Markdown Notes Mcp

io.github.Narashiman-KPublished by GitHub account Narashiman-Kv0.1.4Updated Oct 24, 2026

Convert PDF, Word, Excel, PowerPoint, EPUB and images to Markdown locally, with no upload.

VerifiedSTDIODesktop onlyFiles & StorageMedia & Design

Overview

AI-generated overview

Converts local documents and images to model-friendly Markdown, with optional cloud OCR and audio transcription.

What you can do
Runs on your machine and converts PDF, Word, Excel, PowerPoint, OpenDocument, EPUB, CSV, TSV, plain text and images into tidied Markdown, promoting tables and fencing code. Tools include convert_to_markdown, which saves the .md beside the original, convert_folder_to_markdown for batches, save_summary for assistant-written summaries, and list_supported_formats. Long conversions are saved to disk with only a short preview returned to limit token use.
When to use it
Useful when an assistant needs to read formats it cannot handle natively, such as EPUB, PowerPoint, OpenDocument or scanned images, and you would rather not upload the files. Also suits batch conversion of a folder of documents into Markdown.
Requirements
A local Node.js runtime; install from npm as suprasuta-markdown-mcp or via the .mcpb package for Claude Desktop. Optional secrets ASSEMBLYAI_API_KEY for audio transcription and GEMINI_API_KEY for cloud image recognition. MARKDOWN_MCP_ROOTS optionally limits readable folders; without it any path is allowed. Settings are read only at start-up, so restart the host after changes.
Before you install
Without MARKDOWN_MCP_ROOTS the server may read any path the assistant requests. Conversion writes .md files beside the originals or into a chosen directory. If GEMINI_API_KEY is set, images are sent to Google; if ASSEMBLYAI_API_KEY is set, audio is sent to AssemblyAI. With no keys, no network connections are made. Keys are read from the environment and not written to disk.

Installation

In SourceWeft

  1. Open Markdown Notes Mcp in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.

Other MCP clients

Follow the launch instructions in the repository.

README

Suprasūtā Markdown Notes — MCP server

Let Claude, VS Code and Antigravity read any document on your machine — without uploading it.

Created by Narashiman Krishnamurthy


Licence: free for personal, non-commercial use. See LICENSE.md.

What it does

AI assistants can read plain text, and some can read PDFs. None of them read EPUB, OpenDocument, PowerPoint or a scanned image, and all of them do it by uploading your file to a server.

This converts documents on your own machine and hands back Markdown that has been tidied for a model to read — tables promoted to real Markdown tables, deep indentation flattened, code properly fenced.

DocumentsPDF, Word .docx, Excel .xlsx/.xls/.xlsm, PowerPoint .pptx, OpenDocument .odt/.ods, EPUB, CSV, TSV, plain text
ImagesOCR on the device, with no key and no network
AudioTranscription — the one feature that cannot run locally

Tools

ToolWhat it does
convert_to_markdownConverts one file and saves the .md beside the original automatically
convert_folder_to_markdownConverts several files into a chosen directory
save_summarySaves a summary the assistant wrote as <name>.summary.md beside the document
list_supported_formatsLists the extensions, and which optional cloud features are configured

About tokens

Converting costs you nothing — it all happens locally. But returning a long document into the conversation does cost tokens, at the same rate as pasting it in by hand.

So a long conversion is saved to disk and only a short preview comes back. A 50-page PDF costs roughly 400 tokens instead of 30,000. Short documents are returned in full, since they are cheap either way. Ask for return_content: "full" to override.

Installing

Claude Desktop — the easy way

Download the .mcpb file from Releases and double-click it. Claude Desktop installs it and asks which folders it may read, and optionally for API keys. No Node, no terminal, no JSON.

Claude Desktop — from source

bash
npm installnpm run buildnode scripts/install-claude-desktop.mjs --roots "D:\your\documents"

Then restart Claude Desktop completely. The script patches the existing config rather than replacing it, and writes a timestamped backup first.

--roots is optional but recommended: it limits which folders the server will read. Without it, any path the assistant asks for is allowed.

