About
Two things kept being true. Our assistants could describe the backlog but never change it. And customer feedback lived in a different product from the work it turned into, so the person who reported a bug never found out it was fixed.
01 — Problem one
Most tools added an LLM to summarise a screen. That is a demo, not a colleague. We wanted the product itself to be callable, so we shipped a real MCP server — the same 32 operations the interface exposes, over the same permission checks, reachable by whatever assistant a team already uses.
The 32 tools and their scopes02 — Problem two
Collecting feedback is the easy half. The half everyone skips is telling the person it shipped. So that is automatic here — including for people whose duplicate report got merged into someone else’s, who are the ones normally dropped on the floor.
How the loop closes03 — Construction
A project tool is judged on the boring parts: what happens when two people save at once, what happens when the email should have gone out, what happens on the ten-thousandth task. Four choices, made once, that most of the product then inherits.
3.1
A click in the app, an MCP tool call and a REST request take the same path: same permission checks, same validation, same activity record. There is no faster, looser way in, which is why granting an assistant access is a scoping decision rather than a leap of faith.
3.2
Emails, notifications, realtime updates and the activity trail are not fired off inline and hoped for. They are derived from an outbox row written inside the same transaction as the change itself, then processed separately. If the change committed, the follow-up is owed; if it rolled back, nothing was ever queued.
3.3
Every record carries a version. When two people change the same task at the same moment, the second save is told it is stale and gets a merge path — rather than the quiet last-write-wins that makes people distrust a tool without ever being able to prove why.
3.4
Every list is read with keyset pagination — around twenty rows at a time, with the totals computed server-side rather than counted from whatever happens to be loaded. A list with ten thousand tasks opens like an empty one, and the number at the top is right while the rows are still arriving.
04 — Limits
Every gap below is also written into llms.txt, the plain-text summary answer engines read, so an assistant asked about us cannot infer a capability we do not have. Finding this out in week two is worse for you than reading it now, and worse for us.
Not built yet
Not our claim
Planner.day is for teams who want their assistant to operate the board and their users’ feedback to live in the same system as the work. If that is not the thing you are short of, one of those tools is the better call — and we would rather say so than waste your trial.
Have a look
Free while you evaluate. The security posture, including what is missing from it, is written out in the same tone as this page.