jigripokri/poha

POHA stands for Personal Overnight Helper Agent. It is a setup that turns the Claude Desktop app into a personal assistant that runs while you sleep and emails you a morning brief before your alarm goes off. The name is a pun: poha is also a real breakfast dish, and this agent is meant to be the breakfast you wake up to. The author says it can be forked and put into use in about 30 minutes and runs on top of an existing Claude Pro or Max subscription rather than a separate paid service. The README frames POHA as different from a normal chatbot in three ways. It observes your inbound messages on WhatsApp, Gmail, Calendar, Tasks, and SMS without being asked. It remembers across sessions, using plain markdown files you can open and edit by hand instead of a hidden database. And it acts on a schedule rather than waiting for you to type. The wiring for all three is included in the repository. Out of the box you get seven scheduled jobs. A 5am morning briefing scans your messages, mail, calendar, and tasks, flags commitments and unanswered threads, and emails you the result with one-tap acknowledgment links. A 9pm evening reflection updates memory. A Sunday weekly review prunes stale notes. A 7am sweep turns vague verbal commitments into real calendar events. A Friday noon job builds a get-it-done punch list, and another Friday job suggests weekend activities. A monthly job on the 28th does a financial sweep. There are also five on-demand commands like /draft for writing messages in your voice, /reply for suggested responses, /wildcard for a contrarian take, /gsd for a priority list, and /roast for a group chat joke. The architecture has three layers: a briefing engine of scheduled tasks, a memory system of seven markdown files covering people, commitments, life, insights, finances, health, and a private gated file, and a layer of on-demand skills. These read and write through MCP connectors to outside services like WhatsApp, Gmail, Google Calendar, Google Tasks, Android Messages, Fitbit, and Monarch Money. A clever detail is the acknowledgment loop: each briefing email includes one-tap mailto links that send a coded subject back to yourself, which the next run reads to mark commitments done. To set it up you install Claude Desktop with Cowork mode, clone the repo, mount it in Cowork, fill in placeholders in CLAUDE.md, connect the MCPs you want, and ask Claude to onboard you. The author lists costs, privacy notes, and a frank section on who will not get value from it, like people who want a normal chat bot or who do not want an AI reading their email.