autocli_anthropic

Use when the user wants data fetched from a website, a site turned into a CLI, existing browser login state reused through Chrome, or a supported local CLI/desktop app accessed through AutoCLI. Prefer

SKILL.md

GitHub

AutoCLI

Use this when a site or supported app already has an AutoCLI adapter, or when the user wants “turn this website into a command” behavior without a whole browser-driving session.

What it does

AutoCLI is a fast Rust CLI that turns websites and some local tools/apps into command-line interfaces. It can:

  • fetch structured data from supported sites
  • reuse Chrome login state through its extension for browser-backed commands
  • search for existing community adapters
  • generate adapters for new sites via autocli.ai
  • wrap a few local CLIs like gh, docker, kubectl, and Google Workspace helpers

Think of it as: “try a purpose-built website/app CLI first, then fall back to browser automation if needed.”

URL-to-markdown extraction with curl.md

For one-off webpage-to-markdown extraction, use Hermes curlmd_fetch first when the goal is clean markdown context and no site-specific tool exists. It wraps curl.md with bounded timeouts, retry/backoff, authenticated token use from CURLMD_API_KEY/1Password when available, and a plain curl fallback for stalls, rate limits, weak/blocked pages, or CLI failures.

Current docs/observed behavior:

  • Basic use inside Hermes: call curlmd_fetch(url="https://example.com"); terminal fallback is npx -y curl.md@0.1.1 <url>.
  • Authenticated Mac Studio token is stored in 1Password item CLI / curl.md API Token - Hermes Mac Studio; expose it as CURLMD_API_KEY or pass --token for non-interactive terminal calls.
  • Query-string targets need encoding as a path segment, e.g. PubMed search: https://curl.md/pubmed.ncbi.nlm.nih.gov/%3Fterm=PLGA+drug+delivery; otherwise the target query can be treated as curl.md API params and fetch the home page.
  • Anonymous limits: standard fetches 100/hour; objective= narrowing only 3/hour. Authenticated limits: 1,000/hour and 10/hour. Paid usage skips these limits.
  • Good fits from benchmark: docs, public forum pages like LPSG threads, normal GitHub repos, Apple docs, and PubMed encoded-query pages.
  • Poor fits from benchmark: X profiles return empty shells (use xurl first), Reddit can return verification shells, YouTube can 429, and GitHub repo names containing dots such as wevm/curl.md can fail with upstream 404.
  • Treat x-tokens-count, x-tokens-saved, x-cache, and rate-limit headers as useful benchmark telemetry, but judge extraction quality by content hits and blocked/error markers, not HTTP 200 alone.

See references/curlmd-heavy-site-benchmark-2026-05-13.md for the benchmark matrix, rate-limit receipts, query-string gotcha, and routing notes.

Environment notes

  • Expect the autocli binary to be available on PATH.
  • Keep auth tokens in the machine's normal secure secret store or local AutoCLI auth flow.
  • If a team already has a standard token-retrieval workflow, use that instead of inventing a new one.

Do not paste tokens into files, skills, or committed config.

First checks

  1. Confirm the binary exists:
    • autocli --version
  2. If AI generation/search features need auth, run:
    • autocli auth
  3. If browser-backed commands matter, make sure Chrome is open and the AutoCLI extension is installed.

Default workflow

  1. Check whether the target already has a command:
    • autocli --help
    • or autocli search <url>
  2. If a matching adapter exists, use it directly.
  3. If not, and the user wants extraction from a new site, try:
    • autocli generate <url> --goal '<plain-English goal>' --ai
  4. If the task needs full arbitrary page interaction, rich login flows, or UI testing, switch to browser automation instead of forcing AutoCLI.

Good fits

  • “Get structured data from this website.”
  • “Can you pull this page into JSON/CSV/markdown?”
  • “Is there already an adapter for this site?”
  • “Turn this site into a reusable CLI command.”
  • “Use the site I’m already logged into in Chrome.”

Poor fits

  • Multi-step UI testing
  • Complex forms or flows with lots of clicks
  • Sites with no useful adapter path where browser automation is clearly simpler
  • Tasks where raw browser screenshots/DOM interaction matter more than structured output

Quick commands

autocli --version
autocli search https://example.com
autocli generate https://example.com --goal 'list the latest posts' --ai
autocli hackernews top --format json
autocli gh repo view cli/cli --format json

Notes

  • Public commands can work without browser auth.
  • AI generation/auth features use autocli.ai and store local config in ~/.autocli/config.json.
  • If auth is needed again, prefer the machine's existing secure secret workflow instead of asking the user to retype a token that is already managed locally.