Run Empi from Claude over MCP
Connect an agent over MCP and it stops reading your tasks and starts driving them: filing the backlog, re-estimating what slipped, and booking tomorrow against real capacity.
Most “AI integrations” in task apps are a summarize button. You ask, it reads, it produces a paragraph. Nothing in your data changes.
Empi ships an MCP server instead, and the difference is the verb. Your agent does not read your board. It drives it.
What the agent actually gets
Sixteen tools, the same ones the app itself uses:
| Area | Tools |
|---|---|
| Tasks | list_tasks, get_task, create_task, update_task, complete_task, delete_task |
| Projects | list_projects |
| Notes | list_notes, get_note, create_note, update_note, delete_note, convert_note_to_task |
| Planning | today, planner_view, planner_apply |
So this works:
Turn the review comments on PR 412 into tasks, 45 minutes each, due Thursday, under the Billing project. Then re-estimate “rewrite the docs” – it’s clearly bigger than an hour. Then plan tomorrow.
The agent calls create_task three times, update_task once, then planner_view, and comes back
with a plan you can look at.
The planner is the part it cannot fake
This is the design decision that makes an agent trustworthy here.
A markdown to-do file lets an agent write whatever it likes, including “you have five free hours tomorrow” when you do not. There is nothing to check it against.
planner_view returns the same changeset the app renders: your actual meetings, your actual
capacity, your actual over-capacity warnings. If tomorrow only holds three hours, the plan the agent
gets back only holds three hours. Hallucinated free time is not a value the tool can return.
And planning is deliberately two calls:
planner_viewproposes. It is a dry run: nothing is written, nothing is booked.planner_applycommits, and the agent has to ask for it explicitly.
You keep the veto on the only operation that touches your calendar.
Setting it up
- In Empi, open Settings → Integrations → AI assistant and copy the MCP server URL.
- Add it as a connector in Claude (or ChatGPT, or Mistral, or your own MCP client).
- Authorize once with OAuth. The connection is scoped to your workspace and revocable in a click.
That is the whole setup. There is no API key to paste into a config file, and nothing to rotate.
What it costs
Nothing extra. Agent requests are unlimited on every plan, Free included: usage is recorded so we can watch the system, but a request is never refused for being over a quota. An assistant you have to ration is an assistant you stop reaching for.
Every write the agent makes shows up in the app exactly like a change you made by hand, so there is never a mystery about who moved what.
The shape of the thing
The useful framing is not “AI in a task app”. It is: your agent now has a workspace with a constraint solver in it. It can propose, but the hours are not negotiable, and the calendar pushes back.
That is a much better assistant than one which agrees with you.
Start free and connect your assistant.