Today, computed in your own timezone
Tasks, meetings, follow-ups, bookings and drafts awaiting approval, ranked, with "today" computed in the reader's timezone rather than the server's. Names which sources it could not read.
Works todayEighteen of eighteen capabilities liveVirtual Assistant
Tasks ranked with reasons, drafts that stay drafts until you say otherwise.
A priority score that shows its working, booking pages that cannot double-book, focus blocks that only ever add time nobody has claimed. Every outbound action is a draft until a person approves it, and the approval is bound to a hash of the exact content — edit an approved draft and it drops back to draft, so nothing sends wording nobody read.
Scan the mechanism, not a pitch
Approval bound to a content digestOpt-in checked twice before a sendBooking slots re-checked at write time
18/18specified capabilities work today, at least in part
SHA-256digest an approval is bound to — any edit clears it
90 minthe length of every focus block Mitra plans
What it does
Eighteen things Mitra runs today.
Every card's status label is derived from the backend's own capability module.
A note becomes a task a person confirms
A typed note, a pasted email body, or a transcribed voice memo becomes a reviewable capture. Nothing turns into a task without someone confirming it.
Works todayA score that always shows its reasons
Scored from deadline proximity, stated priority, age and open dependencies, with a written reason per component. A manual override wins outright and says so.
Works todayA brief becomes a draft plan, not a live one
A five-stage plan whose middle tasks are your own sentences with spread deadlines — deterministic, nothing invented, no credential needed — and it stays draft until you accept it. Asking a model to split the brief into work items instead needs an LLM key, and refuses rather than silently falling back if asked for without one.
Needs a keyStatus computed fresh on every read
Plain-English status from this workspace's own tasks, with the counts it was written from. Never stored, so it cannot go stale against the board it describes.
Works todayTrackers that know when a reply lands
Windows, snoozing and a due list all work locally, keyed to a provider message id rather than a subject line. An expired window raises Mitra's own in-app bell notification — no phone number or provider needed. There's no Gmail or Microsoft Graph OAuth in this deployment to watch a real inbox, so resolving a tracker is a route your own mail automation (a forwarding rule, an Apps Script trigger) calls, not something Mitra reaches into your inbox to detect on its own.
Works todayDelivery over Wavy's own consented rail
Scheduling and dispatch are real, and the opt-in gate is Wavy's own consent record, checked at schedule time and again at send. A dry run is recorded failed, never sent.
Needs a keyA booking page that cannot double-book
Public page, real open slots from stored busy time, and availability re-checked in the same transaction as the booking — so two people racing for 11:00 cannot both get it.
Works todayOverlap for who you can actually read
Free time across participants whose busy blocks this workspace holds. An external participant comes back unreadable with a booking-link suggestion, never assumed free. Every query is now saved with its result, so looking it back up later answers with what was actually seen rather than recomputing against busy time that has since changed.
Works todayOnly Mitra's own holds ever move
Tasks placed into open time around existing commitments — now on every task create and update, not only on demand. A task that will not fit is reported unplaced rather than squeezed in; a newly-added fixed event re-slots any of Mitra's own holds it now overlaps. Only Mitra's own calendar is written — two-way sync with a real Google or Microsoft calendar needs OAuth this deployment does not hold.
Works todaySuggests time. Never declines on your behalf.
A weekly goal refilled into 90-minute blocks in open time, with utilisation tracked against it. A new busy block overlapping an active focus block is now flagged and still created — suggest, don't hide — with one call to release the flag. Declining or proposing a new time to a real invitee is deliberately absent — this endpoint has no outbound side at all.
Works todayMinutes per category, week over week
Compared with the prior week, clipped to the window. Meetings held only on an unconnected calendar are not counted, and the response says so rather than reporting a quiet week. A workspace can now define its own named categories, matched first-match- wins so a minute of busy time is never double-counted.
Works todayNotes a person typed, tasks in one click
Action items you listed convert into tracked tasks with a link back to the note, no credential needed. Asking a model to summarise the note and suggest action items of its own needs an LLM key, and a suggestion stays separate from your real list until you promote it — the same capture-then-confirm shape task capture uses. Still no transcript path — joining a call as a recording bot needs every participant's consent.
Needs a keyDraft, edit, approve, send — one gate for all four
This is a draft until a person approves it, and the approval is bound to a SHA-256 digest of the exact content it covers. Edit an approved draft — even one word — and it drops back to draft, because an approval that survived the edit would have approved bytes nobody read. Generating the draft needs a model key; sending needs mail access.
Needs a keyDetects the script, says when it's guessing
Ten Indian scripts, plus Latin. Devanagari is reported as Hindi with a note that Marathi shares the script; Hinglish is reported as indistinguishable from English rather than guessed — and detection plus the confirm-before-create capture both need no credential. Replying in that language, with a plain-English gloss of what it understood so you can confirm at a glance, needs a model key.
