Skip to main content

HTTP

The agent fetches each pattern from this site at generation time — nothing to vendor locally, always current. The skill ships as a single SKILL.md plus a small local catalog; the pattern corpus itself lives on the docs site and is pulled over HTTP only for the patterns a given task actually selects.

What gets fetched, and from where​

HTTP is a retrieval mechanism, not a different brain. The skill's selection logic is identical to every other variant; only the retrieval step changes. Two of the three inputs stay local — only the per-component pattern pages travel over the network:

  • Catalog — local. The skill reads patterns.json from its own directory (${CLAUDE_SKILL_DIR}/patterns.json) to select pattern IDs. It is not fetched over HTTP.
  • Pattern pages — remote. For each selected ID, the skill issues GET https://a11y-context-project.vercel.app/web/react/components/{id} — all fetches in parallel, each URL at most once per session — and extracts Must Haves, Don'ts, Golden Pattern, and Customizable from the response.
  • Foundations rules — local. The skill reads global_rules.md from its own directory and applies each rule by scope.

So HTTP mode is local catalog + local Foundations rules + remote pattern pages. The base URL is baked into the skill; there is nothing to configure.

Install​

Download: a11y-context-web-react-http.zip ↓

Unzip into your tool's skills directory — the same one-step install as Local:

BASH
unzip a11y-context-web-react-http.zip -d .claude/skills/

The skill folder holds three files and no bundled corpus:

your-project/
├── .claude/
│   └── skills/
│       └── a11y-context-web-react-http/
│           ├── SKILL.md          # brain: invocation + selection + fetch step
│           ├── patterns.json     # local catalog (selection only)
│           └── global_rules.md   # local Foundations rules
├── src/
└── package.json

Install locations for other tools (Cursor, Copilot, Codex, …) are on the Downloads page.

When to pick HTTP​

HTTP is the low-friction default whenever the agent has network access. It's the cheapest retrieval path measured and needs no local corpus to maintain:

  • Cheapest retrieval. In a controlled experiment, HTTP used roughly 30–40% fewer subagent tokens than Local or MCP (p < 0.001) — the agent pulls only the pages it selected, nothing more.
  • Always current. Pattern updates land on your next request; there's no bundled copy to refresh and no catalog_revision to track.
  • Smallest footprint. Three small files in your repo, no corpus to vendor.

Needs: network access from the agent to https://a11y-context-project.vercel.app. If a fetch fails, the skill surfaces the failure and never invents pattern guidance in its place.

Not HTTP if…​

  • You're offline or air-gapped, or need a pinned, deterministic snapshot — use Local, which bundles the corpus and reads it from disk.
  • You're already MCP-native and want retrieval composable with your other MCP sources — use the MCP server.
  • You have existing enterprise retrieval infrastructure to reuse — see Custom / Enterprise RAG.

Verifying​

After unzipping, run the prompts on Verify It's Working. The agent should select pattern IDs from the local catalog, then fetch …/web/react/components/{id} for each before generating UI code.

On This Page