Mitra documentation

A note, a pasted email body or a transcribed voice memo becomes a reviewable capture. Confirming it — with a real title and due date a person supplies — is what actually creates the task.

Will not: Parse a date out of free text and schedule it — 'day after tomorrow evening' guessed wrong is the exact failure this two-step design avoids. suggested_title is only the capture's first line, nothing more.

Scored from deadline proximity, stated priority, age and whether it is blocked, each component appending a human-readable reason. A manual override lives in its own column and replaces the component reasons with a single one naming who set it and why.

Will not: Return a score with no explanation — every component contributes a written reason, persisted alongside the task and returned in every response that touches it.

A brief becomes a five-stage draft plan (Define, Plan, Do, Review, Launch) whose middle-stage tasks are the owner's own sentences with spread deadlines — sentence-split, not invented — held as a draft until a person accepts it. This deterministic path works with no credential and is the default on POST /projects/generate. use_llm=true instead asks a model to split the brief into discrete work items, and returns a 503 naming the missing key rather than silently falling back if asked for without one configured.

Needs: LLM_API_KEY, only for the richer use_llm=true path — the default deterministic scaffold needs none

Will not: Invent a task the brief did not mention — an empty brief produces one honest frame task rather than plausible-sounding filler. Activate a generated plan on its own — it stays draft until POST /accept is called, either way it was generated. Silently fall back to the deterministic scaffold when use_llm=true is asked for with no key configured — that is a 503 naming the gap.

Plain-English status computed from the workspace's own tasks on every read, never stored — so it cannot go stale against the board it describes.

Will not: Include a task's private notes in a summary meant to be shared — only titles, counts and dates are exposed.

A weekly focus goal refilled into open slots as fixed 90-minute blocks, with utilisation reported against it and a one-call override to give a block back. focus-conflict-guard now runs on every new busy block: one that overlaps an active focus block is flagged in focus_conflicts and the block is still created — suggest, don't send, never a silent hide — with the override as the one-call release.

Will not: Decline a meeting invite or propose an alternate time on your behalf — that is an outbound action needing calendar OAuth and its own approval, deliberately absent. Leave a side effect when you override a block — releasing a focus block declines nothing and notifies nobody.

Minutes per category for the owner's own week, compared with the prior week, computed from busy time recorded in this workspace. A workspace can now define its own named categories matching by busy-block kind or title, and the weekly report adds a parallel custom_categories breakdown — first-match-wins, so a minute of busy time is never double-counted across two categories.

Will not: Count a meeting that only exists on an unconnected calendar as a quiet week — the response says plainly that it is not counting it, rather than reporting an empty category; that needs calendar OAuth this deployment does not hold.

Today's tasks, meetings, follow-ups, bookings and drafts awaiting approval, ranked together and computed in the reader's own timezone. Regenerating replaces the day's briefing rather than accumulating a second copy.

Needs: Google/Microsoft calendar OAuth or mail OAuth, for the calendar and email sources specifically

Will not: Read Google Calendar, Microsoft Graph or a mailbox without that OAuth connected — those sources report not_connected by name rather than being silently absent. Compute 'today' on the server's clock — every date is resolved in the reader's own saved timezone.

A public page, real open slots from stored busy time, and availability re-validated at the moment of booking — inside the same transaction as the write — so two racing requesters cannot both take a slot. Notice period, booking horizon and a daily cap all apply on this unauthenticated write path.

Will not: Let two people take the same slot — the write path re-checks busy blocks fresh and refuses with a 409 if it was just taken. Say why a given time is unavailable to a stranger booking the page — only open times are ever shown. Mirror the booking onto Google or Outlook — that needs calendar OAuth.

Overlapping free time for participants whose busy blocks this workspace holds, inside the owner's own working hours and timezone. Every query is now persisted as its own row with its computed result, so re-reading it later answers with what the requester actually saw rather than recomputing against busy time that has since changed.

Will not: Assume an external participant is free — they come back reported as unreadable with a booking-link suggestion, never guessed available. Reading a real external participant's free/busy needs their own explicit share, which no credential of ours can substitute for.

auto-scheduler-engine now runs on every task create and update, not only on demand — placing or replacing the task's own flexible hold with the same free-slot search a manual replan runs workspace-wide. conflict-resolver re-slots any of Mitra's own holds a newly-added fixed event now overlaps, or reports one unplaced if no slot remains before its deadline — only Mitra-owned flexible holds ever move.

