Install
Through an AI agent
Paste this into Claude Code, Codex or Claude Desktop. The agent will check your server or Mac, run the installer and help you get a reply from your bot. Instructions for the agent.
Install Iva for me, instructions: https://iva-agent.com/install.txt
On macOS, the same installer uses Homebrew and launchd. Units are named app.iva.*; plists live in ~/Library/LaunchAgents and logs in ~/Library/Logs/Iva/. Log in on the Mac screen before installing. The Mac must stay on and awake for Iva to work. A night missed while the Mac is off is skipped; after sleep, launchd runs the pending calendar job when the Mac wakes. Use iva status and iva logs -f to inspect the installation.
Everything between curl and a working bot. One command on a fresh server: the installer asks your language, walks you through five keys, and ends by messaging you from your own bot.
Requirements
- 🖥️ A server or your own machine — Ubuntu/Debian is the tested path (apt); Fedora (dnf) and macOS (brew) work too. Any always-on box.
- 🧠 512MB RAM is enough — on boxes under 1.5GB the installer adds a 2GB swapfile so the build isn’t OOM-killed (needs ~2.6GB free disk).
- 💾 At least ~1.5GB free for Iva itself — two version directories of ~400MB each plus the npm cache; allow 2GB more when the installer adds a swapfile.
- 🔑 sudo — asked up front, and only if system packages are missing or a swapfile is needed; the Chromium step may ask once more.
Never used a server? The host sends you an address (IP), a login and a password. On Mac or Linux open Terminal, on Windows PowerShell, type
ssh root@YOUR_ADDRESS, enter the password. You’re in. First thing: make yourself a normal user and switch to it —adduser iva && usermod -aG sudo iva && su - iva— and install from there. Iva’s shell tool runs with the permissions of whoever installed it, so don’t hand it root.
If the installer ends with
Failed to connect to bus: No medium found, leave thesu - ivashell. Enable linger and reconnect directly:exit # back to the root session on the server loginctl enable-linger iva exit # back to your own computer ssh iva@YOUR_ADDRESSThen run the installer again.
Step 0 (optional): prepare a fresh VPS
Brand-new server, still logged in as root? One command gets it ready — run it as root, before anything else:
bash <(curl -fsSL https://raw.githubusercontent.com/smixs/iva-agent/main/bootstrap.sh)
It asks three things up front — a login, its password, and the timezone — then: updates the system, installs Iva’s system packages (git gh python3 ffmpeg pandoc poppler-utils), creates that user with sudo and systemd lingering already enabled, turns on a firewall that allows SSH only, starts fail2ban and unattended security upgrades, and hardens sshd last — with a reload, so the session you’re typing in is never dropped. Every step is detect-then-skip, so re-running it is safe. It ends by printing the ssh and install.sh commands to run next. Log: /var/log/iva-bootstrap.log.
No SSH key is involved: you log in as the new user with the password you just set, and hardening stops at PermitRootLogin no. Headless runs take --non-interactive with IVA_USER and IVA_PASS, plus optional IVA_PUBKEY (authorizes a key), IVA_TZ and IVA_DISABLE_PASSWORD_AUTH (key-only login — ignored without a valid key).
Install
curl -fsSL https://raw.githubusercontent.com/smixs/iva-agent/main/install.sh | bash
The first question is your language — English or Russian — before anything touches the system. Input is read from /dev/tty, so the wizard stays interactive even piped through curl. If there’s no terminal at all (Docker, CI), setup is skipped and the script prints how to run it later.
Setup wizard
Five steps. Each key comes with a direct link to where it lives, and each is validated live — a bad key is rejected on the spot, not discovered at runtime. Enter keeps the current value, so re-running the wizard (iva config) changes only what you want.
- Provider and model. Ollama Cloud, OpenCode Go, an OpenAI (ChatGPT) subscription, OpenRouter or your own OpenAI-compatible endpoint (comparison); the key (or the subscription sign-in) is checked, then you pick a model.
