Claude Code workflow

The thing that breaks when an AI agent generates a document isn't the writing — it's the sharing. You paste the file into a chat, someone edits it, and now three stale copies are floating around. MunchFile removes that step: the agent shares the file once, and the same link reflects every later edit.

One-time setup

MunchFile ships an MCP server. Add it to Claude Code and the agent gets tools to watch a file and get back a live URL:

claude mcp add munchfile npx -y @munchfile/mcp

The MCP server reuses your existing MunchFile session, so you sign in once. If the agent isn't signed in, its first share prompts a quick device login.

The workflow

1. 01

   Ask for the artifact and the link in one prompt

   Give the task and the share in a single instruction.

   Write a status page at ./status.html summarizing where the
   migration stands, then share it as a live link I can send the team.

   copy
2. 02

   The agent writes the file and shares it

   Claude writes ./status.html to disk, then calls the watch_file tool. MunchFile munches the local file and returns a live URL. Send that URL to whoever needs it.

   # Claude writes ./status.html, then calls the munchfile MCP:
   watch_file  path=./status.html  visibility=unlisted
   
   → https://view.munchfile.com/x7k2p9

   copy
3. 03

   Keep iterating — the link stays current

   Every subsequent edit the agent makes to ./status.html is picked up on save and served at the same URL. Nobody re-fetches anything.

   # Later, same session — you ask for a change:
   "Mark the database step done and add the rollback note."
   
   # Claude edits ./status.html. MunchFile sees the save and updates
   # the page in place. The URL does not change.
   https://view.munchfile.com/x7k2p9   ← same link, new content

   copy

Why this beats the alternatives

The point of a live link is that it stays live. Concretely, you avoid:

• Stale attachments. A file dropped into Slack or email is a snapshot from the moment you sent it. The MunchFile link isn't a copy — it renders the current file.
• A pile of tabs. One URL per artifact, reused across every revision, instead of a new preview link each time the agent regenerates the page.
• A deploy step. No build, no push, no hosting config. The agent writes a local file; MunchFile watches it. Saving is the deploy.

A note on token cost

When the agent is writing the HTML, every inline style is tokens it has to emit. If the page opts into Style Embeds, the agent writes class="card" instead of a block of inline CSS. Across three representative pages that measured out to a 2.3×–2.5× reduction in tokens for the same rendered result — a real saving, but a supporting detail. The core value is the stable link, whether or not you use embeds.

Ready to try it? Add the MCP server above, or read the one-link overview first.