MarkScribe
A convention-aware MCP server for managing markdown files with YAML schema validation. It enables AI agents to read, write, search, and validate markdown notes while enforcing user-defined conventions across directories like Obsidian vaults and documentation repositories.
README
MarkScribe
A convention-aware markdown MCP server for AI assistants. Point it at a directory of markdown files and it gives the AI read, write, search, wikilink, and validation tools, enforcing your conventions through user-defined YAML schemas rather than hard-coded vault assumptions.
Works with Obsidian vaults, Foam workspaces, Logseq graphs, digital gardens, documentation repos, or any plain markdown directory. Nothing about the format is assumed. If your directory has its own rules — required frontmatter, hub notes, filename patterns, link constraints — you express them as schemas and MarkScribe enforces them.
The distinction matters: conventions are enforced, not assumed. A schema-less directory still gets the full read/write/search/link toolkit; a schema-driven directory additionally gets structural validation, convention-aware note creation, and lint feedback on every file.
What It Does
Read, write, search. 23 tools for AI assistants to operate on markdown: atomic read/write/move/delete, batch reads, frontmatter-aware patching, and full-text BM25 search across body and frontmatter.
Wikilink graph. Backlinks, broken link detection, orphan finding, and plain-text mention discovery. The graph rebuilds on every call, so there is no stale index or cache to invalidate.
Schema validation. User-defined YAML note and folder schemas. Note schemas validate frontmatter fields and content rules; folder schemas classify directories, assign note schemas by role, and enforce structural constraints. _conventions.md files scope schemas to subtrees so the same directory can host multiple conventions.
Path security. .obsidian/, .git/, node_modules/, .DS_Store, and Thumbs.db are always blocked. User config can extend the blocklist, never shrink it. Atomic writes everywhere, so a crashed process never leaves a torn file.
Lite mode. --lite trims the tool surface from 23 to 11, keeping schema validation, the wikilink graph, and directory meta. Note CRUD, frontmatter editing, and search drop out, since harnesses like Claude Code already ship native file tools for those.
Quick Start
Prerequisites
- Node.js v18+
- A directory of markdown files
Install
npm install -g markscribe
Or run directly via npx, no install step. The MCP config below shows both.
Configure Your MCP Client
Add the following to your MCP client config. Works with Claude Code, Claude Desktop, Cursor, or any MCP-compatible client.
Zero-install via npx (recommended):
{
"mcpServers": {
"markscribe": {
"command": "npx",
"args": ["-y", "markscribe", "--root", "/path/to/your/notes"]
}
}
}
Or install globally:
{
"mcpServers": {
"markscribe": {
"command": "markscribe",
"args": ["--root", "/path/to/your/notes"]
}
}
}
--root is the directory MarkScribe will serve. To load your own schemas, add "--schemas-dir", "/path/to/schemas". Otherwise ~/.markscribe/schemas/ is used.
Verify
Ask your AI assistant to call get_stats. If it returns a note count and recent files, you're connected.
CLI flags
| Flag | Default | Description |
|---|---|---|
--root <path> |
Current working directory | Root directory to serve |
--schemas-dir <path> |
~/.markscribe/schemas/ |
Directory to load schema YAML files from |
--log-level <level> |
info |
Log level (debug, info, warn, error, fatal) |
--lite |
off | Trim the tool surface to lint, validation, and link-graph only |
Lite mode
If your AI client already has native file read/write/search (like the Claude Code harness does), the note CRUD, frontmatter, and discovery tools are duplicative. --lite exposes only what MarkScribe uniquely provides — convention enforcement and wikilink analysis — and leaves file manipulation to the harness.
markscribe --lite --root /path/to/your/notes
Kept (11): lint_note, validate_folder, validate_area, validate_all, list_schemas, get_backlinks, find_broken_links, find_orphans, find_unlinked_mentions, get_stats, switch_directory.
Cut (12): read_note, write_note, patch_note, delete_note, move_note, read_multiple_notes, create_note, get_frontmatter, update_frontmatter, manage_tags, search_notes, list_directory.
