Skip to content

Browse your activity log

mxr can record selected successful user-intent requests from supported clients. The activity log is local, can be disabled or paused, and is not a record of every action. Browse a time range, filter by action or client, and redact rows you no longer want to keep.

The log is a local working history, not a compliance audit. By default, subject, recipient email, and search-query context can be recorded; link URL tracking is off. Pause drops ordinary new rows, some rows coalesce, and the write queue can drop rows when full. Change the metadata controls in [activity] or set MXR_ACTIVITY=off to disable writes.

State-changing or intent-expressing actions:

  • Mail: mail.read, mail.archive, mail.trash, mail.star, mail.label, mail.move, mail.snooze, mail.unsubscribe, mail.mark_spam.
  • Send: mail.send, mail.reply, mail.forward, draft.create, draft.assist (the action, model and how much came back, never the text), draft.discard, scheduled sends.
  • Search: search.run, search.save, search.delete, saved.open.
  • Threads: thread.open, thread.summarize, thread.flag_reply_later.
  • Accounts: account.add, account.remove, account.signin, account.sync.
  • Rules: rule.create, rule.update, rule.delete, rule.run, rule.test.
  • Snippets / signatures / screener / reminders: edits and triage decisions.
  • Microsoft task links: todo.microsoft_link, todo.microsoft_sync for confirmed link and manual sync requests. These record only the task ID, not its title, notes or email content. Previews and automatic reconciliation are not captured.

Mapping is closed: every IPC verb that produces activity has an explicit entry in crates/daemon/src/activity/mapper.rs. New IPC verbs default to “no activity” until someone decides what to log.

Terminal window
# Show the last day of activity in a table
mxr activity list --since 24h

What you get: reverse-chronological table with TIME / SRC / ACTION / TARGET / CONTEXT columns. Source badges (tui, cli, web, daemon, agent, mcp, assistant and the rest) make it obvious where each action originated. Changes the assistant made after you approved them show as assistant; filter with --source assistant.

  • Cursor moves, scroll, pane focus, palette opens-then-cancelled.
  • Heartbeats, pings, status polls, background sync, FTS rebuilds.
  • Read-side getters (mxr cat, listing envelopes) — they don’t change state.
  • New activity when [activity].enabled = false or MXR_ACTIVITY=off is set at daemon startup. Pause drops ordinary activity but keeps pause/resume markers.

Shared filter flags apply to list, stats, top, export, and redact. AND-combined; empty means any.

Terminal window
# Last week of mail mutations, JSON for piping
mxr activity list --since 7d --prefix mail. --limit 100 --format json
# Searches you ran for "invoice"
mxr activity list --action search.run --query 'invoice'
# Activity on a specific thread
mxr activity list --target-kind thread --target-id thr_abc123

What you get: each command emits the matching rows; --format json produces { entries: [...], next_cursor: { ts, id } | null } for paging.

Aggregate over a window to see what you actually did.

Terminal window
# Top actions over the last week
mxr activity top --since 7d --limit 10
# Hour-of-day histogram for the last month
mxr activity stats --group-by hour --since 30d --format json | jq
# Daily rollup
mxr activity stats --group-by day --since 30d

What you get: KEY / COUNT table in human mode, { buckets: [...] } in JSON mode.

replay aggregates consecutive same-action rows into readable lines.

Terminal window
mxr activity replay --since 1h

What you get:

Since 1h:
09:42 Read 5 threads
09:43 Searched "invoice 2026"
09:44 Archived 12 threads
09:46 Sent 1 message

recall resolves a curated set of fuzzy-time phrases.

Terminal window
mxr activity recall "yesterday afternoon"
mxr activity recall "last hour"
mxr activity recall "before lunch" --limit 20

Accepted phrases: today, yesterday, this morning/afternoon/evening, lunch, breakfast, night, last <duration>, past <duration>, before|after|since|until <phrase>. Anything else returns an error pointing at the supported grammar.

Terminal window
# CSV → stdout
mxr activity export --format csv > today.csv
# NDJSON → file (preferred for piping into jq/awk)
mxr activity export --format ndjson --out my-week.ndjson
# Filter first, export only the slice
mxr activity export --prefix mail. --since 7d --format ndjson --out mail-week.ndjson

What you get: the daemon writes the file with the matching rows in the requested format. CSV is RFC 4180. NDJSON is one ActivityEntry per line. JSON is a top-level pretty-printed array.

Choose what to record, pause temporarily, or scrub existing rows.

Edit the file printed by mxr config path:

[activity]
enabled = true
track_subjects = false
track_recipient_handles = false
track_search_queries = false
track_link_clicks = false

Restart the daemon after changing these controls. This example keeps actions, counts and opaque IDs without subjects, email handles, search text or URL metadata. Sender-address targets are omitted too; folder destinations and rule names remain. Set enabled = false instead to stop all new activity rows, including administrative markers.

