Routines

A routine is a task a bot runs on a schedule: "every weekday at 7:30, brief me on tech news", "every Sunday evening, review my week". Ask in plain words and the bot creates it (routine_create), or add one in the Routines tab with a schedule picker (every day, weekdays, weekly, every few hours, or custom cron). You can edit or pause a routine without losing it. Schedules are in your local time. A routine a bot creates runs at most every 15 minutes, its prompt is up to 1,000 characters, and a chat holds at most 25 of them (routines you add yourself aren't limited).

Reminders and one-offs. "Remind me in 2 hours to call mum" or "on Friday at 9, check if the invoice was paid" becomes a routine that runs once, at that time, and is then gone (the Routines tab shows it as Once · Fri 16 Oct 09:00). It can be paused or deleted like any routine.

Follow-ups the bot books itself. When a bot says it'll check back ("I'll look Friday whether they replied"), it books that as a one-off follow-up and says when in its reply. The chip under the reply opens it, and the Routines tab shows it as Follow-up · Fri 16 Oct 09:00; delete it to cancel. Follow-ups count like any routine: toward the 25 a chat holds, and toward the daily caps.

Routines that end. "Every morning until the 30th" or "for two weeks" sets an end (Until in the Routines tab, or the bot sets it). Its last run is the last one on or before that day, and then it's gone. If it couldn't run near the end (Mimir wasn't running, or it was paused), it's removed 15 minutes after the end instead of running late.

Reminders for someone else. "Remind Sam at 6 to take the bins out" goes to Sam's own chat, if Sam is in your household (matched by the name of their chat, first name is enough). The first time for each person you're asked (its own kind of approval, never "always"); after that it's set up straight away. It shows under Sam's chat's Routines. Guests can't set reminders for others. The bot that sets it is the one that sends it, so it has to be one Sam paired with.

Timed off your calendar. "An hour before each meeting with someone outside, brief me on who they are" runs off the calendar you linked in Meetings → Calendar instead of a clock. Where a cron goes, the schedule is calendar:<minutes before>, optionally :meetings (only events with someone else invited) or :outside (only ones with someone from another email domain): calendar:60:outside. The bot can make one when you ask, or type it in the Routines tab under Custom. Each run is told the event (its title, time and who's invited), so pair it with the Meeting prep skill. A meeting that's added, moved or cancelled moves the run with it (the calendar is read every 10 minutes); one added with less notice than the lead time gets no run. Such a routine only lives in a private chat, since your calendar is private. For "the evening before tomorrow's first meeting", use a plain cron (0 20 * * *) and have the routine look at tomorrow's calendar.

Goals #

A goal is one thing you want a bot to see through, over days if need be: "find a plumber who can come this week", "get the insurance refund paid". Ask the bot to keep at it and it sets the chat's goal (goal_set): the goal in your words, what has to be true for it to be done, and a progress line it keeps up to date (goal_update). A chat has one goal at a time. It's pinned at the top of the chat, and the chat's header shows it.

While the goal is on, the follow-ups the bot books in that chat, and tasks it hands to another bot, belong to it (marked For the goal in the Routines tab). Routines you ask for stay yours, whatever happens to the goal. Its card has these buttons, and they work in Mimir itself, whatever the AI is doing:

Pause: its follow-ups don't run, and a task it handed to another bot stops, even mid-run. Its result is never told. Resume picks up where it was, without a rush of missed runs.
Done: it's finished. Its follow-ups are removed. After reading outside content, a bot's done asks you first, and the card says what it removes.
Remove: the goal goes, and so do its follow-ups and handed-over tasks.

Blocked. When the bot needs something from you (which quote to take), it marks the goal blocked, says what it needs, and the goal shows in Needs you until you answer in the chat. Its follow-ups keep running.

Done means done. A bot can only mark a goal done when the chat's run log shows something done towards it since the goal was set, like an email sent or an event added. Saying so isn't enough, and neither is booking another follow-up. Otherwise it's refused, and the bot says what's left. You can always mark it done yourself. A done goal stays done: remove it, or ask for a new one.

Outside content. Working on a goal usually means reading pages and email, so a bot can update its progress after reading outside content without asking you. The goal is marked as such (its card says so): from then on bots read it as outside content, never as instructions from you, and a turn that reads it counts as having read outside content. Setting a new goal after reading outside content still asks you first.

