Email is five apps at once
Your inbox mixes five kinds of mail that ask for five different things. A person is waiting on a reply. A bill has a deadline. A build notification wants a glance. A newsletter you chose is there to read when you like. A receipt matters a year from now, when you need its reference number. One list of subject lines treats all five the same and offers the same actions for each: reply, archive, label.
mxr treats email as five modes that share one local mailbox and runtime. Each has its own view, primary object and completion action. The web app, TUI, CLI and agent interfaces use the same daemon-backed state, with different client surfaces. Now brings together a short list of work across the modes; it is a starting view, not a sixth mode.
mxr now # people waiting on you and tasks due soonmxr messages # people and groups, with their topicsmxr todo # tasks ordered by when to actmxr updates # notifications grouped into a briefingmxr reading # newsletters as an editionmxr records # receipts, orders, bookings and billsmxr modes explain # what each shipped mode does and its keysEach mode does one job, with its own verb
Section titled “Each mode does one job, with its own verb”| Mode | What it holds | What you do there | Completion |
|---|---|---|---|
| Messages | People you are in conversation with | Read the request, reply, acknowledge or mark its topic done | Done clears that topic in Messages; a later message can bring it back |
| To do | Dated actions with a source email | Check the evidence, act, schedule or complete one task | Completing a row ticks that task; other tasks remain open |
| Updates | Source-grouped notifications | Glance, open the source email, make a task when needed, then let go | A preview selects the digest messages to mark done in Updates |
| Reading | Newsletters and posts you chose | Read, save to Later, unsubscribe or let go | Let go or expiry clears an item from Reading; Later stays saved |
| Archive | Sourced records: receipts, orders, bookings, invoices | Retrieve a field and follow it to its source email or document | Records stay filed; Archive has no done action |
Inbox stays as the everything view, in arrival order. It is not a sixth mode. The mode answers what an email is for; Inbox shows what arrived.
Each view keeps its own object and action. Use Now to choose a person in Messages, a task in To do, the Updates briefing, or a reading item. Use Archive when you need a filed record and its source.
A conversation can belong to several modes
Section titled “A conversation can belong to several modes”A conversation can contain a reply to write, an open task and a filed record. Each mode tracks its own work. Finishing in Messages leaves an open to-do alone. Archive keeps filed records without holding their email in the inbox.
Inspect membership and preview what finishing in one mode would do:
mxr modes why MESSAGE_ID --format jsonmxr modes done THREAD_ID --mode messages --dry-runWhen the last inbox-holding mode lets go, mxr may archive the conversation at
the provider. This is enabled by default through
modes.archive_on_last_done; the preview shows the provider effect before
you apply mxr modes done THREAD_ID --mode messages. Archive records do not
hold an email in the inbox and filing a record does not perform this action.
An inbox removal observed by sync is evidence, not a mode decision. It does not complete a task or mark Updates or Reading done. Those modes retain eligible items; their own let-go, snooze and expiry rules still apply. Mail first discovered already archived is not added to Updates or Reading just because it is outside the inbox.
Updates and Reading label retained mail “left the inbox · observed
mxr updates --format jsonmxr reading edition --peek --format jsonmxr todo --format jsonSorting rules and categories serve different jobs
Section titled “Sorting rules and categories serve different jobs”Each inbound email has one stored home and deciding reason. Messages, Updates, Reading, the Inbox chip and arrivals read that answer. Your moves and sender choices take precedence over teaching rules and built-in rules; Spam, Trash and blocked senders stay outside those homes.
When no rule decides, possible personal mail goes to Messages and machine
mail goes to Updates, marked Not sure. A missing body remains pending.
Open mxr why MESSAGE_ID --format json for the stored assessment and evidence,
or mxr classify report --format json for bounded fallback counts and clusters.
The stored-assessment details require changes made after
v0.6.158; that release does not include them.
While sync changes the evidence, an email can briefly be still sorting. It
stays in Inbox and arrivals while its mode waits for a stored answer. If
mxr why returns classification_pending, retry once sorting finishes.
