Structured browser perception Runtime snapshots SiFR

Structured browser perception
for LLMs

E2LLM turns a live browser page into structured, model-readable context. Instead of asking an LLM to infer a UI from screenshots, selectors, or static HTML, E2LLM captures runtime state from the supported standard DOM and represents it as SiFR: a proposed Structured Interface Representation.

LLMs fail on web UIs when the perception layer is unstable or incomplete. E2LLM separates perception from reasoning: capture the browser state deterministically within supported scope, then let the model reason over a stable representation.

Why this matters

Browser agents are usually discussed as if they are one category. They are not. E2LLM focuses on a specific layer: structured perception of the actual runtime browser state, so the model receives context that is closer to what the page is doing.

Vision agents

See pixels. They interpret a rendered image of the page.

Accessibility / tree agents

See an accessibility abstraction of the page.

Structured runtime perception

E2LLM: structured perception of the actual runtime browser state.

What E2LLM captures

  • Runtime DOM structure from supported standard DOM surfaces
  • Element attributes and relationships useful to model reasoning
  • Visibility, actionability, disabled/loading state, labels, and relevant UI state where available
  • Snapshots that can be copied, saved, compared, or fed into LLM workflows

SiFR

SiFR is a proposed Structured Interface Representation for browser runtime snapshots. It is designed to make live UI state legible to models and automation workflows without relying on screenshots alone or forcing the model to reconstruct behavior from raw HTML.

Treat SiFR as a public/proposed representation, not an adopted industry standard.

Model ceiling, perception constant.

The core value is repeatable perception. For the same supported browser state, the capture should produce the same structured representation. Model output can still vary above that layer; the perception substrate should not.

Evidence

Public proof is collected in the evidence index: category taxonomy, the SiFR public/proposed spec trail, the product/tool surface, supported-scope limits, the repeatability trail, marketplace listings, GitHub/Dev.to surfaces, and independent category corroboration where accurate.

Open the evidence index

Runtime Snapshots

Runtime Snapshots is the public technical series behind E2LLM and SiFR. It documents the practical problem: LLMs need reliable browser-state context before they can reason well about web UIs.

Read the series index

Product surfaces

Browser extension

Capture runtime browser state as structured snapshots, directly from the page. Snapshots can be copied, saved, compared, or fed into an LLM workflow.

MCP tools

An MCP surface exposes structured capture and page-interaction tools to LLM clients that speak the Model Context Protocol, so an assistant can request a runtime snapshot of the supported standard DOM and reason over it.

Docs & prompts

Implementation details live in the product docs and the Runtime Snapshots series. Curated prompts for QA, accessibility, and security review are published openly.

Workflow use

Structured snapshots become workflow artifacts: capture the rendered state after a deploy, compare meaning rather than pixels, and surface UI regressions an LLM can describe in context.

Boundaries

What E2LLM is

  • Structured perception of live browser state, scoped to supported standard DOM surfaces
  • Deterministic capture within that supported scope
  • A model-readable representation (SiFR) of runtime UI state

What E2LLM is not

  • Not a managed browser cloud
  • Not a Playwright wrapper
  • Not an MCP transport layer
  • Not a claim that autonomous agents are safe or correct
  • Not a canvas, image, or shadow-DOM extractor

Public claims are scoped to supported standard DOM surfaces. Shadow DOM is a deliberate exposed-interface and privacy boundary; there is no public claim for shadow-DOM piercing, canvas perception, or image understanding. Stable perception does not imply autonomous-agent safety or output correctness.

Give your model a stable picture of the page

Install the extension, read the evidence index, or request a technical walkthrough for enterprise evaluation.