Mimir Get Mimir

Your own AI assistants, always on

Bots that remember you, work on a schedule and ask before they act. One app on your computer, and they answer in Telegram, Slack, Discord and Signal, on the Claude, ChatGPT or Grok plan you already pay for. No Mimir account, no Mimir server.

DownloadTake the tourFree and open source · macOS · Windows · Linux

The Mimir app: eight bots in the sidebar, each with the app it answers in, and Ada’s chat, where she asks before sending an email to Sam.

Things you can ask

A tour of the app

The real app, with a week of demo data. It looks like this on Mac, Windows and Linux, and in your browser when it runs on a server.

What a routine sends

Every weekday at 7:30 the Morning briefer writes up your day: calendar, weather, and the one email that needs you.

  • It sees its last three runs, so it says what’s new
  • Sent to the chat you pick, or saved for your daily digest
  • Started by a time, a meeting, a matching email or a webhook
The Morning briefer’s chat: a short brief each morning with today’s meetings, the weather and the email that needs you.

A memory you can read

Everything your bots know about you, line by line, and who wrote it. Tell one bot and they all know.

  • Suggestions wait for your Accept
  • A correction keeps what it replaced; Undo brings it back
  • Pin what should always be in front of them
The Memory view: two suggested changes to accept or dismiss, what was learned this week with an Undo on each, and pinned memories.

Routines, in plain words

Say “every weekday at 7:30, brief me” and it’s scheduled. Each routine shows how it’s doing and what it costs.

  • Quiet checks don’t wake the AI while nothing changes
  • Rehearse a routine before it runs for real
  • Daily caps; failures handled by cause
The Routines view: a weekday morning brief, a quiet check every 15 minutes, inbox sorting and Friday reviews, each with its recent runs and cost.

It asks before it acts

Every approval, and anything that went wrong, in one place. Or answer right in the chat you’re in, on your phone too.

  • Allow once, always on this site, or deny
  • Sending email never becomes “always”
  • No answer means no
Needs you: Ada waiting to send an email, the Trip planner waiting to add a calendar event, memory suggestions, and runs that didn’t work.

Each bot, your way

A name, a face, a role in your words, the AI it thinks with, and the tools it may use.

  • Claude, ChatGPT, Grok, OpenRouter, an API key or a local model
  • Each tool allowed, asks first, or off
  • Bots hand tasks to each other
Ada’s settings: her name, her role, and the AI account she uses, chosen from Claude, ChatGPT, Codex, Grok, OpenRouter, xAI, a local model or an API key.

Skills, from what worked

The steps for one kind of task. From the library, or suggested after a task went well.

  • Twenty-four in the library, in plain text
  • Say “save that as a skill”
  • Run one with its name, like /morning-briefing
The Skills view: a suggested skill for chasing an unpaid invoice, and the Morning briefing skill’s steps.

Everything on the record

Every reply and every routine run: what it did, the tools it used, and the tokens it took.

  • Runs, tokens and what they’d cost on the API
  • Quiet checks that changed nothing cost nothing
  • A daily digest, and quiet hours
The Activity view: 31 runs in 30 days with their tokens and cost, most of them quiet checks that used no AI.

What they work with

Each is optional. Connect what you want your bots to use.

Where you talk to them

  • Telegram
  • Slack
  • Discord
  • Signal
  • WhatsApp
  • iMessage, on a Mac
  • The desktop app
  • Voice notes

What they think with

Claude
Your Claude subscription, through the claude app you’re signed in to.
ChatGPT
Your ChatGPT subscription, through codex.
Grok
Your SuperGrok plan, through grok.
OpenRouter
One key for hundreds of models.
Local
Ollama or LM Studio. Nothing leaves your machine.

What they can use

  • Email: search, read, draft
  • Calendar
  • Web search
  • X search
  • A real browser, with logins the AI never sees
  • Meeting and call transcripts
  • Notes and files
  • 40+ tool servers: Stripe, Notion, Linear, GitHub…
  • A computer, on your server

Questions? Ask Ada

The ones people ask most.

What does it cost?
Mimir is free and open source. Your bots run on the AI plan you already have (Claude, ChatGPT or Grok). With an API key you pay that service for what you use; a model on your own computer is free.
Do I need an API key?
No. For Claude or ChatGPT you sign in once to their official command-line app (claude or codex), and Mimir uses that login.
Where does my data go?
Memories, chats and settings are one file on your computer or server. What a reply needs goes to the AI you chose. There are no Mimir servers.
Is it safe to connect my email?
Reading changes nothing. Every email it would send asks you first, with the whole text, and sending never becomes “always”. It uses an app password you can revoke.
Does my computer have to stay on?
Bots work while the app runs in the menu bar or tray. For bots that work while your laptop sleeps, run Mimir on a server.
How is this different from the Claude or ChatGPT app?
It uses them, and adds bots in your messaging apps, routines that run while you’re away, one memory you can read line by line, and approvals for every action that matters.

On your computer, or a server you own

About ten minutes from download to your first message: Getting started.

On Windows, one extra step the first time

The installer isn’t code-signed yet, so SmartScreen says “Windows protected your PC”. Click More info (under the message), then Run anyway, and install as usual. After that Mimir updates itself. If your browser warns about the download, choose Keep.

The Mac app is signed and notarized. A .deb, .rpm and .msi are on the releases page.

Your own server

For bots that work while your laptop sleeps: any small server with Docker. In an empty folder:

curl -fsSLO https://heymimir.app/docker-compose.yml
echo "MIMIR_COMPUTER_TOKEN=$(openssl rand -hex 24)" > .env
docker compose up -d
docker compose exec mimir mimir admin-token

The last line prints your admin token to sign in with. Step by step.

Open source under the AGPL-3.0, on GitHub. Everything your bots remember is one SQLite file in ~/.mimir. Everything else is in the docs.