A home describes the email’s purpose. Inbox membership, snoozes, done marks and expiry still decide whether it appears as open work. Inspect a message’s placement before correcting it:
mxr why MESSAGE_ID --format jsonmxr reclassify MESSAGE_ID reading --dry-runmxr reclassify MESSAGE_ID readingmxr arrivals list --mode reading --format jsonmxr corrections --format jsonUse --sender with Messages, Updates or Reading when the choice should apply
to that sender’s mail. Rules can automate placement with a
preview of their effect.
Personal categories add your own groupings, such as Travel or Household, across modes and accounts. They keep provider labels and folders unchanged. Category rules add or remove membership; manual inclusions and exclusions win over those rules.
mxr categories --format jsonmxr categories --format json why MESSAGE_ID CATEGORY_IDMeasure how well mxr sorts
Section titled “Measure how well mxr sorts”mxr classify eval grades the mail already in your mailbox against labelled
scenarios and prints how often today’s rules agree with the labels. It reads
stored mail and changes nothing. Only the demo mailbox carries labels. mxr demo
runs in its own profile and leaves your config alone, and every mxr command
uses it until mxr demo stop:
mxr demo --no-tuimxr classify eval # the quick set, in secondsmxr classify eval --set full # every labelled email, up to a minutemxr demo stop # back to your real profileOn a profile without labelled mail it prints No labelled mail here; run mxr demo. In a terminal the output is a table; piped, it is JSON. --format table
forces the table. This is what the quick set printed on v0.6.150:
Classifier eval, quick set (label schema v1)
question evaluated correct accuracywho 99 89 89.9%home 99 89 89.9%todo 99 98 99.0%record 99 95 96.0%
99 of 147 labelled emails are in this mailbox; 49901 stored emails have no label.
25 mismatchessubscriptions-disney-receipt demo-msg-103 who expected machine got person (possible personal mail; no decisive list or automation evidence)Each email gets four questions. who asks whether a person, a list, a
machine or a screened-out sender wrote it. home asks which mode it lands in.
todo and record ask whether a to-do or an Archive record was made from
it. A mismatch names the scenario, the demo message id, the question, the
expected and actual answers, and the rule that decided. [label is a judgement call] marks a label that may be the label’s fault rather than the classifier’s.
Labels say what mxr should do, so a gap shows up as a mismatch.
| Set | Grades | Use it |
|---|---|---|
--set quick (default) |
The 147 labelled emails of the hand-written demo mail and its near-miss twins; 99 of them are in the demo mailbox | After you change how mail is sorted |
--set full |
Every labelled email in the mailbox: on the 50,000-message demo, 38,747 emails across both accounts | Before you rely on a change; CI runs it |
The full set groups its mismatches into clusters, biggest first, with one
example each, and prints a second accuracy table for held-out phrasings. Those
are wordings kept out of the mail used while tuning the classifier. If a held-out
figure falls well behind the overall one, a change fitted the generator’s
wording rather than the mail. --format json lists every mismatch,
--format csv writes one row per mismatch, and --account limits the run to
one account.
Model work will run in two tiers (planned)
Section titled “Model work will run in two tiers (planned)”Today every model feature uses the one model you set under [llm], or a
per-feature override. Model features are off until you turn them on, and
the default endpoint is a model server on your own machine.
LLM features lists what runs today.
The modes plan splits model work by how hard it is:
- A fast tier does the bulk work that touches every incoming message: deciding which mode it belongs in. It runs on your local model by default.
- A smart tier pulls out the fields that need care: the amount and deadline of a bill, the fields of a receipt, what a person asked you. It only sees mail already placed in To do, Archive or Messages. It can use a cloud model you set up with your own API key, or stay local.
- When the smart model can’t find an answer that code can check, the task can step up to a stronger model you name, instead of every call paying for the strongest one.
The current LLM configuration already supports a local endpoint or a cloud provider with your own key. The proposed tier routing is separate from that choice. Amounts and dates a model pulls out must appear word for word in the email, and code checks that. Many tasks won’t use a model at all: where code can do the job, such as quote stripping or ranking people, code does it.
Check whether model features are on, and which model and endpoint the daemon uses:
mxr llm status --format jsonEach mode will search the part of the email it cares about (planned)
Section titled “Each mode will search the part of the email it cares about (planned)”Today semantic search uses header chunks and overlapping windows of the cleaned body and attachment text, plus saved highlights. Archive also adds a chunk of record fields, with PDF text when the attachment is cached. Embedding runs on your machine. Semantic search covers how to run it.
