DEV Community

Tareq Naji
Tareq Naji

Posted on Fully Autonomous

Why language models should never guess prayer times (and how an MCP tool fixes the workflow)

#ai #programming #opensource #mcp

Prayer times look like a simple question for an AI assistant: “When is the next prayer?”

They are not simple static facts. A useful answer depends on where, when, which calculation convention, and sometimes which school of jurisprudence applies. If the model chooses any of those silently, it can return a confident-looking number that does not match a local timetable.

I built Muslim Prayer Reminder MCP to make these choices explicit and put the calculation in a tool instead of asking an LLM to invent the answer.

Five inputs worth making explicit

  1. Location. A city name can be ambiguous. Precise coordinates are preferable when available, but should not be inferred from a server's IP address.
  2. Time zone and date. A prayer schedule should be evaluated for the user's local date, including daylight-saving changes where applicable.
  3. Calculation authority. Regional methods can use different angles or conventions. A returned timestamp without its method is hard to audit.
  4. Asr convention. The selected juristic convention matters. It should be a visible preference, not an invisible model assumption.
  5. High-latitude handling. Some places and seasons require special rules; a generic formula is not enough for every latitude.

A calculation tool helps with repeatability, but repeatable is not identical to a mosque timetable. Local authorities may publish offsets or other adjustments. The honest answer makes that discrepancy understandable.

What an agent-friendly result should communicate

A response is easier to trust when it identifies the requested date and location, the selected method, the calculated times, and what will happen next.

For example, rather than saying “Maghrib is at 6:10” without context, the assistant should be able to give a timetable tied to the user's city, time zone, and selected regional authority, along with a next-prayer countdown.

The model can then concentrate on explaining the result instead of fabricating the calculation.

Try the open-source MCP

Muslim Prayer Reminder MCP exposes prayer-time tools to compatible AI clients. For a local Node.js-based setup, use the project's documented MCP configuration, which launches:

npx -y muslim-prayer-mcp
Enter fullscreen mode Exit fullscreen mode

Source and installation: https://github.com/tareq7/muslim-prayer-mcp

ChatGPT version: https://chatgpt.com/plugins/plugin_asdk_app_6a9bca660b2c8191a1146aa615cad6f5

The local and hosted versions have different privacy characteristics; the repository documents hosted preferences and pseudonymous usage analytics.

I built and maintain this free, open-source project. I'm especially interested in feedback on how much user context a tool should require up front versus how much can safely come from preferences. How would you design that boundary?

Top comments (0)