Mimir v0.5.0

Talking to your bots

• WhatsApp: a bot on a server can be on WhatsApp, with a business number of its own,
  through Meta's official Cloud API. Text, photos, voice notes and files both ways, typing, pairing, and approval cards
  and questions as buttons. It can only write within 24 hours of your last message there: a routine's result after
  that isn't delivered, and shows as not sent in its run. Setup is under Connect.
• Grok without Grok Build: Sign in with your Grok plan (Settings → AI keys) runs bots on the xAI API account
  on your X Subscription, SuperGrok or X Premium+ plan, with no CLI and no API key. Confirm a code at x.ai from any
  browser, so it works on a server too. Bots use Mimir's tools directly and get photos at full size. If xAI doesn't
  offer its API to your plan, the bot says so; the Grok (Grok Build) account still works. See
  docs/ai-accounts.md.
• X search: bots can search recent X (Twitter) posts (x_search) with your own X API bearer token
  (Settings → Web search → X). Good for a morning routine of fresh posts. See
  docs/connect.md for getting a token.
• Telegram guest mode: type @yourbot in any chat, even one it isn't in, and it replies there. Only to you, and
  as in a group: none of your email, calendar, meetings, computer or saved logins, and the message you replied to is
  outside content. A reminder it sets comes to your private chat.
• Branch from here: a new conversation that picks up from a message, with a short "where we left off", to try
  another direction without re-explaining. The branch icon on a reply in the app, or /branch as a reply in your
  messaging app; with Telegram topics on it opens a new topic. See Bots and chats.
• Chat titles and a catch-up line: after a chat's first exchange, and again every ten messages, the model in
  Model for routines names it in 3–6 words and writes a line on where it's got to. The title shows under the bot
  in the sidebar; the line shows at the top of the chat when you come back to it. Not for incognito chats; Start
  fresh and deleting a chat's history remove them.
• A checklist for long runs: on a task with several steps (research, a booking, a form) a bot keeps a checklist
  (progress_set) that it ticks as it goes. It shows above the reply being written, in the app and in the Telegram,
  Slack and Discord draft, and goes when the reply arrives. A reply that ends with steps still open gets one more turn
  to finish them or say what blocks them, so it doesn't stop at "I'll now do X".
• Tool servers' interactive cards: when a tool server answers with a card (MCP Apps: Figma, Notion, Canva…), the app
  and the web UI show it under the reply, in a sandboxed frame. A button in it that changes something asks first, like
  any other write. Chat apps keep the text answer. See Tool servers.
• Webhooks out: Settings → Webhooks out posts a small signed JSON to an address of yours when a routine
  finishes or fails, an approval is waiting, or a reply was sent. You can use it with Home Assistant, ntfy or n8n.
  Message text is only included if you ask for it. A failed send is tried again. See
  docs/routines.md.
• Scanned PDFs read on your computer: set a local document model in Settings → Scanned PDFs (an
  Ollama model such as dots.ocr, or any vision model) and a scan's pages reach the bot as text,
  tables included, instead of photos: less of your plan per page, and the scan stays here. Off until you set one.
• PowerPoint files: a .pptx someone sends, or you drop into the app's chat, is read like a Word or Excel file:
  each slide's text and speaker notes, hidden slides marked.
• At the usage limit, answer later: when a reply hits your AI plan's limit, the message says when the plan resets
  and offers Answer at 15:00, which reruns your message then. With a fallback account there's also Answer now
  with that account.
• Hand a task to a bot's own chat: "Hatoba Marketer, write and show your marketing plan" now has that bot do it in
  its own chat: the task and its whole answer stay in its history, files it writes go to its own folder, and the bot you
  asked replies with a one-line confirmation and a link. Works from the app and your own private messaging chats.
• Approval cards say which bot acts: a card from a bot working on another bot's task reads "Scout, on a task from
  Researcher: …".
• Asks from routines show at the top of the app, with a link to Needs you.
• Allow all covers every waiting card (not signing in or taking over), with a confirmation listing each one.
• Always on every card: sending an email, making a bot, a saved login, a plan, a repeating step, and asks after a web
  page or a sensitive source now offer it too. On those it means never ask, not even after outside content: the app
  confirms first for those guards, Settings → Permissions shows such tools as Always, and the health check lists them.

