An AI knowledge base that runs in your browser and your local folder. Open a URL, pick a folder, the agent is ready to help. No app to install, no account, no cloud. Your folder, your API key, your files.
Framer AI AgentsDesign and publish professional sites with AI
Promoted
Maker
π
Hey Product Hunt π
I'm Yunbiao, and localmd is the thing I built because every answer to "let an AI work on my own files" asked me to install something first β an app, a plugin, a local server, a terminal, or even a virtual machine on my Mac.
localmd is just a web page. Open it, grant it **one folder**, and the agent is already there. There is no backend to upload anything to, and no account to create β the whole app is a static page, and your browser's own permission model is the entire security arrangement.
**What it actually does**
π **Reads what you already have.** Drop in a 300-page PDF or an EPUB. Ask a question. The answer comes back with citations you can click β landing on the exact paragraph, highlighted.
βοΈ **Writes back into your folder.** Those citations get written into your own Markdown. They still work tomorrow, in your folder, in any editor, with or without us. This is the part NotebookLM structurally can't do β it reads your sources, but the result stays in its notebook.
π **It adapts to your folders, not the other way round.** There is no import, because there is no format to import into. Point it at your Zotero storage directory, your Calibre library, the pile of PDFs in Downloads, or a vault you already keep β it reads how you already organize things and files new material the same way. Every other tool in this category asks you to move into its structure first.
β **Never edits silently.** Every write is a line-level diff you approve. In ask-first mode, nothing touches disk until you say yes. Git is built in, `isomorphic-git` running in the tab, reading and writing a real `.git` your terminal recognizes.
π **Your key, your model, no caps.** Anthropic, OpenAI, DeepSeek, Gemini and more, each hitting its own endpoint directly from the page. You pay them at their prices. We don't meter anything, because we can't see anything.
The rest is what you'd expect and I won't spend your time on it: wikilinks and backlinks, a graph view, ctrl+p search, skills, MCP, memory, and a deterministic health check for broken links and orphans β a linter, no token cost.
**One thing that isn't obvious:** there's a companion Chrome extension, localmd Connect, that hands the agent *your own logged-in browser* β open a tab, read a page, click, plus ready-made adapters for 30+ sites (arXiv, Hacker News, Wikipedia, Reddit, YouTube, β¦), around 300 actions in all. So "find that paper on arXiv, save it to my papers folder, index it, and tell me what's new since the 2023 version" is one sentence. The extension is free to install; connecting it to the app is the paid part β and it's the clearest example of where I drew that line.
**Three honest things before you click**
- **Chrome or Edge only.** The API that lets a web page read and write a folder you choose (File System Access) exists only in Chromium. That limit is also why this slot was empty β nobody who needs Safari, mobile and enterprise procurement is going to build here.
- **"Nothing leaves your machine" is not quite true, and I won't say it.** Your files stay in your folder. But the text you send the agent goes to the model provider *you* configured, at your key. That's the one thing that leaves, and it goes where you pointed it.
- **It's early.** There are rough edges. Scanned PDFs with no text layer produce nothing (no OCR yet). I'd rather you find these and tell me.
**On money:** the core is free forever, with no usage caps β web search included β because it's your key and your money. Paid is only what reaches past your folder and your model: the browser extension, MCP servers, GitHub sync. $39 once, verified offline, no account. It isn't on sale yet β there are 100 free early-access slots instead, and I'd rather give you one than take your money before it's earned.
I'll be here all day. I'm most interested in where the first 60 seconds goes wrong for you β that's the thing I can't see from my own machine.
π https://localmd.app