The flag is a startup decision — restart the server to toggle it. Default behavior is unchanged for clients that don't pass --lite.
Per-directory config
Place a .markscribe/config.yaml in your root directory:
paths:
blocked:
- private/
- drafts/
allowed_extensions:
- .md
- .markdown
- .txt
search:
max_results: 50
excerpt_chars: 40
The built-in security blocklist (.obsidian/, .git/, node_modules/, .DS_Store, Thumbs.db) is always enforced on top of user config.
Schemas (the short version)
Schemas are YAML files defining conventions for notes and folders. Note schemas validate frontmatter and content; folder schemas classify directories and assign note schemas by role.
Note schema. Validates frontmatter fields and content rules:
name: blog-post
description: Blog post with required metadata
type: note
frontmatter:
fields:
title: { type: string, required: true }
tags: { type: list, required: true }
content:
rules:
- name: has-outgoing-link
check: hasPattern
pattern: "\\[\\[.+?\\]\\]"
Folder schema. Enforces structural rules on directories:
name: project-folder
description: Project folder with hub note
type: folder
noteSchemas:
default: blog-post
hub: project-hub
classification:
supplemental: [assets, templates]
skip: [archive]
hub:
detection:
- pattern: "_{{folderName}}"
required: true
Notes opt into a schema via note_schema: <name> in frontmatter, or inherit one from a _conventions.md file higher in the tree. The convention cascade resolves schema on a per-note basis.
Full schema reference, all field types, all check types, and the cascade resolution order: docs/schemas.md.
Tools
| Tool | Description |
|---|---|
list_directory |
List files and subdirectories |
get_stats |
Note count, total size, recent files |
switch_directory |
Change the active root directory |
read_note |
Read a note with parsed frontmatter |
write_note |
Create or update a note |
patch_note |
String replacement within a note |
delete_note |
Delete a note (with confirmation) |
move_note |
Move/rename with optional link updates |
read_multiple_notes |
Batch read up to 10 notes |
create_note |
Convention-aware note creation |
get_frontmatter |
Read YAML frontmatter only |
update_frontmatter |
Merge or replace frontmatter fields |
manage_tags |
Add, remove, or list tags |
search_notes |
Full-text BM25 search |
lint_note |
Validate a note against its schema |
validate_folder |
Classify and validate a folder |
validate_area |
Recursive subtree validation |
validate_all |
Full directory tree validation |
list_schemas |
List all loaded schemas |
get_backlinks |
Find notes linking to a note |
find_broken_links |
Find wikilinks to non-existent notes |
find_orphans |
Find notes with no incoming links |
find_unlinked_mentions |
Find plain-text mentions that should be wikilinks |
Compatible viewers
MarkScribe works with any tool that reads markdown files:
- Obsidian: PKM app with graph view and community plugins
- Foam: VS Code extension for linked notes
- Logseq: outliner with bidirectional links
- Any text editor or static site generator
Architecture
MarkScribe is stateless at runtime. There are no persistent indexes, caches, or file watchers; search and the link graph rebuild on every call, so results are always correct and never stale. Services (file, frontmatter, search, schema engine, link graph) are constructed via buildServices() and injected through a mutable ServiceContainer, which lets switch_directory rebuild the full service stack at runtime without re-registering tools. All file writes go through atomicWrite (write-to-temp-then-rename) so a crashed process never leaves a torn file. Convention knowledge is schema-driven: the server hard-codes no directory assumptions, only the inviolable path-security defaults.
Development
# Build
npm run build
# Test (vitest)
npm test
npm run test:watch
npm run test:coverage
# Lint and format
npm run lint
npm run lint:fix
npm run format
npm run format:check
# Type check
npx tsc --noEmit
Stdio transport: stdout is reserved for JSON-RPC, all human/debug output goes to stderr. Run tests after changes to services (src/services/) or the schema engine.
Acknowledgements
Built with Claude Code.
License
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。