Routines

• Quiet checks can watch a page's data: pages that build themselves in the browser (booking calendars, prices, stock)
  gave a quiet check nothing to compare. While you take over the bots' computer or show it something, Mimir notes the
  JSON the page loads, and the routine form offers Watch this instead: no browser, and no AI until a value
  changes. The check uses the bot's own cookies for that site, kept in the secret store and never shown to the AI.

Memory

• Suggests only what keeps coming up: the tidy-up no longer proposes a memory after one conversation unless you said
  it yourself; it waits until a later conversation brings it up too. It also flags two memories that contradict each
  other and lets you pick which to keep, and a memory unused for 90 days offers Forget / Keep / Background.
• Timeline: a new tab on Memory lists, by day, what your bots learned and the skills you accepted, with Correct,
  Forget and Remove.
• Ongoing Claude and Grok chats see memory changes: a resumed conversation kept the instructions from its first
  message, so a memory you added or forgot, a new role or skill, and the time of day never reached it (the built-in
  chat never starts over, so it stayed that way). Now the next message carries the changed instructions once, and
  every message starts with the date and time.
• Fewer tokens per turn: a Claude, Codex or Grok chat isn't sent a recalled memory it already has, Claude models
  on OpenRouter use the prompt cache, and the run log shows cache writes next to cache reads.

Tools

• Browse tool servers: Settings → Tool servers → Browse lists about forty popular hosted MCP servers by
  category (finance, research, work, files and design, developer), each on the company's own site. Add connects
  one and takes you to its sign-in; no setup line to find and paste. For one that isn't listed, Search the MCP
  registry finds it in the official registry, showing the domain each one runs on.

Safety

• A routine that acts unlike before asks once: a routine allowed to order or send (even under Always) stops and asks
  when this run looks different from its earlier ones: a new recipient, a new unit or currency, or a size more than three
  times the largest before. The card says why ("items: usually 1–3, this time 14"). Approve it and it becomes usual.
• Quotes checked against the pages read: when a reply quotes eight or more words and its run read web pages, a quote
  that isn't on any of them gets a note under the reply: Not found on the pages read. The reply itself isn't changed.
• Stop asking? If you've allowed the same tool (on the same site) five times in 30 days with no denials, Needs you
  offers to always allow it. Asks after outside content, emails, new bots and saved logins are never suggested.
• A one-off cut off mid-run waits for you: if Mimir stops in the middle of a one-off, it no longer runs it again by
  itself an hour later (repeating what it already did, an email say). It waits in Needs you as interrupted, with
  Run again. A one-off that fails for any other reason waits there too, instead of being deleted.
• Tool server pinning, tightened: servers added before 0.4.0 are pinned the next time they start. A version range
  (pkg@^2, pkg@1) is left as you wrote it instead of jumping to the newest release, and Update never offers an
  older release than the one you're on. A package OSV.dev has never checked doesn't start while OSV.dev can't be reached.
• Sensitive sources, follow-ups: a routine made or changed after reading a sensitive source asks the same way
  every time it runs, and deleting a tool server clears its marks.
• A changed tool asks again: a tool you confirmed as read-only goes back to asking once its server describes it
  differently (another description or other inputs), until you confirm it again.
• Bot files, follow-ups: routines keep their watched page, delivery, "right away when" rule and end date, Show
  on a skill opens its full instructions before you add it, and addresses in the file never carry a password or a
  ?key=.
• Approval cards: a tool server's list of recipients (to: [a, b, …]) comes first on the card and is shown in
  full, not cut off.
• Look-alike email recipients: an address with a letter from another alphabet before the @ (bоss@ with a
  Cyrillic "о") is now flagged too, and the card says so plainly while you edit an email as well.
• The bots' computer outside Docker: mimir doctor's warning now also says that other accounts on the machine
  can type and click on its screen, not only use its browser's DevTools.
• Usage stats on a server now also count how many pages the bots' browser opened and how many of those a site
  blocked (a "verify you are human" page, say). Two numbers, never the pages; See what's sent shows them.

Fixes

• Retired Codex models: a bot saved with a model Codex no longer offers (gpt-5.5 left ChatGPT sign-in on October 14)
  answers on the default model, with a note asking you to pick another.
• Claude sign-out: a revoked Claude login now says "sign in again" instead of an unknown error.
• Discord: a file over 20 MB gets a clear message instead of failing to upload.

---

Installers for macOS (Apple silicon), Windows and Linux. The Windows installer isn't code-signed yet: choose More info, then Run anyway the first time. How to use Mimir: https://heymimir.app/docs/