Budget. On top of the daily caps, a goal can have a token budget of its own ("don't spend more than 200k on this"). It counts the goal's own runs since it was set: its follow-ups, the tasks it handed over and their reports, and replies in its chat while it's on. Other routines in the chat don't count. When the budget runs out, the goal pauses and Needs you asks: Go on gives it the same budget again.

Started by something happening #

A routine can start when something comes in instead of at a time. In the Routines tab, pick it under When:

When email arrives: a rule like from:acme.com subject:invoice (from: an address or domain, subject: a word or "a phrase", and any other words the email must contain, in the subject or the text; case doesn't matter). Mimir looks at new mail every 2 minutes itself and only a matching email starts a run, so mail that doesn't match costs nothing. Mail from before the routine was made doesn't count. The run is told each matching email (who, when, the subject and the start of the text). It watches the first mailbox in Settings → Email, and works only in your own private chat.
When called (on mimir server only): the routine gets its own address, https://<your server>/hook/<id>, and a secret (Show its address and secret). A caller sends the secret as Authorization: Bearer <secret>, or signs the body with it as X-Hub-Signature-256: sha256=<HMAC-SHA256 hex> (how GitHub signs, so a GitHub webhook's secret field works as is). A wrong secret, or an unknown address, gets 401; a paused routine 409, and so does a signed body sent a second time (a captured call can't be replayed). The body (up to 256 KB; the run sees the first 4,000 characters) is passed to the run.

What came is someone else's words: the run counts as having read outside content. Several emails or calls that come before the run starts go to one run together (up to 20). In chat, "tell me when the invoice from Acme arrives" makes a mail one (cron is mail:<rule>); it ends after the first match unless you say "every time". Only the first one does the same in the Routines tab. Bots can't make webhooks: the secret is yours to hand out.

The scheduler checks every 30 seconds, so a routine runs within half a minute of its time, as long as Mimir is running. On a laptop, Settings → Keep this computer awake helps; for 24/7, self-host.

Pause everything #

The pause button by Bots in the sidebar pauses all bots; a bot's right-click menu pauses just that one. Either lasts until you resume, for an hour, until tomorrow 8:00 or until Monday 8:00. While paused, nothing runs on its own (routines, the digest, the weekly review, the memory tidy-up), other bots can't hand it tasks, and in Telegram, Slack or Discord it answers "I'm paused" instead of replying (/memory, /routines and /stop still work). The Built-in chat and Run now still work: those are you asking directly.

Its routines show as Paused, not missed. When the pause ends (you resume, or its time comes), each routine moves on to its next time instead of catching up on the runs in between, so there's no burst of runs. A one-off reminder that fell inside the pause still runs, once.

Rehearse before it runs for real #

Rehearse on a routine (or on a skill, in the Skills tab) runs it once for real, except that every step that would change or send something (a message, an email, a memory, a new routine, a click on a site) or would ask you first is only described. You see what it found, the list of what it would have done, and what it would have said. Nothing lands in the chat; Activity keeps the run as a rehearsal, and it doesn't count as the routine's run. Reading (web pages, your email, memory) does happen, so a rehearsal uses tokens like a run.

Is it working, and what does it cost? #

Each routine in the Routines tab shows its last runs as dots (green ran, red failed, hollow missed; hover one for when and how many tokens) and its last 30 days: how many runs, about how many tokens each, the API-equivalent cost where the AI reports it, and how many failed or were missed. A routine that quietly uses a lot, or keeps failing, shows on the routine itself.

A routine missed its time when it ran more than 15 minutes late: Mimir wasn't running then (the app was closed), the computer was asleep, or a long run before it held it up. It still runs once as soon as it can (not once for every slot it missed), and the run log gets a line saying which times it missed and why ("Missed 3 runs, 15:00 to 17:00: Mimir wasn't running then. It ran at 17:17 instead."). Missed and failed runs show in Needs you. The late run is told when it was due, so a 7:30 briefing that runs at 13:10 leaves out what has already passed instead of saying "this morning".

