商业产业 Hacker News AI 2026-07-20T22:01 正文 ok 379 words

🔶 Can AI agents use ur site?

英文标题:🔶 Can AI agents use ur site?

⬆️ 5 pts | 💬 9 comments | by manupareekk
打开原文相关 3 · 机会 1
media460 B media1.2 KB media1.2 KB media1.2 KB media43.3 KB

中文

⬆️ 5 pts | 💬 9 comments | by manupareekk

中文全文翻译正在补齐,可先查看下方英文原文。

English / 原文

OpenIngress

Self-hosted web operability analysis.
Crawl coverage, accessibility-tree inspection, and agent navigation break points.

MIT Python 3.10+ Vue 3

openingress.dev · Architecture · Quick start · Self-host


AI agents are starting to browse the web for people — opening pages, clicking buttons, filling forms. A site can look fine in a browser and still be hard for an agent to use: buttons without labels, pages that only load in JavaScript, links that never show up in the page structure agents read.

Being findable (SEO, citations, llms.txt) is not the same as being usable by an agent on the live site.

OpenIngress crawls a site like a browser would, inspects what an agent can actually see and click, then runs LLM-guided tasks against that crawl. The output is a report: which pages were reached, what broke, and where navigation stalled.

flowchart TB
  subgraph crawl["Crawl (Playwright)"]
    URL[Site URL] --> BFS[BFS page discovery]
    BFS --> CAP[HTML · screenshots · aria snapshots]
  end

  subgraph analyze["Analyze"]
    CAP --> STATIC[Static operability]
    CAP --> GRAPH[Navigation graph]
    STATIC --> GAPS[Gap taxonomy]
  end

  subgraph explore["Explore (BYOK LLM)"]
    GRAPH --> CATALOG[Page catalog]
    CATALOG --> JOBS[Exploration jobs]
    LLM[OpenAI-compatible API] --> JOBS
    JOBS --> TRACE[Job traces · blockers]
  end

  subgraph report["Report"]
    GAPS --> OUT[Coverage + break points]
    TRACE --> OUT
  end
Loading Stage Input Output
Crawl Public URL Page graph, DOM, screenshots, Playwright accessibility snapshots
Analyze Crawl artifacts Operability scores, label coverage, llms.txt, hydration checks, gap taxonomy
Explore Catalog + LLM key Reachability jobs, task attempts, blocker evidence
Report Analysis + job results Readiness metrics, break points, markdown export

Default crawl bounds: depth 3 , max 100 pages.

Layer Components
API Flask, Gunicorn
Automation Playwright (Chromium), Python 3.10+
Exploration OpenAI-compatible LLM (Azure, Gemini, Ollama, local proxy)
UI Vue 3, Vite
Persistence Local filesystem; optional Supabase
cp .env.example .env
cp backend/.env.example backend/.env
cp frontend/.env.example frontend/.env

# LLM_API_KEY required in backend/.env

make install
make backend    # terminal 1 — API
make frontend   # terminal 2 — UI

Configuration, Docker, and deployment: SELF_HOST.md.

Variable Required Description
LLM_API_KEY yes Provider key (OpenAI or compatible)
LLM_BASE_URL no Custom endpoint
LLM_MODEL_NAME no Default gpt-4o-mini
AUTH_DISABLED no Default 1

Runs and exploration require a configured LLM key.

Path Description
backend/ Flask API, crawl/explore workers, report generation
frontend/ Vue dashboard
docs/ Agent brief, Playwright recording notes
Dockerfile Playwright + Gunicorn container

MIT