claude-skills is a plugin, not a one-off script — skills live inside it. The first one I built, webagent, gives whichever agent it's dropped into full browser automation: find content, scrape data, research a topic, download media, or navigate and interact with a site, all from a plain-language request.

One Plugin, Any Platform

Skills follow the open SKILL.md standard, so the plugin isn't locked to one tool. According to its own compatibility table, it works across every major coding agent without modification:

  • Claude Code — skills + hooks supported
  • OpenCode — skills + hooks supported (Claude Code compatible config)
  • Codex CLI (OpenAI) — skills + hooks supported
  • Copilot CLI — skills + hooks supported
  • Antigravity CLI (Google) — skills supported, hooks not yet ported

Drop webagent into any of these and that agent can suddenly browse, scrape, and download — it's not a Claude Code-only trick.

What webagent Can Actually Do

Every request gets classified into one of five task types before anything happens:

  • find_content — videos, articles, examples, demos
  • scrape_data — structured data extraction from a page
  • research — gather and compare information across sources
  • download — video/audio/file download (saved to ~/Downloads/webagent)
  • navigate — a deterministic URL + action, e.g. "go to github.com/settings and find 2FA"

In practice that looks like: "go to youtube and find me a video on Claude Code hooks", or "download this YouTube video", or "go to docs.aws.amazon.com/bedrock and find all Moonshot AI model documents" — webagent figures out which mode it's in and takes it from there.

It Plans Before It Acts — Unless You Already Have One

Most agents fail because they act before they understand what you actually meant. webagent runs three phases: clarify, plan, execute.

  • Clarify — one question at a time (topic, then goal, then format) until the request is unambiguous
  • Plan — a step list with a token estimate per step, shown to you before anything runs
  • Execute — runs one step at a time and reports tokens spent after each one
Step 1: Navigate to youtube.com           [~180 tokens]
Step 2: Search for query                  [~150 tokens]
Step 3: Scan results                      [~300 tokens]
Estimated total: ~630 tokens
Proceed? (yes / adjust)

If the task is already deterministic — like a direct navigate request — it skips the clarification questions entirely and goes straight to the plan. You don't get interrogated for something you already spelled out.

Token Tracking Is the Specialty

This is the part I care most about. A PostToolUse hook fires after every browser action, reads the actual tool output, and counts real tokens — not an upfront guess — accumulating a running session total.

2026-06-15T10:23:44Z | browser_navigate | 142 tokens | session total: 4821
2026-06-15T10:45:00Z | HARD STOP | session total 50142 exceeded 50000 cap

At 50,000 tokens the hook injects a hard stop and the agent halts all further browser tool calls — so a research task can't quietly burn through your context window without you noticing.

Installing It

# Mac / Linux
git clone https://github.com/Aakashjammula/skills
cd skills
bash scripts/install.sh
# Windows (run as Administrator for symlinks)
git clone https://github.com/Aakashjammula/skills
cd skills
.\scripts\install.ps1

The install script symlinks every skill into each platform's discovery path automatically — no manual config wiring per tool.

More Skills Coming

webagent is one skill living inside the claude-skills plugin — I'm adding more over time. Adding a new one is intentionally low-friction: drop a SKILL.md into skills/<name>/, no extra config required.

claude-skills GitHub Repo