context-hub-mcp
A local-first MCP server that turns a .context/ folder of markdown files into a searchable knowledge layer for AI coding agents.
README
context-hub-mcp
A local-first MCP server that turns a .context/ folder of markdown files into a searchable knowledge layer for AI coding agents.
- write domain knowledge in markdown, keep it in git
- index it locally for fast retrieval
- expose it as MCP tools for Claude Code, Copilot, and other MCP clients
The markdown is the source of truth. SQLite is a local runtime index only — don't commit it.
Quickstart
# 1. Scaffold the .context/ folder
npx context-hub-mcp init
# 2. Build the local index
npx context-hub-mcp reindex
# 3. Generate MCP client config
npx context-hub-mcp config --target claude-code --cwd /absolute/path/to/project
Paste the generated JSON into your MCP client config. The client will then invoke serve automatically — you don't need to run it manually.
.context/ Structure
.context/
├── schema.md ← authoring conventions
├── domains/ ← business rules, workflows, state machines
├── integrations/ ← third-party APIs, auth, webhooks
└── pitfalls/ ← gotchas, lessons learned, failure modes
Writing Knowledge
Every document needs YAML frontmatter:
---
title: Payment Rules
domain: payments
tags: [payments, line-pay]
last_verified: 2026-03-28
confidence: high
---
# Payment Rules
## Key Files
- `src/payments/service.ts` - entrypoint
## Payment State Machine
```mermaid
stateDiagram-v2
pending --> paid
```
Fields: title, domain, tags, last_verified, confidence (high / medium / low), related (optional list of related doc paths)
Use Key Files sections for source references and state machine headings so structured parsing picks them up.
The related field accepts paths relative to .context/:
related:
- integrations/stripe-webhooks.md
- pitfalls/payment-idempotency.md
Importing Docs with Claude Code
If you use Claude Code, the repo includes a /import_context command that lets you turn any URL, wiki, or spec into a properly formatted .context/ document:
/import_context https://stripe.com/docs/webhooks
/import_context the spec at ./docs/payment-rules.md
Claude will fetch the source, distill the relevant knowledge, pick the right subdirectory, and write the file with correct frontmatter. Run npx context-hub-mcp reindex afterward to update the index.
MCP Tools
| Tool | Purpose |
|---|---|
list_domains |
Discover available knowledge areas |
search_context |
Full-text search across indexed documents |
get_context |
Read a specific document |
get_context_structured |
Read a document as structured data (keyFiles, stateMachines, pitfalls, sections, related) |
get_pitfalls |
List pitfalls, optionally filtered by domain |
list_tags |
List all unique tags with document counts; optional domain filter |
get_stale_docs |
List documents with old or missing last_verified, or confidence: low |
get_related |
Return documents related to a given path (bidirectional, depth 1 or 2) |
annotate_context |
Leave a note on outdated or missing docs |
rate_context |
Mark whether a document was helpful |
list_annotations |
Review accumulated annotations |
reindex_context |
Force a fresh index rebuild |
search_context — filtering params
In addition to a query string, search_context accepts optional filters:
| Param | Type | Description |
|---|---|---|
tags |
string[] |
OR-match: return docs that have any of the listed tags |
confidence |
string |
Minimum confidence level (low, medium, high) |
verified_after |
string |
Only docs with last_verified on or after this date (YYYY-MM-DD) |
verified_before |
string |
Only docs with last_verified on or before this date (YYYY-MM-DD) |
get_stale_docs — params
| Param | Type | Default | Description |
|---|---|---|---|
domain |
string |
— | Limit to a specific domain |
days_threshold |
number |
90 |
Docs not verified within this many days are considered stale |
limit |
number |
— | Cap the number of results |
get_related — params
| Param | Type | Description |
|---|---|---|
path |
string |
Path of the source document (relative to .context/) |
depth |
1 | 2 |
1 returns directly linked docs; 2 also follows their links |
CLI
npx context-hub-mcp init # scaffold .context/ workspace
npx context-hub-mcp add # create a new context document with frontmatter
npx context-hub-mcp reindex # rebuild the local SQLite index
npx context-hub-mcp doctor # inspect workspace health
npx context-hub-mcp config # generate MCP client config JSON
npx context-hub-mcp serve # run the MCP server (usually invoked by the client)
Run any command with --help for available flags.
add — flags
npx context-hub-mcp add [--domain <name>] [--title <title>] [--tags <a,b>] \
[--confidence high|medium|low] [--template domain|integration|pitfall] \
[--no-edit]
| Flag | Description |
|---|---|
--domain |
Domain subfolder to place the file in (e.g. payments) |
--title |
Document title (used for the filename and frontmatter) |
--tags |
Comma-separated list of tags |
--confidence |
Initial confidence level; defaults to medium |
--template |
Starter template: domain, integration, or pitfall |
--no-edit |
Skip opening the file in $EDITOR after creation |
Client Setup
Claude Code
npx context-hub-mcp config --target claude-code --cwd /absolute/path/to/project
<details> <summary>Example output</summary>
{
"mcpServers": {
"context-hub": {
"command": "npx",
"args": ["-y", "context-hub-mcp@latest", "serve", "--cwd", "/absolute/path/to/project"]
}
}
}
</details>
GitHub Copilot
npx context-hub-mcp config --target copilot --cwd /absolute/path/to/project
<details> <summary>Example output</summary>
{
"mcpServers": {
"context-hub": {
"type": "local",
"command": "npx",
"args": ["-y", "context-hub-mcp@latest", "serve", "--cwd", "/absolute/path/to/project"]
}
}
}
</details>
See also: examples/claude-code.mcp.json, examples/copilot.mcp.json
Configuration
context-hub.config.json is optional. Defaults:
{
"contextDir": ".context",
"dbPath": ".context/context_hub.db",
"watch": true,
"reindexDebounceMs": 250,
"includeGlobs": ["**/*.md"],
"excludeGlobs": ["**/.git/**"]
}
CLI flags override config file values. Relative paths resolve from cwd.
Troubleshooting
Search returns no results — run npx context-hub-mcp reindex, then npx context-hub-mcp doctor. Common causes: missing .context/, malformed frontmatter, file outside indexed globs.
Documents fail to parse — check that all 5 frontmatter fields are present and valid.
Database files appear in git — ensure .context/.gitignore contains context_hub.db, context_hub.db-shm, context_hub.db-wal.
Client cannot connect — verify the client uses stdio MCP, the --cwd path is correct, and the project has been indexed at least once.
License
MIT
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。