The problem
My to-do list grew faster than I could work through it. Looking at it, a lot of the list was work an AI could do: fix a failing build, write up some research, tidy my notes. The rest (call the dentist) still needs me.
What I built
Hermes is an assistant that runs on my Mac. It’s built on Hermes Agent, an open-source agent framework, and it hands the real work to Claude Code.
- Every 15 minutes it reads my Todoist and rates each open task: it can do it, it can prep it, or it’s mine.
- It picks the next task it can do (highest priority first) and texts me a short plan: the goal, three steps, the rough cost and the risk.
- Nothing runs until I reply yes. No means skip it; anything else means change the plan and ask again.
- Claude Code does the work on its own branch and opens a draft pull request with a screenshot review.
- If a check fails, it fixes it and runs the checks again.
- It only merges when I say “merge it”, and only if every check passed. Then it ticks the task off in Todoist.
The limits that make it safe
| Limit | Setting |
|---|---|
| Tasks a day | 5 at most, 2 at once |
| Hours | 7am to 11pm |
| Spending | A daily cap; it stops when the cap is hit |
| Code | Always a draft PR, never pushed to main or deployed |
| Merging | Only when I say so, after green checks |
| Sending, booking, buying | Always asks first |
| Off switch | One command halts everything |
Build your own, step by step
- Start with the agent. Install Hermes Agent from its GitHub repo and get it answering you in one chat channel you already use.
- Give it your task list, read-only. Connect Todoist through its official MCP server and allow only the read tools at first.
- Write the rating rule. One short prompt that sorts each task into can do, can prep or mine. Leave anything that sends, books or buys in mine.
- Make “plan, then yes” the only way in. The agent writes a three-step plan and waits for your reply before it starts any job.
- Run the work in Claude Code, fenced. Each job gets its own folder or git branch and a turn and spending limit, and code ends as a draft pull request.
- Add the caps last: tasks a day, jobs at once, quiet hours, a daily spending cap and a kill switch.
- Run it for a week on easy tasks (notes, research) before you let it near code.
Comment AGENT on the reel and the DM sends you straight back to this page.