The plan gives each mode its own recipe. Messages indexes each message’s new text without the quoted history. To do indexes one line per task. Updates indexes one fact per message and folds repeats. Reading embeds articles only when you read them. Archive indexes record fields and leaves reference numbers to exact keyword search. Archive’s recipe ships: a record’s emails get one field chunk beside the header, plus their PDF text. The other recipes replace today’s chunking only if a retrieval test on real mail shows they find things at least as well.
See what today’s index holds:
mxr semantic statusWhat you can use today, mode by mode
Section titled “What you can use today, mode by mode”| Mode | Status today | Shipped features that already serve it | Planned |
|---|---|---|---|
| Now | Built | Now (mxr now, g h): People, Due soon, the Updates card and an evening Reading pick, at most ten things; first-run progress, arrivals (mxr arrivals), corrections (mxr corrections), the rail (mxr modes rail), done here per mode (mxr modes done) and “Also in” (mxr modes why) |
Background classification by your model and a model-history horizon |
| Messages | Built (g m) |
Messages (mxr messages): people as rows with their topics inside, Your turn by closeness and pace, Got it (mxr messages ack), thread-aware quote stripping, manual merges with a suggestion; Focus & reply |
Incoming Gmail reactions as a mark on the message; the Messages index recipe once the retrieval eval favours it |
| To do | Built (g x) |
To do (mxr todo): bills, renewals, forms and promises with act-by dates; add and edit tasks, review the first-run catch-up (mxr todo catchup), review later evidence, and optionally link Microsoft tasks; calendar invites (mxr invites) |
An accuracy-gated rail badge; the planned mxr modes eval command is not available yet |
| Updates | Built (g u) |
Updates (mxr updates): one line per source at the 08:00 and 16:30 cuts, code-computed changes, parcels, builds and incidents as trackers, let go in one key (mxr updates let-go); a new sign-in, failed payment or delivery problem leads Needs a look as a suggested to-do that t adds, and explains why automatic creation was held when receiver evidence is unavailable; deliveries (mxr deliveries) |
One fact per message in the semantic index, and model-written facts for generic subjects (phase 7) |
| Reading | Built (g r) |
Reading (mxr reading): an edition banded since your last visit, earlier and fading, a 66-character reader with the linked article on request, a Later shelf, let go with e and E, unsubscribe with evidence, highlights you can edit and remove, and Marginalia on O |
Section-aware search chunks and lazy embedding for whole issues |
| Archive | Built (g e) |
Archive (mxr records): receipts, orders, bookings, bills and documents as records, an answer box with mxr ask fallback, CSV/PDF export, recurring subscriptions (mxr records subscriptions), renewal reviews, coming-up moments and record-field search chunks |
Record fields read by the smart tier, and learned templates per issuer |
Classification by your own model is phase 7. Phase numbers come from the plan and can change; this page changes when each phase ships.
Deleting an email clears its derived data
Section titled “Deleting an email clears its derived data”When mxr observes a provider deletion, it removes the local message and its per-message data, invalidates affected summaries and briefings, and removes extracted facts that have no remaining evidence. Managed cache files and in-memory index entries are cleaned through a durable queue. File cleanup has bounded retries; an in-memory index failure waits for its next rebuild or daemon restart. Failures are reported in daemon logs. Security & Privacy explains the timing and the user-authored state and no-text receipts kept deliberately. Moving a message to Trash is not a provider deletion.
Find a mode’s explanation again
Section titled “Find a mode’s explanation again”mxr modes explain todo --format jsonmxr modes hint todo.runway --showmodes explain lists the mode’s purpose, actions, keys and hint IDs.
modes hint ID --show makes a dismissed hint eligible to appear again;
without --show, it dismisses that hint across clients. Use the mode’s
? help in the app when you want the explanation beside your mail.
Read the plan
Section titled “Read the plan”- 22. Email is five apps: the design, the data each mode needs, and the phases in order
- Decision log, D107 to D114: the settled choices behind the modes
- Roadmap: which phases are done
See also
Section titled “See also”- Work through your day: today’s workflow, start to finish
- Clear the desk
- Clear Reading and Paper trail
- Archive
- Glossary