Self-hosting

The desktop app runs your bots while it's open. To keep them online 24/7, run Mimir on a machine you own: a small VPS (any cloud's cheapest Linux server is enough), a home server, or a Raspberry Pi 4 or 5. It runs all your bots, keeps your memory, and serves the same app in your browser. You can also point the desktop app at it.

It comes as a ready-made Docker image (ghcr.io/currentbits/mimir, for Intel/AMD and ARM), started with Docker Compose. You don't need the source code or to build anything.

What you need #

• A Linux machine that stays on, with about 2 GB of memory and 5 GB of disk. A Mac or Windows PC with Docker Desktop works too.
• Docker with Compose. On Linux the quickest way is Docker's install script (curl -fsSL https://get.docker.com | sh, see docs.docker.com/engine/install); on a Mac or Windows, Docker Desktop. Check with docker compose version.
• A terminal on that machine (over SSH for a VPS).

1. Get Mimir's compose file #

In an empty folder on the server:

mkdir mimir && cd mimir
curl -fsSLO https://heymimir.app/docker-compose.yml
echo "MIMIR_COMPUTER_TOKEN=$(openssl rand -hex 24)" > .env

docker-compose.yml (the same file as in the repository) describes the two containers Mimir runs as (below); .env holds a random key they use to talk to each other.

2. Start it #

docker compose up -d                          # downloads the image (about 800 MB) and starts Mimir
docker compose exec mimir mimir admin-token   # prints your admin token: copy it

Mimir now runs at http://localhost:8080 on the server, and starts again after a reboot. On purpose it only answers on the server itself, so nobody else can reach it until you choose how (next step).

3. Open it in your browser #

Pick one:

• Through an SSH tunnel (simplest, nothing exposed). On your own computer: ssh -L 8080:localhost:8080 you@your-server, then open http://localhost:8080 and paste the admin token.
• With HTTPS on your own domain (to use it from anywhere, and to connect the desktop app over the internet). Point a DNS name (say bots.example.com) at the server, open ports 80 and 443, then in docker-compose.yml uncomment the caddy service and the caddy-data volume and put your name in. docker compose up -d again, and Caddy gets a certificate on its own: open https://bots.example.com.
• At home or on Tailscale. On a home network, change the port line to "8080:8080" and open http://<the server's address>:8080 from your own devices. With Tailscale on both, the server's 100.x address or .ts.net name works the same way, from anywhere.

Then the app is the same as on the desktop: make a bot, connect Telegram, Slack or Discord, and set up email and web search in Settings (Getting started).

4. Sign in to your AI accounts #

The image includes the official claude and codex apps, so bots can run on your Claude or ChatGPT plan. Sign in once, in the server's terminal:

• Claude: docker compose exec -it mimir claude, sign in, then /exit (or set CLAUDE_CODE_OAUTH_TOKEN from claude setup-token on any machine).
• ChatGPT: nothing in the terminal. Settings → AI keys → Continue with ChatGPT; after you sign in, your browser lands on a page that doesn't load (http://127.0.0.1:…). Paste that address into Mimir. (Or the Codex CLI: docker compose exec -it mimir codex login.)
• Grok: not in the ready-made image (its installer is a script from x.ai without a fixed version). Build the image yourself with --build-arg GROK_CLI=1 (below), then Sign in to Grok in Settings.

It's your server and your login, the same as on your laptop. Bots on an API key (OpenRouter, OpenAI, xAI) or a model running elsewhere need nothing extra.

Updating #

docker compose pull && docker compose up -d

Your data is kept: it's in the mimir-data volume, not in the container. To stay on one version, put MIMIR_VERSION=0.1.0 in .env.

When a new version changes how the database is laid out, it first copies it aside as mimir.db.before-<n> in the same folder (the last 2 are kept), so going back means stopping Mimir and putting that copy back as mimir.db. An older version won't open a database a newer one has upgraded: it stops with "This data was written by Mimir X" instead of changing it. The same holds for moving a database between the desktop app and a server.

The server doesn't update itself, but it tells you: twice a day it looks for a newer release, and when there is one it says so once per version in your most recent private chat with your bots (Telegram, Slack or Discord; never a group or a guest's chat), with the commands above, and shows it at the top of Needs you until you've updated. Switch it off in Settings → This server → Tell me when there's a new version. Releases say what changed.

