fresh-fetch

fresh-fetch

A Claude Code plugin that stops Claude from answering web-content questions with stale training data.

Whenever a URL appears in your message — or you ask about content/docs/APIs/prices/versions that live on the web — fresh-fetch routes through the best available fetcher (specialized MCP → WebFetch → Chrome MCP → Playwright) and refuses to guess when no fetcher works.

Why

Out of the box, Claude often answers URL-bearing questions from training data. That's confidently wrong when:

• The library/API/docs changed since the training cutoff
• The page is behind auth (Notion, Linear, Figma, internal tools)
• The site is JS-heavy and didn't pre-render

fresh-fetch makes the right tool obvious and the failure mode loud: a clear refusal, not a hallucination.

Install

In Claude Code:

/plugin marketplace add brandonhall/fresh-fetch
/plugin install fresh-fetch@fresh-fetch

Then open a new session — fresh-fetch will trigger automatically on URL-bearing requests.

If /plugin isn't available, update Claude Code to the latest version.

Updating

/plugin marketplace update fresh-fetch

Uninstalling

/plugin uninstall fresh-fetch@fresh-fetch
/plugin marketplace remove fresh-fetch

What it does

Every URL-bearing request triggers a short procedure:

1. In-session cache — if Claude already fetched this URL in the last 15 minutes, reuse it
2. Specialized MCP — if a matching MCP is connected (GitHub, Linear, Supabase, Vercel, Render, Slack, etc.), use it
3. WebFetch — Claude Code's built-in fetcher
4. Chrome MCP — uses your existing Chrome session (handles auth, cookies, JS)
5. Playwright — headless fallback, no auth state
6. Refuse + offer install — if nothing works, Claude explains the situation and offers to install Chrome MCP or Playwright

Every fetch prints one short status line so you can see what just happened:

[fresh-fetch] react.dev/learn — source: WebFetch (fresh, 1.2s)
[fresh-fetch] github.com/foo/bar — source: github MCP
[fresh-fetch] notion.so/abc — WebFetch failed (auth wall) → Chrome MCP

What it doesn't do

• Doesn't take screenshots — text/markdown extraction only
• Doesn't cache across sessions — uses WebFetch's existing 15-minute in-session cache
• Doesn't replace action tools — if you say "open this PR" or "push to this URL", Claude uses the right action tool instead of fetching

Recommended companion tools

For best results, install one or both:

• Claude for Chrome (recommended) — official Chrome extension, uses your existing browser session
• Playwright — npm install -g playwright && npx playwright install chromium

If neither is installed, fresh-fetch will offer to install Playwright for you the first time a fetcher chain runs out.

How it triggers

The skill description is intentionally aggressive: it fires on any URL in your message, and on questions about web-hosted content even without a URL. If it triggers when you'd rather it didn't (e.g., the URL is incidental and you wanted Claude to use a different tool), just say so — the skill will defer to action tools when you're asking Claude to do something with the URL rather than read it.

License

MIT — see LICENSE.

Design notes

See DESIGN.md for the design rationale and PLAN.md for the implementation plan.