Embedded workflow instructions for AI agents — enabling efficient web UI automation without repeated DOM analysis.
Korean Patent Application No. 10-2026-0071716 (filed 2026-04-21) · IETF Internet-Draft draft-han-ai-manifest-00 · MIT License · FRAND terms for patent claims
LLM-based AI agents (Claude Code, MCP clients, Puppeteer/Playwright-driven tools) currently parse entire DOMs or screenshots on every page, wasting tokens and failing on complex multi-step transactions. Website operators have no standardized way to declare "AI-Ready" workflows.
AI Manifest is a JSON specification that websites embed to tell AI agents exactly how to operate their UIs — step by step, with CSS selectors and actions.
Without manifest: AI analyzes full DOM/screenshots repeatedly → high token cost, frequent failures
With manifest: AI reads instructions → executes directly via CSS selectors → done
Benchmark: ERP-style 2-step order entry transaction, n=30 iterations, tiktoken(cl100k_base).
| Metric | Baseline (DOM analysis) | AI Manifest | Improvement |
|---|---|---|---|
| Mean input tokens | 1887.6 | 341.0 | −81.9% |
| Task success rate | 20.0% (6/30) | 100.0% (30/30) | +80.0 %p |
| Mean LLM calls | 1.4 | 1.0 | −0.4 |
Reproducible benchmark code lives in validation/.
Three embedding methods (sites can combine):
<meta name="ai-manifest" content="/.well-known/ai-manifest.json">Serve the JSON at /.well-known/ai-manifest.json per IETF RFC 8615.
Method B — Hidden DOM Element
<div id="ai-manifest" style="display:none" aria-hidden="true"
data-manifest='{"version":"1.0", ...}'></div>X-AI-Manifest: url=/.well-known/ai-manifest.json; hash=sha256:...
Agents fetch the manifest, compute its SHA-256 hash over a canonical (key-sorted, UTF-8) byte sequence, and verify against a central trust registry before executing steps[] sequentially.
A Central Trust Registry prevents malicious sites from injecting harmful instructions into AI agents (prompt injection defense):
AI Agent → detects manifest → queries registry.aimanifest.io/verify
↓
Trusted ✅ → Execute steps
Unknown ⚠️ → Warn user, fall back to DOM analysis
Blacklist ❌ → Refuse and alert user
The registry also performs static analysis of submitted manifests — detecting suspicious selector patterns (iframe injection, external form submissions), disallowed actions, and untrusted value sources — automatically blacklisting risky manifests.
See ai-manifest.schema.json for the full JSON Schema (Draft-07).
| Action | Description |
|---|---|
click |
Click a button or link |
fill |
Type text into an input field |
select |
Choose an option from a dropdown |
upload |
Upload a file to a file input |
wait |
Wait for a condition or timeout |
scroll |
Scroll to an element |
navigate |
Go to a URL |
assert |
Verify a condition before proceeding |
hover |
Hover over an element |
| Example | File |
|---|---|
| Academic manuscript submission (ScholarOne-like) | examples/academic-submission.json |
| SAP Material Master creation (MM01) | examples/sap-material-create.json |
| Standard | Purpose | Covers step-by-step UI? |
|---|---|---|
robots.txt |
Crawling permissions | ❌ |
llms.txt |
LLM-friendly documentation | ❌ |
agents.txt / agent-permissions.json |
Agent capability & permission declarations | ❌ |
OpenAI ai-plugin.json |
API integration manifest | ❌ |
ai-manifest |
Step-by-step UI workflow instructions with trust verification | ✅ |
This protocol is subject to:
Korean Patent Application No. 10-2026-0071716 · Filed 2026-04-21 · Applicant: Won-pyo Han
- Code in this repository (including the JSON Schema, validation harness, and examples) is licensed under the MIT License.
- Patent claims are offered under FRAND (Fair, Reasonable, and Non-Discriminatory) terms to promote adoption as an open standard. See
docs/FRAND.mdfor the full declaration. - Early publication of the Korean application has been requested (public publication expected within ~2–3 months of filing).
International counterparts (PCT, direct national filings) may be pursued within 12 months of the Korean priority date (until 2027-04-21).
- v0.1 schema and reference implementation
- IETF Internet-Draft:
draft-han-ai-manifest-00— posted 2026-04-21 - SDK: auto-generate manifests from existing web UIs (Python, Node.js)
- Central Trust Registry deployment:
registry.aimanifest.io - AI-Certified badge program for compliant websites
- Collaboration with W3C / IETF standardization tracks
Issues, schema proposals, and implementation PRs welcome. Please read docs/FRAND.md before contributing patentable improvements.
- Code & schema: MIT License (see
LICENSE) - Patent claims: FRAND terms (see
docs/FRAND.md)
Won-pyo Han · pk102h@naver.com