Needs a keyClassify and label mail, without reading the body
Rule creation, the classifier and retroactively labelling recent mail under a newly-active rule all run locally, storing metadata only — no message body retained, archiving opt-in per rule. There's no Gmail or Microsoft Graph OAuth in this deployment, so applying a label or archiving inside a real inbox needs your own mail automation to call this route with your own platform credential — nothing here writes back to a real mailbox on its own.
Needs a keyTurn a plain-text mention into a calendar draft
Deterministic date, time and attendee extraction from free text — day-after-tomorrow, weekday names, explicit times, email addresses — always shown back as a draft to confirm or discard, nothing auto-created. Confirming writes to Mitra's own calendar and re-slots anything it now overlaps. A real Google or Outlook invite, instead of a local block, needs calendar OAuth this deployment does not hold.
Works todayA recurring series, without double-booking a slot
The next four occurrences generate on creation, each its own local calendar block. A colliding slot nudges inside the rule's flexibility window and is marked skipped rather than double-booked when no room remains. Rescheduling an occurrence with external participants is refused until you explicitly confirm, so a client's standing call is never silently moved. Only Mitra's own calendar is written; a real invite needs calendar OAuth this deployment does not hold.
Works todayDraft → approved → sent
What each action does to the approval.
The most common support question this mechanism produces is answered by this table: an edit clears an approval, on purpose.
| Action | What happens to the approval |
|---|---|
| A draft is created | A content digest is set. Status: draft. |
| A person approves it | The current digest is stored as approved_digest. Status: approved. |
| The draft is edited — even one word | The digest is recomputed, approved_digest is cleared, and status drops back to draft. |
| Send is attempted and the digest still matches | Sends, under "approval" authority. |
| Send is attempted after an edit broke the match | Refused with approval_stale — approve the new version first. Nothing goes out. |
| An auto-send rule is enabled | Only WhatsApp and SMS may ever be covered, never email, and never a message that reaches outside the business — that still needs an explicit approval. |
How it looks
Two screens, and what makes each one honest.
Limits
Real limits, read from the code.
Every figure here is a constant or a default in the codebase.
| Priority score range | 0–100 |
|---|---|
| Focus block length | 90 minutes |
| Booking page — default minimum notice | 120 minutes |
| Booking page — default horizon | 30 days ahead |
| Booking page — default daily cap | 20 confirmed bookings |
| Free-slot search — granularity and page size | 15-minute steps, up to 40 slots returned |
What it refuses to do
Six things Mitra will not do, on purpose.
- It will not send anything without a per-action approval, unless a narrow auto-send rule explicitly covers it.
- It will not let auto-send cover email, an invite response, a schedule change, or anything reaching outside the business.
- It will not keep an approval alive past an edit — one changed character clears it.
- It will not message a contact with no recorded opt-in — checked at scheduling and again at the moment of send.
- It will not let two people book the same slot — availability is re-checked in the booking transaction, not just when the page loaded.
- It will not decline or propose a new time to a real invitee to protect your focus time — it only ever adds blocks to open time.
Before you switch
See your own week ranked, with the reasons shown.
Tell us how your day gets planned today and we will show you exactly what Mitra carries — and what still needs a credential this deployment hasn't set.
Questions
Straight answers.
I approved a draft, fixed a typo, and now I can't find the send button. Is it broken?
No — it worked as designed. Every outbound action is a draft until a person approves it, and the approval is bound to a SHA-256 digest of the exact content it covers. Edit an approved draft — even one word — and it drops back to draft, because an approval that survived the edit would have approved bytes nobody read. Approve the corrected version and the send option returns. This is deliberate: an approval that survived your edit would have approved wording you never actually read.
Can Mitra send messages on its own, with nobody reviewing them?
Only in one narrow case: an auto-send rule you turned on, and only for WhatsApp or SMS — never email, never an invite response, never a schedule change. Even then, anything that would reach someone outside the business still needs an explicit approval.
What does the priority score actually mean?
A number from 0 to 100 built from deadline proximity, stated priority, how long a task has waited, and whether it is blocked — with a written reason for every component that moved it. Set your own number instead and it wins outright, labelled as your decision, and survives every recompute.
Can two people book the same slot on my booking page?
No. The page shows open slots from stored busy time, but the moment someone submits, availability is re-checked and the booking is written in the same transaction — so a second person racing for the same time finds it already taken, not double-booked.
Will Mitra text a customer who never agreed to hear from us?
No. WhatsApp reminders reuse Wavy's own consent records — the same ones Wavy's broadcasts check — and the opt-in is checked twice: once when the reminder is scheduled, and again at the instant it would send.
Will Mitra decline a meeting invite to protect my focus time?
No. Focus planning only ever adds blocks into time nobody has claimed yet. The specification's own rule is "suggest, don't send", and this endpoint has no outbound side at all — nothing is declined, and nobody but you is written to.