These controls affect new rows only. To remove old context, use the dry-run and tombstone commands below. Disabling or pausing recording also stops Reading engagement, but you can still save Later items and highlights. Link-click tracking is off by default; no current producer emits link.click rows, so enabling it does not add browser tracking.

Terminal window
MXR_ACTIVITY=off mxr daemon

The kill switch overrides config for the lifetime of the daemon, including forced pause/resume markers. Set it when starting the daemon, not just on a client command talking to an already-running daemon.

Terminal window
# Stop recording for 2 hours; auto-resumes
mxr activity pause --for 2h
# Indefinite — stays paused until `resume`
mxr activity pause
# Bring it back
mxr activity resume

Pause writes one activity.paused marker, then drops new writes until the window elapses or you resume. Auto-resume also writes a synthesized activity.resumed marker so you can always see when the gap began and ended.

Always dry-run first. Tombstones are irreversible (the audit-trail columns survive but context_json is cleared).

Terminal window
# Preview what would be tombstoned
mxr activity clear --last 1h --dry-run
# Tombstone the last hour (preserves important-tier rows by default)
mxr activity clear --last 1h --yes
# Nuke everything, including important rows
mxr activity clear --last all --include-important --yes
# Surgical: tombstone two specific rows
mxr activity redact --ids 42,43 --yes
# Retention prune (hard delete, not tombstone)
mxr activity prune --before 90d --dry-run
mxr activity prune --before 90d --yes

Saved activity filters mirror saved searches.

Terminal window
# Save the current filter under a slug
mxr activity saved save mail-week --name "Mail this week" --since 7d --prefix mail.
# Run a saved filter
mxr activity saved open mail-week
# Manage
mxr activity saved list
mxr activity saved delete mail-week
Terminal window
mxr activity replay --since 30m

Quick prose narrative of the last half-hour. Catches you up after a context switch.

Which threads did I archive without reading?

Section titled “Which threads did I archive without reading?”
Terminal window
mxr activity list --since 7d --action mail.archive --format json \
| jq -r '.entries[] | select(.context.read_then_archive != true) | .target_id'

What you get: thread IDs you archived without the read_then_archive flag — candidates for unarchive if you went too fast.

Terminal window
mxr activity list --since 30d --action search.run --format json \
| jq -r '.entries[].context.query' \
| sort | uniq -c | sort -nr | head -20

What you get: top-20 queries you ran in the last 30 days, ranked.

Audit what mxr did the last time it crashed

Section titled “Audit what mxr did the last time it crashed”
Terminal window
mxr activity list --since 24h --include-redacted --format json \
| jq -r '.entries[] | select(.action | startswith("activity.")) | "\(.ts)\t\(.action)\t\(.context)"'

What you get: synthesized markers (activity.paused, .resumed, .pruned, .redacted, .exported, .cleared) — the daemon’s diary about itself.

"Use `mxr activity list --since 24h --format json` to summarise what I did
in the last day, then suggest one filter I should save with
`mxr activity saved save`. Don't run any mutating commands. If you find
sensitive context, propose a `mxr activity redact --ids ... --dry-run`
without executing it."
"Run `mxr activity replay --since 4h` and paste the output. Then run
`mxr activity stats --group-by hour --since 7d` and tell me my peak
inbox-processing hour."
  • CLI: mxr activity ... (alias mxr act). Scripts use --format json; mxr activity export also takes --format ndjson.
  • TUI: press g a from any screen to open the activity modal. j/k to navigate, p to toggle pause, Esc to close. Also visible inside the Diagnostics page (G d) as the Activity pane.
  • Web: navigate to /activity in the web app, or open the Diagnostics page and switch to the Activity tab. Filter sidebar, bulk-select with checkboxes, pause/resume from the top bar.

Three tiers with separate retention windows. Configure in the file printed by mxr config path:

[activity]
enabled = true
track_link_clicks = false # opt-in; URLs reveal a lot
track_subjects = true
track_recipient_handles = true
track_search_queries = true
[activity.retention]
ephemeral_days = 30 # views, palette opens, navigation
standard_days = 90 # searches, thread opens, attachment views
important_days = 365 # mail mutations, sends, account changes

A daily prune sweep hard-deletes rows older than the cutoffs. Each deletion writes one activity.pruned marker so you can audit what was removed.

Activity producers omit credential material and mail content:

  • OAuth/refresh tokens, password hashes, API keys, session ids.
  • Attachment bytes (filenames + sizes only).
  • Mail bodies and draft text. Send activity keeps content kind and counts, not body excerpts.

A PII-audit test asserts the table never contains the forbidden keys after a full mapper run.

  • No encryption at rest. Relies on OS-level disk encryption (FileVault, LUKS).
  • No cross-device sync. By design.
  • The TUI modal is read-only in v1 (browse + pause). Redaction and export currently flow through the CLI; the web surface has bulk-redact via the data table.