide: docs/release.md; agent quick start: docs/agent-quickstart.md; publisher AI guide: docs/publisher-ai-guide.md; publisher Studio: studio/README.md; agent skill: skills/aifeed/SKILL.md; full guide: docs/penjelasan-aifeed.html; security policy: SECURITY.md; governance and open core policy: GOVERNANCE.md; site deployment: docs/deploy-site.md; arXiv submission instructions: paper/ARXIV-SUBMISSION.md.
License and contact The specification uses CC BY 4.0, the reference code and plugins use MIT, and the vectors use CC0. Contact email: contact@aifeed.md; for security reports, please follow SECURITY.md.
