cal-scheduler
A thin, timezone-correct MCP server for scheduling against CalDAV calendars, providing tools for creating, updating, and managing events with proper timezone handling and recurrence support.
README
<!-- mcp-name: io.github.limey/cal-scheduler-mcp -->
If you are an agent installing this MCP, read AGENTS.md first.
cal-scheduler
A thin, timezone-correct MCP server for scheduling against any CalDAV calendar. It gives an LLM agent a small, strict set of calendar tools and does the deterministic, fiddly parts of iCalendar correctly so the model doesn't have to.
It is not an NLP layer: the agent phrases the request, the server validates
it, rejects bad input loudly, and persists clean zoned .ics.
Why it exists
Most calendar tooling an LLM reaches for gets three things subtly wrong. cal-scheduler fixes them by construction:
- Stores zoned, not bare UTC. Events are written with
TZID/VTIMEZONE, so a weekly 9am stays 9am across a daylight-saving boundary instead of drifting an hour. Storing bare UTC is the classic cause of that drift. - Validates recurrence. It rejects an
RRULEwhose anchor contradicts it (e.g. a series starting on the 30th but set to repeat on the 1st), and normalisesUNTILto UTC under a zonedDTSTARTas RFC 5545 requires. - Does real single-occurrence edits. Exclude one instance (
EXDATE) or move one instance (RECURRENCE-ID) without disturbing the rest of the series — the operations naive wrappers tend to lack.
How it works
A small uv Python package that composes mature libraries rather than
implementing a calendar engine:
| Module | Role |
|---|---|
config.py |
environment config (CALDAV_*, CAL_DEFAULT_TZ) |
timezones.py |
parse datetimes; naive → assume default-zone wall time, offset → normalise into the zone; report what was assumed |
ical.py |
build/parse VEVENTs (icalendar), expand ranges (recurring-ical-events), recurrence validation, EXDATE/RECURRENCE-ID ops |
store.py |
CalDAV transport (caldav) — list/create/delete calendars, get/put/delete events, read-modify-write with etag |
server.py |
the FastMCP stdio server and the tool surface |
It runs as a stdio MCP server that an MCP host (Claude, an agent harness, etc.) spawns as a subprocess.
Install
Requires Python ≥ 3.11. The package is on PyPI as
cal-scheduler-mcp:
uv tool install cal-scheduler-mcp
(or pip install cal-scheduler-mcp, uvx --from cal-scheduler-mcp cal-scheduler,
etc.)
For editable dev work, install from a clone:
# run straight from the repo with uv (no global install)
uv run --directory /path/to/cal-scheduler-mcp cal-scheduler
# or install the console script into a tool environment
gh repo clone limey/cal-scheduler-mcp
uv tool install --editable /path/to/cal-scheduler-mcp
# or, with SSH GitHub access, in one step:
uv tool install git+ssh://git@github.com/limey/cal-scheduler-mcp
Don't
uv add cal-scheduler-mcpfor the MCP —uv addwrites into whatever project you're sitting in, not into the tool environment. For an MCP server (spawned as a subprocess),uv tool installis the right shape.
Configure
All configuration is via environment variables:
| Variable | Required | Default | Meaning |
|---|---|---|---|
CALDAV_BASE_URL |
✅ | — | CalDAV server URL, e.g. http://127.0.0.1:5232 |
CALDAV_USERNAME |
— | CalDAV account user | |
CALDAV_PASSWORD |
— | CalDAV account password | |
CAL_DEFAULT_TZ |
Pacific/Auckland |
IANA zone every event is stored in |
Many MCP hosts strip inherited environment from stdio servers, so set these in the
host's per-server env block rather than relying on the ambient shell.
Example MCP host config
The MCP runs as a stdio subprocess that the host spawns. Many hosts strip
inherited PATH from that subprocess, so wire the uv run --directory
form rather than relying on the cal-scheduler shim being on the spawn
host's PATH:
{
"mcpServers": {
"cal-scheduler": {
"command": "uv",
"args": ["run", "--directory", "/abs/path/to/cal-scheduler-mcp", "cal-scheduler"],
"env": {
"CALDAV_BASE_URL": "http://127.0.0.1:5232",
"CALDAV_USERNAME": "me",
"CALDAV_PASSWORD": "secret",
"CAL_DEFAULT_TZ": "Pacific/Auckland"
}
}
}
}
/abs/path/to/cal-scheduler-mcp is the absolute path to a local clone of
this repo (see Install above).
If you installed from PyPI with uv tool install cal-scheduler-mcp, use the
installed shim directly:
{
"mcpServers": {
"cal-scheduler": {
"command": "/path/to/cal-scheduler",
"env": {
"CALDAV_BASE_URL": "http://127.0.0.1:5232",
"CALDAV_USERNAME": "me",
"CALDAV_PASSWORD": "secret",
"CAL_DEFAULT_TZ": "Pacific/Auckland"
}
}
}
}
Find the shim path with which cal-scheduler or uv tool list --show-paths.
Pair it with any CalDAV server. A simple self-hosted option is
Radicale (plain http://, no TLS needed for local use).
Tool surface
Events — list_events(start, end, [calendar]), create_event(summary, start, [end, calendar, description, location, rrule]), update_event(uid, …),
delete_event(uid, [calendar]), exclude_occurrence(uid, occurrence, [calendar]),
move_occurrence(uid, occurrence, new_start, [new_end, calendar]),
mark_done(uid, [calendar]) — tag an event done and record the timestamp.
Calendars — list_calendars, create_calendar(name), delete_calendar(name).
Helper — resolve_datetime(value) — preview how a datetime will be interpreted,
without writing anything.
The timezone rule (the whole point)
Every event is stored zoned to CAL_DEFAULT_TZ.
- A naive datetime (
2026-06-30T21:00) is assumed to be wall time in that zone, and the tool response says so ("assumed Pacific/Auckland wall time"). - An offset-qualified datetime (
…+12:00) is honoured as an instant and re-expressed in the zone — same wall clock when the offset matches, a correct conversion otherwise. - A date-only value (
2026-06-30) is an all-day event.
Develop
uv sync # install deps + dev tools
uv run ruff check # lint
uv run pytest # unit tests (no server required)
The unit tests cover the pure layers (timezone resolution, recurrence validation,
EXDATE/RECURRENCE-ID construction) and need no running CalDAV server. To exercise
the full stack end to end, point CALDAV_BASE_URL at a throwaway CalDAV account.
License
MIT © 2026 Robert Clark
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。