To remove it again: node scripts/install-claude-desktop.mjs --remove

From npm — any MCP host

bash
npm install -g suprasuta-markdown-mcp

Then point your host at the suprasuta-markdown-mcp command, or run it without installing:

json
{  "servers": {    "suprasuta-markdown": {      "command": "npx",      "args": ["-y", "suprasuta-markdown-mcp"],      "env": { "MARKDOWN_MCP_ROOTS": "D:\\your\\documents" }    }  }}

VS Code and Antigravity

Both read a standard MCP configuration — the block above works as-is. Antigravity looks in ~/.gemini/config/mcp_config.json, or .agents/mcp_config.json inside a workspace.

To run a checkout rather than the published package, replace command and args with node and the absolute path to dist/server.js.

Optional cloud features

Both are dormant until a key is present. Everything else works without either.

VariableEnablesWhat leaves your machine
GEMINI_API_KEYCloud OCR, which can also describe charts and diagramsThe image, to Google
ASSEMBLYAI_API_KEYAudio transcriptionThe audio file, to AssemblyAI

Keys are read from the environment, never written to disk and never logged.

Settings are read once, at start-up. An MCP server is a long-running process, so a key or folder saved while it is running does not reach it until the host restarts. If a key you have just entered reports as not configured, restart the host before looking for anything more complicated.

Development

bash
npm install          # .npmrc sets include=dev — do not delete itnpm run build        # stages the OCR engine, then bundles with esbuildnpm test             # 19 tests against the built bundlenpm run smoke        # converts every file in samples/ and reportsnpm run typecheck

Why the build is esbuild rather than tsc

src/lib/convert/ is kept byte-identical to the copies in the Windows, web and Android apps, so that drift can be detected rather than discovered. Those files use extensionless imports and one Vite-specific ?url import, neither of which Node's own resolver accepts. esbuild resolves both, so the shared files never need Node-shaped edits.

The runtime hook

Three things genuinely differ between a browser and Node, and src/lib/convert/runtime.ts asks the host rather than sniffing for process:

HookWhy
tesseractPathsA browser needs URLs and its own worker; Node needs filesystem paths and must not be given the browser worker
docxSourcemammoth reads arrayBuffer in browsers and buffer under Node
ocrImageInputtesseract's Node build cannot decode a Blob; it wants a Buffer

Everything else is shared. src/runtime/node.ts supplies the Node answers and shims exactly one browser global, DOMParser.

A caution learned the hard way: that file used to define document too, and it broke OCR. tesseract.js sniffed for a DOM, found one, took the browser branch and died on window is not defined. Half a browser is worse than none — shim the minimum.

Keeping the copies honest

bash
npm run sync:check              # list differences against all three appsnpm run sync:pull web           # bring the web app's version innode scripts/sync-converters.mjs --push web

This package is the canonical copy: it is the only one with the runtime hook that lets identical files run in both environments, so fixes start here and flow outward. sync:check reports and exits zero — drift is information, not a build failure.

Privacy Policy

This software collects nothing. No account, no analytics, no telemetry, and no server of its own — document conversion happens entirely on your machine.

Two optional features send data to a third party, and only when you supply your own API key: cloud image recognition sends the image to Google, and audio transcription sends the audio to AssemblyAI. With no keys configured, the software makes no network connections at all. API keys are read from the environment, never written to disk and never logged.

Full policy, covering collection, storage, retention, third-party sharing and contact details: PRIVACY.md.

Related

Windows appMicrosoft Store · source
Web appmarkdown-notes-psi.vercel.app · source
AndroidInstall the web app from Chrome · older native build, no longer maintained

Source: README.md at commit bca560f

Tools

0
SourceWeft has not started this package in its sandbox yet, so its tools are not known.

Version history

1
  1. v0.1.4LatestSep 16, 2026

Score

40% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 days25 pts

    Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.

  • Auth declared10 pts

    Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.

  • Tools listed15 pts

    Answer tools/list without a credential, or ship a package that starts in a sandbox without one.

  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts

    Declare a license in the repository or the registry entry.