Caps #

Settings → Limits sets daily caps, in tokens: for each routine, for each bot, and for all bots together (none by default). A routine whose run would go over one pauses until midnight instead of running (one routine for its own cap; all of a bot's routines for the bot's or the day's), says so where it speaks, and Needs you asks: Let it go on today lifts that cap for the rest of the day. Otherwise the routine runs once after midnight, and the run log says which times it skipped and why. Replies to your own messages are never stopped by a cap, though they count towards the bot's and the day's.

Tokens read back from the AI's cache (Claude Code and Codex send the bot's instructions that way) count a tenth, as they're billed, so a cap isn't eaten by the same instructions read again; each routine shows what it uses on average on that scale, which is a good guide for its cap. An AI that doesn't report what it used (some local models) gets an estimate from the text, about four characters a token, shown with ≈ in Activity, so caps still apply to it. Where the AI reports them (Claude Code does), Limits also shows your plan's own windows: how much of the 5-hour and weekly windows is used, and when each resets.

Model for routines (also in Limits) runs every routine on a lighter model of the same account, e.g. haiku on Claude or gpt-5-mini on ChatGPT, while replies to you keep the bot's own model. Routines are usually the bulk of what a plan's windows go to, and most don't need the biggest model. A routine's Test run uses it too, so you can see whether it's good enough first. With bots on different kinds of account, there's one field for each kind.

When a run doesn't work #

What happens next depends on why:

The AI plan's usage limit. The routines of every bot on that AI login (your Claude, ChatGPT or Grok sign-in is one for all your bots; bots on an API key only themselves) wait until the limit resets (the time Claude Code reports; an hour when the AI doesn't say), then each runs once, and the one that hit the limit runs again. You get one message where that routine speaks, and Needs you shows "routines wait until 15:00" with Run them now (if you raised the limit or switched account).

A fallback account (the bot's page, When it hits a usage limit; off by default) runs the bot's routines on another account instead: a different login (ChatGPT, Codex, Grok, Claude) or a key (OpenRouter, xAI, an API key, which bill per token, so that costs money). The routine that hit the limit runs again there on the next pass, the others go on as usual, and at the reset time they're back on the bot's own account. Each run on the fallback says so in the run log ("on the fallback account until 15:00"), and Needs you doesn't list the bot as waiting. Daily caps still apply. If the fallback hits its own limit, its routines wait for the bot's own account after all. Replies to your messages stay on the bot's own account.
Couldn't reach the AI (the network, a local model that's starting): it tries again after 5, 15 and 30 minutes, without messaging you, and only then says it failed.
The AI account isn't set up (not installed, signed out): Mimir checks this before each run, so nothing is spent; it retries like a network error, so signing in again is enough.
Anything else: it says so where the routine speaks, as before.

A routine that keeps failing the same way is one item in Needs you ("4 times"), and once it works again its failures leave Needs you by themselves. A one-off (a reminder) that fails this way is kept and tried again, not lost.

Activity shows where each result went: sent to telegram, in the digest, on its card, only in the log, in the app. When the messaging app didn't take it (tried twice), that's an item in Needs you, and the result is still in Activity.

Every run starts a fresh conversation with the AI, so a routine that runs every 15 minutes doesn't grow its context (or its cost) run after run.

Quiet checks #

A routine can be quiet: it runs with read-only tools and messages you only when something needs your attention; otherwise it stays silent (the run log still shows it). "Keep an eye on X and only tell me if it changes" makes one. The Routines tab also offers ready-made ones: a morning task check, an evening look-ahead, watching a page.

A quiet check about one page, feed or your inbox can watch it (the page's address, or mail for new unread email; bots set it when you ask them to keep an eye on something). Mimir then looks itself before each run: when nothing changed, the AI doesn't run at all (the run log shows "nothing changed, no AI run", and the routine counts these as free). When something did change, the AI runs with what changed: the lines that are new or gone (for mail, the new unread messages), up to about 3,000 characters. So a check every 15 minutes costs a fetch, not 96 AI runs a day. The first look runs the AI as usual, since there's nothing to compare with yet. If Mimir can't reach the page, the check runs as it would without a watch.

