Web Context Interface (WCI) makes sites natively readable for LLM-based agents through
structured data-wci-* markup, a semantic Distiller Engine,
and a typed Interactivity Protocol — all in <8 KB.
The same web — two radically different representations.
| Feature | 🧑 Human-Friendly (Standard) | 🤖 Agent-Friendly (WCI) |
|---|---|---|
| Structure | Deeply nested, CSS-dependent divs | Flat, schema-typed node list |
| Navigation | Visual cues — icons, colors, hover | Descriptive data-wci-desc |
| State | Implicit — URL changes, CSS classes | Explicit ActionResult payloads |
| Noise | High — ads, navbars, hero images | Low — only functional nodes |
| Errors | Red border, toast notification | Typed error.code + error.hint |
| Context cost | 50–500 KB raw HTML | 1–10 KB distilled JSON/MD |
| Side effects | Re-parse full DOM | sideEffects[] in ActionResult |
| Auth | Implicit redirect flow | Auth-Required in wci.txt |
| Scoping | None — full DOM exposed | Landmark scopes bound reasoning |
Interact with the form or click the Agent Actions to dispatch typed commands and watch the distilled output update in real time.
Click to dispatch typed actions via WciBridge:
Loading…Loading…[]
Paste browser-interpreted DOM (for example from Amazon) and infer an WCI-like view
even when there are no data-wci-* tags.
Open the target page in your browser console and run:
copy(document.documentElement.outerHTML)
then paste it below.
Ready.
Metrics will appear here after parsing.
No page loaded yet.No page loaded yet.
Multi-step task grounding on 50 scenarios (OpenRouter).
See evals/README.md for methodology and limitations.
Multi-step grounding on 50 scenarios — all five evaluation approaches.
Performance comparison across 50 multi-step grounding scenarios evaluated over six LLMs (OpenRouter, May 2026).
See evals/README.md for methodology and limitations.
| Model | Success rate (%) | Tokens per step (est.) | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Raw HTML | DOM outline | Candidates | WCI full | WCI grounding | Raw HTML | DOM outline | Candidates | WCI full | WCI grounding | |
| Loading benchmark results… | ||||||||||
Like robots.txt — but for agents.
Three files at the web root give LLMs site-wide grounding before they touch a single page.
wci.txt
Directive file — allow/deny scopes, rate limits, auth requirements, human-confirmation guards. Machine-parsed by the agent runtime.
View file ↗wci.json
Structured site manifest — capabilities, authentication flows, multi-step task flows, scope descriptors. For tool-calling agents.
View file ↗wci.md
Rich Markdown narrative injected into the LLM's system prompt. Explains purpose, allowed actions, task flows, and error recovery.
View file ↗Loading site context files…
data-wci-* Attribute Reference12 attributes. Zero breaking changes. HTML5-compatible.
| Attribute | Type | Description |
|---|---|---|
data-wci-id | string | Stable unique identifier for this node |
data-wci-role | enum | action · form · display · nav · status · landmark |
data-wci-desc | string | LLM-optimised description of what this node does |
data-wci-action | enum | click · fill · select · check · upload · submit · navigate |
data-wci-state | JSON | Current observable state snapshot |
data-wci-precondition | string | Natural language guard condition |
data-wci-required | boolean | Whether this field must be satisfied before form submission |
data-wci-options | JSON array | Enumerated choices for select/radio/checkbox groups |
data-wci-emit | string | Custom DOM event fired after interaction |
data-wci-scope | string | Parent landmark scope ID |
data-wci-hidden | boolean | Flag decorative nodes for Distiller pruning |
data-wci-priority | 1–5 | Importance ranking within scope (1 = primary CTA) |