The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Markdown Notes MCP listing page.
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.
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.
| Documents | PDF, Word .docx, Excel .xlsx/.xls/.xlsm, PowerPoint .pptx, OpenDocument .odt/.ods, EPUB, CSV, TSV, plain text |
| Images | OCR on the device, with no key and no network |
| Audio | Transcription — the one feature that cannot run locally |
| Tool | What it does |
|---|---|
convert_to_markdown | Converts one file and saves the .md beside the original automatically |
convert_folder_to_markdown | Converts several files into a chosen directory |
save_summary | Saves a summary the assistant wrote as <name>.summary.md beside the document |
list_supported_formats | Lists the extensions, and which optional cloud features are configured |
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.
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.
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
Then point your host at the suprasuta-markdown-mcp command, or run it without
installing:
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.
Both are dormant until a key is present. Everything else works without either.
| Variable | Enables | What leaves your machine |
|---|---|---|
GEMINI_API_KEY | Cloud OCR, which can also describe charts and diagrams | The image, to Google |
ASSEMBLYAI_API_KEY | Audio transcription | The 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.
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.
Three things genuinely differ between a browser and Node, and
src/lib/convert/runtime.ts asks the host rather than sniffing for process:
| Hook | Why |
|---|---|
tesseractPaths | A browser needs URLs and its own worker; Node needs filesystem paths and must not be given the browser worker |
docxSource | mammoth reads arrayBuffer in browsers and buffer under Node |
ocrImageInput | tesseract'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
documenttoo, and it broke OCR. tesseract.js sniffed for a DOM, found one, took the browser branch and died onwindow is not defined. Half a browser is worse than none — shim the minimum.
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.
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.
| Windows app | Microsoft Store · source |
| Web app | markdown-notes-psi.vercel.app · source |
| Android | Install the web app from Chrome · older native build, no longer maintained |