Will not: Move an event it does not own — only Mitra's own flexible hold blocks are ever touched; a real meeting or an external invite is never moved. Squeeze a task in past its deadline — a task that will not fit is reported unplaced, with the reason, rather than forced into a slot. Two-way sync with a real Google or Microsoft calendar — only Mitra's own busy-block calendar is written; that needs calendar OAuth this deployment does not hold.

Notes a person typed, the action items they listed, and one-click conversion into tracked tasks that keep a link back to the note all work with no credential. transcript-summarizer and action-item-extractor now run too, over the same typed notes rather than a real transcript: POST .../summarize asks a model for an abstractive summary and candidate action items, held in suggested_action_items separate from the real list until POST .../accept-suggestion promotes one — the same capture-then-confirm shape task-capture uses.

Needs: LLM_API_KEY, for the summarizer and extractor only — typing notes, listing action items and converting one to a task need no credential

Will not: Infer an action item from the prose without asking — only what is explicitly listed in action_items becomes a task automatically; a model-suggested one sits in suggested_action_items until a person accepts it. Join or record a call — there is no transcript path at all; joining as a recording bot needs every participant's consent and a transcript API, deliberately absent.

Deterministic date, time and attendee extraction from free text — day-after-tomorrow, weekday names, explicit times, email addresses — always shown back as a draft for a person to confirm or discard; nothing is ever auto-created. Confirming writes a busy-block row, Mitra's own calendar, and re-slots any of its own holds the new event now overlaps.

Will not: Auto-create a calendar event from parsed text — every extraction is a draft first; confirming it is a separate, explicit step. Send a real Google or Outlook invite — confirming writes a local block only; a real invite needs calendar OAuth this deployment does not hold.

recurring-meeting-resolver generates the next four occurrences on creation, each its own local calendar block. occurrence-conflict-checker nudges a colliding nominal slot inside the rule's flexibility window and marks it skipped, never double-booked, when no room remains. Rescheduling an occurrence whose rule has external participants is refused with a 409 until confirm_external=true is sent explicitly the first time.

Will not: Silently move a client's standing call — an occurrence with external participants needs confirm_external=true stated explicitly before a reschedule is accepted. Send a real invite to any participant — only Mitra's own calendar is written; that needs calendar OAuth this deployment does not hold.

Trackers, reminder windows, snoozing and due listing all work locally. reminder-dispatcher (POST /followups/dispatch-reminders) raises an in-app bell notification — needing no phone number or provider — for every tracker whose window expired. reply-detector (POST /followups/reply-signal) resolves a tracker by thread or message id only, never a subject line.

Will not: Watch a real inbox on its own — there is no Gmail or Microsoft Graph OAuth in this deployment, so reply-signal is a trigger meant to be called by whatever the workspace points at its own mailbox (a forwarding rule, an Apps Script trigger, a Zapier step) with the workspace's own platform credential, not ours. Resolve a tracker by matching a subject line — only a thread or message id counts, to avoid a false match on a common subject.

Scheduling, the opt-in gate and dispatch through Wavy's existing sender all work. Consent is checked against Wavy's own records at schedule time and rechecked at the moment of dispatch.

Needs: a connected WhatsApp channel

Will not: Schedule a reminder to a contact with no recorded opt-in for that category — that is a 403 naming consent_missing, before any draft is even created. Deliver an email reminder at all — Wavy contacts hold a phone number, not an email address, so this is a structural 503, not a missing key away. Record a dry run as sent — an unconnected channel dispatches to status = blocked or failed, never sent.

The whole draft-edit-approve-send machinery is real, including the content digest that invalidates an approval the moment the draft is edited — 'approve a polite note, then send something else' is not a state the schema can hold.

Needs: an LLM key to generate the draft in the owner's voice, and Gmail or Microsoft Graph mail OAuth to actually send it

Will not: Send an email without mail OAuth connected — every draft, approve and discard step works; send always returns a 503 naming the missing credential. Let an approval survive a later edit — changing the subject, body, target address or payload recomputes the content digest and drops the action back to draft, clearing who approved it. Auto-send an email reply, an invite response or a schedule change under any rule — those three kinds are outside the closed auto-sendable set entirely.

