MCP 服务器

Mimir's Vault

An Obsidian-compatible memory graph and context management MCP server that lets AI agents navigate project architecture and relationships via wikilinks before scanning raw code, reducing token waste.

README

<p align="center"> <b>ᚱ • ᚢ • ᛖ • ᛋ • ᚨ • ᛗ • ᛟ • ᛚ</b><br> <i>"Just as Odin drank from Mímir's Well to gain infinite cosmic wisdom,<br> AI agents require targeted knowledge—not a chaotic ocean of raw files."</i> </p>

Key Pillars • System Architecture • Installation • MCP Client Setup • Tool Reference • Agent Workflow Guide

</div>

The Problem: Wandering AI Coding Agents

When AI coding agents (Claude Code, Cursor, Antigravity) work on complex software repositories, they frequently suffer from Context Inflation & Wandering:

They blindly grep directories, re-scan full codebases, and load 30+ unrelated files into memory.
They waste thousands of LLM tokens, increase execution latency, and lose track of core architectural rules.

The Solution: Mímir's Vault Memory Graph

Mímir's Vault (Mimirs_Vault) is a high-performance Model Context Protocol (MCP) server built in Python (FastMCP).

It maintains a human-readable, interlinked Obsidian-Wikilink memory graph inside ./.context/ for any software repository.

AI agents follow the Memory-First Scanning Rule:

Inspect .context/overview.md via get_root_node.
Follow Obsidian wikilinks ([[modules/auth]]).
Inspect only the targeted source code files mapped in that module note.
Edit zero unnecessary files.

Key Pillars

Rune	Pillar	Rationale
ᛗ	Memory-First Protocol	Mandates AI agents to consult `.context/overview.md` before reading project source code.
ᛟ	Obsidian-Wikilink Graph	Interlinked notes (`[[modules/auth]]`) viewable natively in Obsidian, VS Code, and Foam graph visualizers.
⚡	Zero Token Waste	Reduces context consumption by up to 80%, eliminating agent codebase wandering.
⚔️	Surgical Node Updates	Tools to append, update, and log Architectural Decision Records (ADRs) without clobbering existing notes.
ᛚ	Human & Agent Shared Memory	Clean Markdown files (`.context/`) readable by both human developers and LLMs alike.

System Architecture

<your-project-root>/
└── .context/                        <-- Memory Graph Root (Ignored in .gitignore)
    ├── overview.md                  <-- ROOT NODE (Scope, File Structure Map, [[modules/*]] index)
    ├── decisions.md                 <-- Architectural Decision Records (ADRs)
    └── modules/                     <-- LINKED MODULE NODES
        ├── auth.md                  <-- Module Note (Scope, Mapped Source Files)
        ├── database.md              <-- Module Note
        └── ui/                      <-- Submodule Directory
            └── components.md        <-- Submodule Note ([[modules/ui/components]])

Installation

Prerequisites

Python 3.10+

1. Clone the Repository & Initialize Environment

# Clone repository
git clone https://github.com/GopikChenth/Mimir-s_Vault.git
cd "Mimir-s_Vault"

# Create virtual environment & install dependencies
python3 -m venv .venv
.venv/bin/pip install "fastmcp>=3.4.0" "mcp>=1.2.0" "pydantic>=2.0.0"

2. Verify Installation & Test Graph Engine

.venv/bin/python test_graph.py

Expected Result: ALL TESTS PASSED SUCCESSFULLY!

MCP Client Setup

Add Mimirs_Vault to your preferred AI Client configuration:

Google Antigravity IDE (`agy.json` / Settings) or Claude Desktop (`claude_desktop_config.json`)

{
  "mcpServers": {
    "Mimirs_Vault": {
      "command": "<path-to-repo>/.venv/bin/python",
      "args": [
        "<path-to-repo>/main.py"
      ]
    }
  }
}

Note: Replace <path-to-repo> with the absolute path where you cloned Mimir-s_Vault (e.g. /home/user/projects/Mimir-s_Vault or C:/Projects/Mimir-s_Vault).

Cursor IDE

Open Cursor Settings → Features → MCP Servers.
Click + Add New MCP Server.
Name: Mimirs_Vault
Type: command
Command: <path-to-repo>/.venv/bin/python "<path-to-repo>/main.py"

MCP Tool Reference

Mimirs_Vault exposes 7 core tools for AI agent navigation:

Tool	Signature	Purpose
`get_root_node`	`(workspace_path: str = ".")`	Fetches `.context/overview.md` (Project Scope, Master File Structure Map, [[modules/*]] index).
`navigate_node`	`(node_path: str, workspace_path: str = ".")`	Traverses a wikilink node (e.g. `modules/auth`, `decisions`). Returns content & mapped source files.
`search_nodes`	`(query: str, workspace_path: str = ".")`	Performs keyword and topic search across all `.context/` markdown notes.
`init_context_graph`	`(project_name, description, scope, workspace_path)`	Bootstraps `.context/` folder with starter `overview.md` and `decisions.md`.
`create_module_node`	`(module_name, scope, source_files, parent_node)`	Creates a module note and registers `[[modules/name]]` in the parent node.
`update_node`	`(node_path, content, section_heading, mode)`	Surgically appends or replaces content under a specific section heading.
`log_decision`	`(title, context, decision, consequences)`	Appends an Architectural Decision Record (ADR) in `.context/decisions.md`.

System Prompts & Resources

MCP Prompts

memory_first_workflow: Instructs AI agents: "Consult get_root_node first. Read ONLY the target source files mapped in module wikilinks."
bootstrap_project_context: Guides AI to auto-scan an un-documented repository and build .context/ notes.