mcps
F G O

Recipe

Give Claude your files, repo, and notes

Local filesystem, Git history, and your Obsidian vault — so Claude works from what you actually have, not guesses.

for Founders and analysts who want answers without writing SQL

Hand-checked recipe · curated by Ilya Gindin

10 minutes

to run

beginner

difficulty

3

connectors

What you'll need

Filesystem Scoped read/write access to a folder you choose on your machine.
Optional
Git Read commit history, diffs, and blame for a local repository.
Optional
Obsidian Search and read your Obsidian vault of markdown notes.
Optional

This recipe runs in 3 steps

  1. 1

    Scope the access

    Say to your AI

    Install the filesystem MCP and point it at this one project folder only. Confirm what you can and can't reach.

    You'll get: Claude confirms the single allowed directory and refuses paths outside it.

  2. 2

    Get oriented

    Say to your AI

    Summarize what this project is from its files: structure, entry points, and what each top-level folder does.

    You'll get: A short orientation map of the project.

  3. 3

    Do real work on it

    Say to your AI

    Find every place we read an env var and list the ones that are used but never documented in the README.

    You'll get: A list of undocumented env vars with file:line references.

You're done when

Claude answers from your actual files and history — fewer hallucinations, no more copy-pasting code into chat.

Why this workflow exists

The fastest way to make an AI useful on your project is to let it read the project. A filesystem server (scoped to one folder), a Git server for history, and a notes server for your own knowledge base give Claude the same context you have — without you pasting files into chat one at a time. Keep the filesystem scope tight: one project folder, not your home directory.

People ask their AI

“let Claude read my files”“give Claude access to my repo / notes”“what MCP for filesystem or git”