Day to day #

• Logs: docker compose logs -f mimir. Stop: docker compose down (your data stays).
• Is everything working? docker compose exec mimir mimir doctor checks each bot's AI sign-in, that each Telegram, Slack and Discord token and each mailbox still signs in, that the bots' computer answers, free disk space (it warns under 1 GB), and that the database is sound, not from a newer Mimir, and backed up in the last two days. It also warns about what leaves the server open: the web app reached over plain HTTP from the internet, Permissions at "Allow all", tool servers without a fixed version, a newer Mimir being out, and the bots' computer running in Mimir's own process on a host without Docker. One line each (✓ fine, ! worth a look, ✗ broken, with what to do), and it exits non-zero when something is broken, so a monitoring script can run it. Nothing is spent or sent. Settings → Your data → Check everything shows the same.
• A new admin token: docker compose exec mimir mimir admin-token --new, then docker compose restart. Signed-in browsers and connected apps need the new one.
• Backups: Mimir keeps a checked copy of its database every day in /data/backups (7 daily, 4 weekly), and Settings → Your data restores one. To restore a copy from elsewhere, put it there as mimir-<anything>.db. For a copy off the server, back up the mimir-data volume: it holds the database (memories, bots, chats) and the secrets. For example docker run --rm -v mimir_mimir-data:/data -v "$PWD":/out busybox tar czf /out/mimir-backup.tgz /data (the volume's name starts with the folder's name; docker volume ls shows it).
• Two containers: mimir (bots, web app, your data) and computer (the bots' computer: a screen and Chromium, as its own user, without your data). They talk on their own network only; the computer publishes no ports. Both drop all Linux capabilities and can't gain privileges.

Building the image yourself #

In a clone of the source (git clone https://github.com/currentbits/mimir && cd mimir), with the .env from step 1: docker compose up -d --build. Build options: --build-arg GROK_CLI=1 adds Grok, --build-arg AI_CLIS="" leaves out the Claude and Codex apps, --build-arg COMPUTER= leaves out the bots' computer. The AI apps, the browser tool and the base images are fixed versions, so a rebuild doesn't change what runs your bots. Update with git pull && docker compose up -d --build.

Without Docker #

cargo build --release -p mimir && (cd ui && npm ci && npm run build)
MIMIR_UI_DIR=ui/dist mimir server --listen 127.0.0.1:8080

Connecting the desktop app #

Settings → Your server: enter the server's address and admin token. The app then shows the server's bots, memory and routines, and the bots on this computer pause, so nothing runs twice. Disconnect switches back. The token is kept in the keychain. Use an https:// address for a server on the internet; plain http:// is accepted only for addresses at home or on Tailscale (localhost, 192.168.x, 10.x, 100.64–127.x, .local, .ts.net), since the token would otherwise travel in the clear.

Some things stay on the computer even when connected, because they belong to it: meeting recording and the meetings list, the quick-ask shortcut, keeping the computer awake.

Chat from Shortcuts, Home Assistant and other apps #

Settings → Chat API makes a token for one bot. Give an app that speaks OpenAI's chat API the base URL https://<your server>/v1, the token as its API key, and any model name. It can then ask that bot (POST /v1/chat/completions, GET /v1/models) and do nothing else: the token doesn't open /api or any other bot.

The bot answers in its built-in chat, so what the app asked shows up there. Mimir keeps the conversation itself: only the last user message of each request counts, and the app's own history is ignored. "stream": true gets the whole reply as one chunk once it's done. Anything the bot needs approved waits in Needs you as usual, so set up such a bot with the tools it needs allowed. The token is shown once and only its hash is kept; delete it there to cut the app off, and deleting a bot deletes its tokens.

Moving from the desktop app #

Copy ~/.mimir/mimir.db into the server's data volume before its first start, then re-enter messaging-app tokens on the server (the desktop keeps them in the OS keychain, not the database).

How the server works #

mimir server runs the same host as the desktop app: every bot's Telegram / Slack / Discord connection, the routine scheduler and approval cards, plus an HTTP API (POST /api/<command> with the admin token, the same commands the desktop UI uses) and the web UI. Secrets are stored in its database instead of a keychain; the data folder is owner-only (mode 0700). The API checks the token before reading a request; the web UI can't be framed by other sites and loads only its own scripts. GET /healthz answers ok for health checks. A MIMIR_ADMIN_TOKEN you set must be at least 24 characters. POST /hook/<id> is a routine's webhook (Routines): it takes that routine's own secret, never the admin token, so put it on the internet only if callers need to reach it.

Email, web search, the file workspaces and the skill library work the same on a server: set them up in the web UI's Settings (Connect your email, calendar and web search). The calendar link is set under Meetings, which is in the desktop app only, so a server's bots can't read a calendar yet (add-to-calendar links work everywhere). The server doesn't update itself: see Updating.