An AI chief of staff is an AI system that handles the coordination work a human chief of staff would: triaging what needs attention, tracking open commitments, and keeping one current record of what is in progress. The setups people are actually building do this well for email and calendars. Most of them handle tasks badly, because task tracking gets bolted onto a second app the AI cannot properly read or write.
What people mean by "AI chief of staff"
The term describes an AI system, usually running on Claude or a comparable model, that operates with standing access to your email, calendar and task list rather than answering one question at a time in a chat window. It reads what is happening, flags what needs a decision, and drafts or completes the small stuff on its own.
A widely shared write-up by consultant Charlie Grinnell describes exactly this: an assistant named Ava running on a dedicated Mac Mini, with her own email address and logins, working through email triage, meeting prep and overnight research on a fixed schedule. The build cost $583 in hardware plus roughly $200 a month in AI subscriptions, mostly for Claude API access, and it replaced a $949-a-month human executive assistant. It is a genuinely useful account of what the pattern looks like end to end, and it is also more machine than most people setting out to try this actually need.
Why the task board is the weak point
Look closely at what these builds connect to and a pattern shows up. Grinnell's Ava manages her own task board with columns for Inbox, Backlog, In Progress, Waiting on Charlie and Done, but that board lives inside Basecamp, alongside a separate personal task app for the human side. Alfred, a commercial AI assistant aimed at founders and lawyers at $29.99 a month, triages email and calendars but has no task board at all. Cortex's AI chief of staff answers ownership and incident questions for engineering teams over Slack, which is a different job entirely.
None of that is wrong, but it means the task list is either split across two apps that do not talk to each other, missing outright, or scoped to one team's engineering metrics. An AI chief of staff is only as good as the one place it reads and writes tasks. If that place is a Basecamp project the human never opens, or a Things list the AI cannot touch, the AI ends up managing a to-do list nobody else uses.
A shared board fixes the split
A kanban board that both you and Claude can read and write solves this directly, because there is exactly one record instead of two. Claude does not need its own separate task system distinct from yours. It reads the same columns you look at, and anything it creates shows up where you would have put it yourself.
Set it up with Claude and MCP
You do not need a Mac Mini running around the clock. Claude connects to a kanban board through MCP, the Model Context Protocol, which the official architecture guide describes as a standard way for AI applications to reach external tools and data. Connecting Claude to your kanban board covers the full setup, but the shape of it is short:
- Add the EasyKanban MCP server (
https://easykanb.app/api/mcp) in Claude's connector settings, or runclaude mcp add --transport http easykanban https://easykanb.app/api/mcpin Claude Code. - Sign in with the account that owns your board and approve the connection. No API key changes hands.
- Ask a read-only question first, such as "list my boards and summarize what's overdue," to confirm the connection before letting Claude write anything.
Reading boards and getting a cross-board to-do summary is free and unmetered on every plan. Creating, moving, editing, archiving or deleting cards is free for 50 actions a calendar month, and unmetered on EasyKanban Pro (€6/month or €60/year). That is the entire cost of the task-tracking half of an AI chief of staff, next to Alfred's $29.99-to-$34.99-a-month subscription for email and calendar alone.
What a realistic daily rhythm looks like
Grinnell's build runs a 6:30am briefing and an end-of-day recap. You can get most of the value from two much smaller habits built around the board:
- Morning: ask Claude for a short work plan from your open cards, prioritizing anything overdue or due soon, without changing anything yet.
- Evening or after a meeting: turn meeting notes or a rough idea into a proposed card, with a title and short description, and confirm it before Claude creates it.
A concrete example beats a generic one. After a call with a vendor about a late invoice, the prompt is "turn this into one card on the Ops board, title and short description first, don't create it until I confirm: follow up with Acme Supplies about the September invoice, they said Thursday." Claude proposes the card, you check the board it lands on, and only then does it get written. That two-step pattern, propose then confirm, is what keeps this from turning into an AI quietly restructuring your week.
What it should not do without asking
The permission model in Grinnell's write-up is read-only by default, write access earned gradually: his assistant can post to a shared board but cannot email his clients or spend money without approval. The same split works with a kanban board.
- Let Claude read everything: cards, due dates, comments, the full board.
- Let it propose cards, moves and edits, but review anything destructive before it happens.
- Never let a vague instruction like "clean up my board" stand in for archiving or deleting cards. Ask what "clean" means first, or ask Claude to list what it would remove and wait for your answer.
EasyKanban's board and account permissions apply to the MCP connection the same way they apply to you, so a viewer on a shared board cannot suddenly gain edit access through Claude. You can revoke the connection from Settings at any time, for any client, without touching your password.
Frequently asked questions
What is an AI chief of staff?
An AI system, usually built on Claude or a similar model, that handles ongoing coordination work rather than answering isolated questions: triaging email, tracking commitments, and keeping one current record of what is in progress across your day.
Do I need to buy a dedicated computer or several paid subscriptions?
No. The Mac Mini and multi-app setups you see written up online are one way to build this, not a requirement. Claude connected to a single shared kanban board through MCP covers the task-tracking half for free, with 50 monthly write actions included before EasyKanban Pro is needed.
What can Claude actually do with my kanban board through MCP?
It can read your boards and cards, summarize what needs attention, and, once you approve it, create, move, edit, archive or delete cards. Reading is always free and unmetered. Writing is free up to 50 actions a month and unmetered on Pro.
Is it safe to let an AI chief of staff read my whole board?
It is reasonable if you start read-only, ask for a preview before anything gets written, and keep destructive requests explicit rather than implied. Your existing board and account permissions still apply to what the connection can see and change.
Can I use ChatGPT instead of Claude for this?
Yes. EasyKanban's MCP server works with any client that supports remote MCP over Streamable HTTP and OAuth, including ChatGPT. ChatGPT for project management covers that setup specifically. The board stays the same regardless of which AI client reads it.
Start with the board, not the hardware
The build that gets the most attention online is the one with a dedicated Mac Mini and a long list of connected services. It works, and it costs real money and real setup time to get there. Most of what an AI chief of staff needs to do with your tasks, read what is open, flag what is overdue, turn a rough note into a proper card, works today with Claude and one shared kanban board.
Connect Claude to a board, ask it for a read-only summary, and compare that against what you already know is true. If it holds up, let it propose the next card instead of writing your own.
About EasyKanban
EasyKanban is a free, minimal kanban board built for exactly this kind of setup: a remote MCP server, free unlimited board reads, and 50 free card writes a month before Pro is needed. Start instantly without signup, then save your board when you are ready.
Try EasyKanban Free →References and Further Reading:
- Model Context Protocol: architecture - the open specification behind how Claude reaches external tools like a kanban board
- How I Built an AI Chief of Staff That Actually Works - a detailed account of the hardware-and-multi-app version of this pattern