Pages say things that change on their own. Times of day, timestamps and "5 minutes ago" are never counted. When a page did change, Mimir looks again a few seconds later: lines that differ between those two looks (a visitor counter, a session id, a rotating ad) are the page's own churn, and lines like them are left out from then on. The routine shows "ignoring 2 lines that change on every load" (hover for which), with Reset to notice them again. A line that only moved isn't a change.

When routines speak #

Each routine is sent:

right away: a message and a notification;
in your daily digest: one message a day at the time you set (Settings → Daily digest), with a line like "4 quiet checks ran; nothing needed you";
on a card: nothing is sent; the result replaces a card pinned at the top of the chat in the app (bills due, this week's to-buy list). A first line # Title names the card. Each run is shown the card as it is now, including items you ticked, and told to keep your ticks unless an item is done with. Remove the card and the next run makes a new one. See Pinned cards.
just logged: nothing is sent; Activity shows what it found.

Every run's reply is kept in Activity.

Come through right away if… A routine that would wait (in the digest, just logged, or in quiet hours) can have one line of your own, like "my flight is delayed or cancelled". The AI is told your rule and starts its reply with URGENT only when it holds; that result is then sent right away, starting with "🔔 Came through because: my flight is delayed or cancelled", instead of waiting. At most 3 results a day come through this way, all bots together; past that, the result waits as usual and says it matched. You set the rule in the Routines tab (Edit); a bot can't set or change it, so nothing it reads on a web page can widen it. A page can still try to make the AI think your rule holds, which is what the daily cap is for. Rehearsals and Run now don't use it.

You can reply to what a routine sent ("skip the crypto stuff"), and the same goes for the digest and the weekly review. A new routine's first three results end by asking "Anything here you didn't need, or something missing?": reply, and the bot rewrites the routine's brief to match (routine_update), keeping what you didn't ask to change. You can ask for changes any time after that too ("move the news to 9", "make it shorter").

In Telegram, Slack and Discord a result also comes with buttons, so you can tap instead of typing. Mimir handles taps itself, with no AI turn:

A one-off (a reminder): Done, Snooze 1h (it comes again in an hour) or Tomorrow (same time tomorrow).
A routine: 👍, 👎, Less like this, Pause (pauses the routine), For tomorrow… (asks what to change; your reply goes to the bot like any other reply).

👍, 👎 and Less like this never edit the routine. They are evidence for the memory tidy-up, which proposes changes for you to accept, the same as when you correct a reply (Memory). Each message's buttons work once, and in a group only the person who paired the chat can press them. Results that go into the digest, and failures, have no buttons. A routine runs in its own session, so on Claude, Codex and Grok the bot's next turn gets that last message with your reply. If the routine had read a web page, that turn counts as having read it too; so does a reply to the digest or the weekly review, which are put together from routine results and learned memories.

Quiet hours #

Set quiet hours under Settings → Daily digest (e.g. 22:00–07:00). They hold everything proactive: routines that would message right away go into the digest, the digest and the weekly review wait until quiet hours end, and approval cards from routines wait until morning. Replies to your own messages still come, and so does a result that matches its routine's "come through right away if" rule (above).

Needs you #

The first item in the sidebar collects what's waiting on you: approvals, runs that didn't work (in plain words, with a fix button: "Couldn't reach the AI. Is it running?" → Check the AI account), bots that can't connect (a rejected token, an AI account that's signed out; the bot's chat says so too, with a way to its settings), approvals nobody answered in time, memory suggestions, and goals that are over their budget or blocked on you.

Routines run with nobody watching, so an approval a routine needs is parked there for 24 hours instead of being denied after 5 minutes; when you approve, that exact action runs and the chat says so. The menu-bar / tray icon shows the count, and its menu lists what's waiting with Allow and Deny, so you can answer without opening the window.

One chat for decisions #

A parked approval's card goes to the chat its routine belongs to, so one made in the app chat waits until you open the app. Settings → Decisions picks one of your own private chats in Telegram, Slack or Discord (never a group or a guest's chat) that also gets them: a parked approval from any other chat comes there as a card saying where it's from ("From: This computer"), a routine that failed for good says so, and new memory suggestions are mentioned (accept or dismiss them in Needs you). Answer the card in either place and the other one shows the outcome, without buttons. Quiet hours hold these too. Only the bot that asked sends the card, using its own app connection, so that bot needs a connection to the same app as the chat.