mcp-agent
Calendar and task management with kernel-level scraping for robust parsing, and integrates with DocMost wiki as source of truth.
README
mcp-agent
SimpleMCP Agent — Calendar & Task management with kernel-level scraping.
Architecture
┌─────────────────────────────────────────────────────────────────────┐
│ USER / CLI │
└───────────────────────────────┬─────────────────────────────────────┘
│
┌──────▼──────┐
│ MCP Host │ mcp_host/host.py
│ (REPL/API) │
└──────┬──────┘
│
┌──────▼──────┐
│ Agent │ mcp_host/agent.py
│ (ReAct loop)│ DocMost-first system prompt
└──────┬──────┘
│
┌────────────────┼────────────────┐
│ │ │
┌──────▼──────┐ ┌──────▼──────┐ │
│ MCP Client │ │ Anthropic │ │
└──────┬──────┘ │ LLM │ │
│ stdio └─────────────┘ │
┌──────▼────────────────────────┐ │
│ MCP Server │ │
│ Tools │ │
│ ├── calendar (list/search) │ │
│ ├── tasks (CRUD) │ │
│ └── docmost (source-of-truth)│ │
└───────┬───────────────┬───────┘ │
│ │ │
┌───────▼──────┐ ┌──────▼────────────┐ │
│Kernel Scraper│ │ DocMost REST API │ │
│ (httpx+BS4) │ │ integrations/ │ │
└───────┬──────┘ │ docmost/client.py │ │
│ └──────────────────-┘ │
┌───────▼──────┐ │ │
│Calendar Feed │ ┌────────▼──────────┐ │
│ (ICS / HTML) │ │ DocMost Instance │ │
└──────────────┘ │ (self-hosted wiki)│ │
└───────────────────┘ │
Layer Responsibilities
| Layer | Module | Role |
|---|---|---|
| Config | config/ |
Pydantic settings, .env loading |
| Scraper — Kernel | scraper/kernel/ |
Raw async HTTP, no browser |
| Scraper — Session | scraper/kernel/session_manager.py |
Cookie jar, auth tokens, idle recycling |
| Scraper — Time | scraper/kernel/time_manager.py |
Rate limiting, politeness delays |
| Scraper — Parser | scraper/parsers/calendar_parser.py |
BS4 kernel parsing (JSON-LD → Microdata → data-* → ICS) |
| DocMost Client | integrations/docmost/client.py |
REST API client for DocMost wiki |
| DocMost Models | integrations/docmost/models.py |
Page, Space, SearchResult dataclasses |
| DocMost Content | integrations/docmost/content.py |
ProseMirror JSON → Markdown converter |
| MCP Server | mcp_server/ |
Tool & resource registration, MCP protocol |
| MCP Client | mcp_client/ |
Server connection, tool invocation, schema conversion |
| LLM | llm/ |
Abstract BaseLLM + Anthropic implementation |
| Agent | mcp_host/agent.py |
ReAct loop; DocMost-first system prompt |
| Host | mcp_host/host.py |
Lifecycle, REPL, programmatic API |
| Use Case | use_cases/calendar_tasks/ |
Demo workflow, entry point |
DocMost — Source of Truth
DocMost is the team wiki. The agent is instructed to treat it as the authoritative source of truth for all factual questions. The integration has two modes:
Mode 1 — REST API (default, any DocMost version)
DOCMOST_BASE_URL=https://docs.yourcompany.com
DOCMOST_API_KEY=<personal API key from Settings > API keys>
DOCMOST_DEFAULT_SPACE_ID=<optional space UUID>
The REST layer (integrations/docmost/) handles:
- Full-text search →
POST /api/search - Page read (ProseMirror JSON → Markdown) →
POST /api/pages/info - Page create/update (Markdown input) →
POST /api/pages/create,/api/pages/update - List spaces and pages →
GET /api/spaces,POST /api/pages
Mode 2 — Native MCP (DocMost enterprise)
DocMost exposes its own MCP endpoint at {BASE_URL}/mcp. If you have an
enterprise license and MCP enabled (Settings > AI & MCP > MCP tab), set:
DOCMOST_USE_NATIVE_MCP=true
Then add to claude_desktop_config.json:
{
"mcpServers": {
"docmost": {
"command": "npx",
"args": ["-y", "mcp-remote", "https://docs.yourcompany.com/mcp",
"--header", "Authorization: Bearer YOUR_API_KEY"]
}
}
}
DocMost tools exposed via MCP
| Tool | Description |
|---|---|
docmost_search |
Primary source-of-truth lookup — call before answering factual questions |
docmost_get_page |
Retrieve full page content as Markdown |
docmost_list_spaces |
Discover available knowledge spaces |
docmost_list_pages |
Browse pages in a space (paginated) |
docmost_create_page |
Write new wiki page (meeting notes, summaries) |
docmost_update_page |
Append or replace content on existing page |
docmost_delete_page |
Remove a page |
Agent source-of-truth flow
User asks a question
│
▼
docmost_search("keywords from question")
│
├─ Results found → docmost_get_page(page_id) → ground answer in wiki
│
└─ No results → fall back to calendar/task data + flag as inferred
│
▼
Produce response citing wiki page title + URL
│
▼
(Optional) docmost_create_page / docmost_update_page to persist findings
Kernel Scraping vs Layout Scraping
This project deliberately uses kernel-level scraping — we operate at the semantic/structural layer of HTML, not the visual/CSS layer.
