Work log → Telegram

What this is

MCP server that logs what you got done throughout the day via a log_work tool, keeping one Markdown file per day with entries grouped by project. The day's log goes to Telegram two ways: automatically at 11 PM, or on demand when you ask Claude to send it.

Related MCP server: Daily Work MCP Server

Architecture

• server.py — MCP server. Two tools: log_work(entry, project) files a bullet under ## <project> in logs/<date>.md; send_summary(day) posts that day's log to Telegram on request.

• send.py — standalone script, reads a day's log, posts to Telegram. Also the shared implementation the MCP tool calls.

• Scheduled via Windows Task Scheduler at 11 PM daily (register-task.ps1)

• ~/.claude/CLAUDE.md — the parent prompt telling Claude to log as it works, in every project

The scheduled task only needs send.py, so the nightly summary keeps working even when no MCP server is running. That script is stdlib-only for the same reason — the task doesn't have to activate the virtualenv.

Setup

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt

Copy-Item .env.example .env   # then fill in the two Telegram values

Telegram setup

Bot token and chat_id are obtained via BotFather + getUpdates:

1. Message @BotFather → /newbot → copy the token.

2. Send your new bot any message (a bot can't start the conversation).

3. Open https://api.telegram.org/bot<TOKEN>/getUpdates and copy result[0].message.chat.id.

4. Put both in .env as TELEGRAM_BOT_TOKEN and TELEGRAM_CHAT_ID.

.env is gitignored. The scheduled task reads it from disk, so real environment variables aren't required — but if both are set, the environment wins.

Wiring the MCP server to Claude Code

claude mcp add worklog -- "E:\Reels graphics\code logs\.venv\Scripts\python.exe" "E:\Reels graphics\code logs\server.py"

Or add it to .mcp.json by hand:

{
  "mcpServers": {
    "worklog": {
      "command": "E:\\Reels graphics\\code logs\\.venv\\Scripts\\python.exe",
      "args": ["E:\\Reels graphics\\code logs\\server.py"]
    }
  }
}

Add it once at user scope (claude mcp add --scope user ...) so every project logs to the same folder, not just this one.

The parent prompt

Tool descriptions alone make logging optional. ~/.claude/CLAUDE.md loads into every session and carries the standing instruction:

• log after each meaningful unit of work, one call per item, project set to the working directory's name

• don't log routine steps, don't batch a whole session into one entry, don't ask permission first

• call send_summary when you ask for it, never unprompted

Edit that file to change the logging behaviour — it's the lever, not the tool descriptions.

Scheduling

powershell -ExecutionPolicy Bypass -File .\register-task.ps1

Registers WorkLogSummary at 23:00 daily, running under your account while logged on. Override with -At "22:30" or -TaskName "...".

Start-ScheduledTask -TaskName "WorkLogSummary"              # fire it now
Get-ScheduledTaskInfo -TaskName "WorkLogSummary"            # last run + result
Unregister-ScheduledTask -TaskName "WorkLogSummary" -Confirm:$false

The task passes --skip-if-empty, so days with no logged work send nothing. Drop that flag in register-task.ps1 if you'd rather get a "Nothing logged today" ping as proof the job ran.

Sending on demand

Just ask Claude — "send me today's log", "send it", "text me what I did yesterday" — and it calls the send_summary tool. From the shell:

python send.py --dry-run              # print, don't send
python send.py --date 2026-08-01      # re-send an older day
python send.py                        # send today's

Log format

logs/2026-08-02.md:

# 2026-08-02

## Reels graphics

- 09:14 Rebuilt the beat cards in comic style
- 11:02 Added project grouping to the daily log

## contextvolt

- 14:30 Fixed the bolt glyph in the logo

New entries are inserted at the end of their project's section, so a day spent switching between projects still reads as one block per project. A section is created the first time that project is logged.

Plain Markdown on purpose — readable on its own, easy to grep or feed somewhere else later. send.py re-reads the bullets and reformats them; it does not condense with an LLM. If you want summarizing instead of verbatim bullets, that's the place to add it.