The language of a typed request is detected from its script across ten Indian scripts, and the request becomes a capture for confirmation — nothing auto-created. POST /chat/message now also generates a reply in the detected language plus an interpreted_intent_en gloss when a model is configured, so a translated request shows its interpretation back for a quick confirm. PUT /users/{user_id}/language-preference lets an owner set a language default for any user in their workspace, no credential needed.

Needs: LLM_API_KEY, for the reply and interpreted_intent_en gloss only — detection, capture and setting a language preference need none

Will not: Guess between Hindi, Marathi and Nepali from Devanagari script alone — reported as Hindi with a note that the script is shared. Claim Hinglish is distinguishable from English by this method — Latin script is reported as English at 0.5 confidence with that ambiguity named explicitly. Reply in the detected language without an LLM key configured — that field comes back null with reply_status: requires_configuration.

Rule CRUD, the classifier and label-rule-backfill (a newly-created active rule retroactively labels recent triaged mail) all run locally, storing metadata only — no body text retained, archive is opt-in per rule not default. POST /inbox/ingest is the trigger, authenticated with the workspace's own platform credential rather than ours, meant to be called by the workspace's own mail-forwarding automation.

Needs: a Gmail or Outlook mailbox OAuth grant (gmail.modify or Mail.ReadWrite) to apply a label or archive inside a real inbox — rule CRUD, classification and metadata storage need no credential of ours

Will not: Retain a message's body text — only metadata is stored, by design. Write back to a real Gmail or Outlook inbox — applying a label or archiving there needs that mailbox's own OAuth, which this deployment does not hold.

Workspace member. A note, a pasted email body or a transcribed voice memo becomes a reviewable capture. Confirming it — with a real title and due date a person supplies — is what actually creates the task.

{ "raw_content": "call the printer tomorrow about the flex banner", "source": "voice" }

{ "id": "…", "status": "pending_review", "extracted": { "suggested_title": "call the printer tomorrow about the flex banner" }, "task_id": null }

Workspace member. A note, a pasted email body or a transcribed voice memo becomes a reviewable capture. Confirming it — with a real title and due date a person supplies — is what actually creates the task.

{ "title": "Call printer re: flex banner", "due_at": "2026-08-17T12:00:00+05:30", "priority": "normal" }

{ "capture": { "status": "converted" }, "task": { "id": "…", "origin": "capture" } }

Workspace member. Scored from deadline proximity, stated priority, age and whether it is blocked, each component appending a human-readable reason. A manual override lives in its own column and replaces the component reasons with a single one naming who set it and why.

{ "task": { "id": "…", "title": "Send proposal to client", "score": 83,
    "score_reasons": ["Due within a day (+38)", "Priority high (+20)", "Waiting 6 days (+6)"] },
  "explanation": "Send proposal to client — Due within a day (+38); Priority high (+20). Close behind: 'Reply to vendor', 'Update pricing sheet' — override if one of those matters more today.",
  "blocked_count": 1 }

Workspace member. Scored from deadline proximity, stated priority, age and whether it is blocked, each component appending a human-readable reason. A manual override lives in its own column and replaces the component reasons with a single one naming who set it and why.

{ "score": 95, "note": "Client called, this jumped the queue" }

{ …task, score: 95, score_reasons: ["Set to 95 by the owner — Client called, this jumped the queue"] }

Workspace member. A brief becomes a five-stage draft plan (Define, Plan, Do, Review, Launch) whose middle-stage tasks are the owner's own sentences with spread deadlines — sentence-split, not invented — held as a draft until a person accepts it. This deterministic path works with no credential and is the default on POST /projects/generate. use_llm=true instead asks a model to split the brief into discrete work items, and returns a 503 naming the missing key rather than silently falling back if asked for without one configured.

{ "name": "Rebrand website", "brief": "Redesign homepage. Update logo everywhere. Migrate blog.", "weeks": 4 }

{ "status": "draft", "tasks": […], "review_required": true,
  "generated_by": "deterministic scaffold — no language model was involved, so nothing here was invented; review the dates before sharing." }

Workspace member. Plain-English status computed from the workspace's own tasks on every read, never stored — so it cannot go stale against the board it describes.

