Use Fizzdoc from AI coding agents: a local MCP server and Claude Code / Codex plugins #57

Why

AI coding agents are often asked to "merge these PDFs", "cut this MP3" or "strip the metadata before I send this". Today that means either uploading the file to a web service or writing a one-off script.

Fizzdoc's engines already do these jobs well, and they don't need a server. If agents can call them as local tools, the file never leaves the user's machine. That is the same promise the website makes.

Plan

This is tracked as smaller issues so several people can work on it in parallel:

• MCP server for the tools that already run in Node (PDF page tools and audio) — see the sub-issue
• Claude Code plugin that bundles the server with a skill
• Codex setup: docs and config for Codex CLI
• Node backends for the image, OCR, redaction and Office engines, so they can be added to the server too

Ground rules

• Local only. No network calls, no telemetry, no uploads. Inputs are file paths on the user's machine; outputs are written next to them, or to a folder the user chooses.
• Same engines. Reuse src/engine/* rather than re-implementing tools, so the website and the agent tools stay in step.
• Tested. Every tool gets a test that runs it on a fixture from tests/fixtures/ and checks the output, the same way the web tools are tested.

What already works without a browser

src/engine/pdf.ts (qpdf compiled to WebAssembly) and src/engine/audio.ts have no DOM dependencies. So a first version can offer merge, split, rotate, delete, protect, unlock and clean PDF, plus trim, split and merge MP3/M4A.

The image, OCR, redaction and Office engines use OffscreenCanvas / createImageBitmap and need a Node backend first.

Comment on a sub-issue to claim it. Questions about the design are welcome here.