| Approach | Targets | Stability |
|---|---|---|
| Layout scraping | CSS classes, visual position | ❌ Breaks on every redesign |
| Kernel scraping ✅ | JSON-LD, Microdata, data-*, ICS, ARIA |
✅ Survives redesigns |
Parser strategy (priority order)
1. JSON-LD <script type="application/ld+json"> Event schema
2. Microdata itemtype="https://schema.org/Event"
3. data-* attrs data-event-id, data-start-time, data-event-title …
4. ICS/iCal Raw text; no external library needed
Tools exposed by the MCP Server
| Tool | Description |
|---|---|
list_events |
Upcoming events within N days |
get_event_by_uid |
Single event by UID |
search_events |
Full-text search across title + description |
get_todays_schedule |
Today's events |
create_task |
Create a task (optionally linked to an event) |
list_tasks |
List tasks with status/priority filters |
update_task |
Update any task field |
delete_task |
Delete a task |
Quick Start
# 1. Install
bash scripts/install.sh
# 2. Configure
# Edit .env — set ANTHROPIC_API_KEY and CALENDAR_FEED_URL
# 3. Run interactive agent
source .venv/bin/activate
mcp-agent
# 4. Or run the demo sequence
mcp-agent --demo
# 5. Or run just the MCP server (for use with Claude Desktop etc.)
mcp-server
Project Structure
mcp-agent/
├── config/ # Pydantic settings
│ └── settings.py
│
├── integrations/ # ← NEW: third-party integrations
│ └── docmost/
│ ├── client.py # DocMost REST API async client
│ ├── models.py # Page, Space, SearchResult dataclasses
│ └── content.py # ProseMirror JSON → Markdown converter
│
├── scraper/ # Kernel scraper layer
│ ├── kernel/
│ │ ├── http_client.py
│ │ ├── session_manager.py
│ │ └── time_manager.py
│ └── parsers/
│ └── calendar_parser.py
│
├── mcp_server/ # MCP Server
│ ├── server.py
│ ├── tools/
│ │ ├── calendar.py
│ │ ├── tasks.py
│ │ └── docmost.py # ← NEW: DocMost tool handlers
│ └── resources/
│ └── calendar_resource.py
│
├── mcp_client/
│ └── client.py
│
├── llm/
│ ├── base.py
│ └── anthropic_llm.py
│
├── mcp_host/
│ ├── agent.py # DocMost-first system prompt
│ └── host.py
│
├── use_cases/
│ └── calendar_tasks/
│ └── workflow.py
│
├── tests/
│ ├── test_scraper.py
│ ├── test_session_time.py
│ ├── test_tasks.py
│ └── test_docmost.py # ← NEW
│
├── scripts/install.sh
├── pyproject.toml
└── .env.example
Running Tests
pytest tests/ -v
All tests are offline — no network or API key required.
Claude Desktop Integration
Add to claude_desktop_config.json:
{
"mcpServers": {
"calendar-agent": {
"command": "python",
"args": ["/path/to/mcp-agent/mcp_server/server.py"],
"env": {
"ANTHROPIC_API_KEY": "sk-ant-...",
"CALENDAR_FEED_URL": "https://your-calendar.example.com/feed.ics"
}
}
}
}
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。