{ "text": "Rebrand website is active. 3 of 8 tasks are done. 1 overdue, the oldest being 'Update logo' (due 2026-08-15). Next up: 'Migrate blog', due 2026-08-22.",
  "counts": { "total": 8, "done": 3, "open": 5, "overdue": 1, "blocked": 0 },
  "review_before_sharing": true }

Workspace member. A weekly focus goal refilled into open slots as fixed 90-minute blocks, with utilisation reported against it and a one-call override to give a block back. focus-conflict-guard now runs on every new busy block: one that overlaps an active focus block is flagged in focus_conflicts and the block is still created — suggest, don't send, never a silent hide — with the override as the one-call release.

{ "created": [ { "start_at": "…", "end_at": "…" } ], "achieved_minutes": 90, "target_minutes": 300 }

Workspace member. Minutes per category for the owner's own week, compared with the prior week, computed from busy time recorded in this workspace. A workspace can now define its own named categories matching by busy-block kind or title, and the weekly report adds a parallel custom_categories breakdown — first-match-wins, so a minute of busy time is never double-counted across two categories.

{ "categories": { "meeting": { "minutes": 420, "prior_minutes": 360 }, "focus": { "minutes": 180 } },
  "coverage": "Computed from busy time recorded in this workspace. Meetings held only on an unconnected Google or Outlook calendar are not counted." }

Workspace member. Today's tasks, meetings, follow-ups, bookings and drafts awaiting approval, ranked together and computed in the reader's own timezone. Regenerating replaces the day's briefing rather than accumulating a second copy.

{ "local_date": "2026-08-17", "items": [
  { "type": "followup", "title": "Ping ABC Corp re: pricing", "priority_score": 60, "due_at": "…" } ],
  "sources": { "tasks": { "status": "connected", "items": 6 },
               "google_calendar": { "status": "not_connected", "detail": "…needs Google Calendar OAuth…" } },
  "next_delivery_at": "2026-08-18T07:00:00+05:30" }

Unauthenticated. A public page, real open slots from stored busy time, and availability re-validated at the moment of booking — inside the same transaction as the write — so two racing requesters cannot both take a slot. Notice period, booking horizon and a daily cap all apply on this unauthenticated write path.

{ "slug": "vivek-consult", "duration_minutes": 30, "slots": [ { "start_at": "…", "end_at": "…" } ] }

Unauthenticated. A slot taken by someone else in the meantime is a 409: "That slot has just been taken — please pick another."

{ "requester_name": "Amit Shah", "requester_contact": "+91-98xxxxxxx", "start_at": "2026-08-19T11:00:00+05:30" }

{ …booking, status: "confirmed", "confirmation_note": "A confirmation message is drafted for the owner to approve; Mitra does not message a contact without a recorded opt-in." }

Workspace member. Overlapping free time for participants whose busy blocks this workspace holds, inside the owner's own working hours and timezone. Every query is now persisted as its own row with its computed result, so re-reading it later answers with what the requester actually saw rather than recomputing against busy time that has since changed.

{ "duration_minutes": 30, "participant_user_ids": ["u_204"], "external_emails": ["client@external.com"] }

{ "candidate_slots": [ { "start_at": "…", "end_at": "…" } ],
  "participants_unreadable": [ { "email": "client@external.com", "reason": "Their calendar is not connected. Send them a booking link so they can pick a time, or ask them to share availability." } ] }

Workspace member. auto-scheduler-engine now runs on every task create and update, not only on demand — placing or replacing the task's own flexible hold with the same free-slot search a manual replan runs workspace-wide. conflict-resolver re-slots any of Mitra's own holds a newly-added fixed event now overlaps, or reports one unplaced if no slot remains before its deadline — only Mitra-owned flexible holds ever move.

{ "horizon_days": 14, "allow_outside_working_hours": false }

{ "placed": [ { "task_id": "…", "start_at": "…" } ],
  "unplaced": [ { "task_id": "…", "reason": "No free slot inside working hours before the deadline." } ] }

Workspace member. Converting the same action item twice is a 409.

{ "note": { … }, "task": { "id": "…", "origin": "meeting_action_item" } }

Workspace member. Deterministic date, time and attendee extraction from free text — day-after-tomorrow, weekday names, explicit times, email addresses — always shown back as a draft for a person to confirm or discard; nothing is ever auto-created. Confirming writes a busy-block row, Mitra's own calendar, and re-slots any of its own holds the new event now overlaps.

