Got tired of re-grepping details in my project, tools, and Claude's OWN capabilities to Claude Code every session. Built MCR (Model Context Retrieval). Even searching took too long and shoves in unnecessary context by trifling through files
— it auto-loads relevant Obsidian notes into context based on what you're
working on.
The vault is just markdown files with YAML frontmatter (tags, keywords,
priority). A Python script indexes them. Two Claude Code hooks do the rest:
1. You send a message → hook matches your prompt against the index → injects
relevant notes before Claude sees it
2. Claude reaches for a tool (Read, Grep, WebSearch) → hook treats that as a
knowledge gap → injects relevant notes alongside the tool results
Layer 2 is the interesting part. Tool calls are need signals — Claude is
saying "I don't have this info." The hook intercepts that and fills the gap
automatically. Claude never knows it was helped.
No vector DB, no MCP server. 4 Python scripts, pure stdlib,
~20ms per hook. Obsidian organizes the vault, hooks deliver it.
So I added files on current claude code capabilities, project files (gave higher abstracted files higher priority as to not overflow context), and MCR itself so that it adds to the vault over time. I'm gonna add details about myself next! Maybe this can be a method for cloning yourself as an ai assistant...?
Made this in like a couple hours, so let me know what you think!
So I added files on current claude code capabilities, project files (gave higher abstracted files higher priority as to not overflow context), and MCR itself so that it adds to the vault over time. I'm gonna add details about myself next! Maybe this can be a method for cloning yourself as an ai assistant...? Made this in like a couple hours, so let me know what you think!