0

Building a Zero-Dependency Local stdio MCP Server for Plain-Text OKF Markdown Meeting Wikis

Most meeting tools trap spoken context inside proprietary cloud dashboards. Even when they offer a one-off Markdown export, the exported files are static snapshots—unconnected to prior decisions, people, or project specs, and invisible to the AI coding agents developers already use.

When building open tooling around OneMira Desktop and the OneMira C1 AI Voice Recorder, we wanted spoken context to stay 100% local ($0/month, zero forced cloud uploads) while remaining immediately queryable from Claude, Cursor, and ChatGPT through a clean Model Context Protocol (MCP) interface.

Here is the minimal architecture behind okf-wiki-mcp and the Open Knowledge Format (OKF) Local Wiki.

1. Plain-Text Schema First (OKF v0.2 Frontmatter)

Instead of hiding relationships in an opaque SQLite blob, every meeting transcript, topic page, and person profile is a plain UTF-8 .md file with deterministic YAML frontmatter and [[wikilink]] references:

---
kind: session
session_id: s_20261121_093000
created_at: 2026-11-21T09:30:00Z
title: Q4 Firmware & Local AI Pipeline Sync
speakers:
  - "[[Alex Chen]]"
  - "[[Maya Patel]]"
topics:
  - "[[Local Wiki Architecture]]"
---

## Decisions
- Keep local transcription and built-in offline AI completely subscription-free.
- Expose read/search/patch tools over local `stdio` MCP without network sockets.

Because every record is ordinary Markdown, you can edit it in VS Code, Obsidian, or Git—and retain original stereo WAV audio only if you opt in through Settings.

2. Why Zero-Dependency stdio Beats Local HTTP Daemons

For a local knowledge base, exposing an unauthenticated localhost HTTP port adds port-collision bugs and unnecessary attack surface. A pure stdio JSON-RPC 2.0 server spawned directly by the client solves both:

  • Zero background daemon management: the AI client starts and terminates the process cleanly.
  • Explicit directory scoping: pass a single --vault path so the server can only read or patch files inside your designated local wiki folder.
  • Fast cold start: zero runtime npm dependencies keep startup under 50 ms.

Example MCP client configuration (documented in the CLI & MCP Integration Guide):

{
  "mcpServers": {
    "okf-wiki": {
      "command": "npx",
      "args": ["-y", "okf-wiki-mcp", "--vault", "/path/to/local-wiki"]
    }
  }
}

3. Core Read, Search & Compounding Wiki Primitives

Rather than dumping raw 10,000-word transcripts into an LLM context window, the server exposes narrow, composable tools:

  1. wiki_search — scores frontmatter metadata (kind, topics, speakers) and body matches locally.
  2. wiki_read — returns a single topic or session file with resolved [[wikilink]] backlinks.
  3. wiki_backlinks — lists every session that referenced a given person or project slug.

This lets compatible agents pull only the decisions and commitments relevant to the current task—while keeping every note editable on disk and usable offline with built-in Local AI or your own API key provider.

Open-Source Packages & Documentation


All rights reserved

Viblo
Hãy đăng ký một tài khoản Viblo để nhận được nhiều bài viết thú vị hơn.
Đăng kí