{ "raw_text": "Call with ABC Corp day after tomorrow 3pm, invite priya@abc.com" }

{ "id": "…", "extracted": { "start_at": "…", "attendees": ["priya@abc.com"] }, "status": "pending_review" }

Workspace member. Deterministic date, time and attendee extraction from free text — day-after-tomorrow, weekday names, explicit times, email addresses — always shown back as a draft for a person to confirm or discard; nothing is ever auto-created. Confirming writes a busy-block row, Mitra's own calendar, and re-slots any of its own holds the new event now overlaps.

{ "draft": { "status": "confirmed" }, "busy_block": { "id": "…" } }

Workspace member. recurring-meeting-resolver generates the next four occurrences on creation, each its own local calendar block. occurrence-conflict-checker nudges a colliding nominal slot inside the rule's flexibility window and marks it skipped, never double-booked, when no room remains. Rescheduling an occurrence whose rule has external participants is refused with a 409 until confirm_external=true is sent explicitly the first time.

{ "title": "Weekly 1:1", "cadence": "weekly", "duration_minutes": 30, "external_participants": [] }

{ "id": "…", "occurrences_generated": 4 }

Workspace member. Trackers, reminder windows, snoozing and due listing all work locally. reminder-dispatcher (POST /followups/dispatch-reminders) raises an in-app bell notification — needing no phone number or provider — for every tracker whose window expired. reply-detector (POST /followups/reply-signal) resolves a tracker by thread or message id only, never a subject line.

{ "subject": "Pricing proposal for ABC Corp", "channel": "email", "reminder_window_hours": 72 }

{ "id": "…", "status": "waiting", "remind_at": "…" }

Workspace member. sent_at is never touched by a snooze, so 'waiting eleven days' cannot read as 'waiting one'.

{ "hours": 24 }

{ …follow-up, status: "snoozed", remind_at pushed forward… }

Workspace member. No recorded utility opt-in for this contact is a 403 naming consent_missing.

{ "title": "Appointment tomorrow 4pm", "channel": "whatsapp", "send_at": "2026-08-17T09:00:00+05:30",
  "contact_id": "…", "consent_category": "utility" }

{ …reminder, status: "scheduled", "approval_required": true, "approval_note": "Approve the linked draft before this can be delivered." }

Workspace member. If the draft's live digest no longer matches what you read, this is a 409: "This draft changed since you read it. Re-read it before approving."

{ "content_digest": "9f2c…", "note": "Looks good, send it" }

{ …action, status: "approved", approved_digest: "9f2c…" }

Workspace member. No mail OAuth configured: 503 {"code":"requires_configuration", "message":"…Gmail or Microsoft Graph mail OAuth…"}.

{ …action, status: "sent", sent_authority: "approval", provider_message_id: "…" }

Workspace member. The language of a typed request is detected from its script across ten Indian scripts, and the request becomes a capture for confirmation — nothing auto-created. POST /chat/message now also generates a reply in the detected language plus an interpreted_intent_en gloss when a model is configured, so a translated request shows its interpretation back for a quick confirm. PUT /users/{user_id}/language-preference lets an owner set a language default for any user in their workspace, no credential needed.

{ "text": "kal shaam ko printer wale ko call karna hai" }

{ "detected_language": "en", "confidence": 0.5,
  "note": "Latin script — this could be English or Hinglish; the two are not distinguishable by script alone.",
  "capture": { "status": "pending_review" }, "reply": null, "reply_status": "requires_configuration" }

Workspace member. Rule CRUD, the classifier and label-rule-backfill (a newly-created active rule retroactively labels recent triaged mail) all run locally, storing metadata only — no body text retained, archive is opt-in per rule not default. POST /inbox/ingest is the trigger, authenticated with the workspace's own platform credential rather than ours, meant to be called by the workspace's own mail-forwarding automation.

{ "name": "Vendor invoices", "match": { "from_contains": "billing@" }, "archive": false }

{ "id": "…", "name": "Vendor invoices", "active": true, "backfilled": 3 }

Workspace platform credential (not Mitra's own mailbox OAuth). Meant to be called by the workspace's own mail-forwarding automation — the same no-credential-of-ours shape as follow-up-reminders' reply-signal trigger.

{ "triaged": { "id": "…", "matched_rule": "Vendor invoices", "archived": false } }