- Voice, search, hybrid memory. Deepgram key (free starter credit) — Enter skips it, voice notes stay untranscribed until you add the key in
/menu→ 🎤 Voice; recognition languagemultiauto-detects ru/uz/en. The same step picks a web-search provider — Tavily, Exa, Parallel or Brave; Enter skips and search stays off — and offers optional hybrid memory with an embedding key. - Telegram bot. Paste the token from @BotFather; the wizard validates it via
getMeand detects the bot’s username itself. - Access. Send your new bot any message — “hi” works. The wizard reads
getUpdates, shows who wrote, and you pick yourself. Iva answers only these IDs; an empty list means it answers nobody. - Timezone, vault, port. IANA timezone so nightly jobs run on your clock, the vault directory, and the port — default 8723, probed for conflicts.
What install.sh does
- 📦 System packages —
git gh python3 ffmpeg pandoc poppler-utils(poppleron brew):ghbacks your vault up to a private GitHub repo, pandoc and poppler extract text from incoming docx/pdf files, ffmpeg converts media the transcriber can’t take directly. - 🐍 uv — runs the vault’s Python maintenance scripts.
- 🟢 Node 24 via nvm — no root needed; 24 is a hard floor because memory search uses the built-in
node:sqlite. - 🌐 agent-browser + Chromium — headless browser for web tasks; the longest step, 1–3 minutes of visible download output.
- 🗂️ Vault init — your memory is created from
vault-template/as a separate git repo, so personal data never enters the code repo. - ⚙️ Background runtime — two systemd user services, two systemd watchdog timers and seven in-process eve schedules, with linger enabled so they survive logout. Details: deploy.md.
- 🧰 The
ivacommand — installed into~/.local/bin:iva status,iva doctor,iva update. Full reference: cli.md. - ✅ Telegram confirmation — the last thing the installer does is message you from your own bot: “Iva is installed and online. Send me a message — I’ll reply.” That’s the success signal.
Re-running the same command later is safe: over an installation that already exists the installer updates nothing itself — it hands the installation to the one updater, exactly as iva update and repair.sh do. A checkout is put back onto its release first (edits to Iva’s own code are removed, .env, data/ and the vault are not touched), a versioned installation goes straight to its own updater, and a checkout you marked with .iva-dev is refused.
Google Workspace CLI (gws) is installed and updated under ~/.local without sudo. Services prefer ~/.local/bin over older global copies. Existing binaries are left in place; Google configuration and authorization stay under the same user account.
Every stage checks whether its work is already done and skips it, so a run after a failure costs seconds instead of minutes:
| Stage | Skipped when |
|---|---|
| System packages | the command is already there (git, gh, python3, ffmpeg, pandoc, pdftotext) |
npm ci | node_modules matches package-lock.json — npm’s own record, including the pinned eve archive |
| agent-browser | the binary is installed and really opens a page |
gws | the binary is installed (iva update keeps it current) |
| Build | .output carries this installer’s stamp for the current commit, local edits and .env |
The wizard, the vault check, the iva command and the systemd units are cheap, so they run every time. A run that fails is undone: the copy it made of .env goes back, the build it replaced is put back, and both copies are deleted — on Ctrl-C and on a dropped SSH session too. The code is never this script’s to move.
If the undo itself cannot finish — a read-only checkout, a full disk — nothing it saved is thrown away. Whatever is still the only copy of something stays where it is, and the installer prints each one by name before it exits: the copy of .env under data/update-backups/, and the previous build under .output.iva-install-backup-*. Read those lines before running anything else. An installation unpacked from an archive instead of cloned has no commit to compare against, so it rebuilds every time; so does one with a file the build cannot read.
Only one installer runs in an installation at a time: a second one is refused by name, with the process id of the one already working. A lock left by a run that no longer exists is taken over, so a power cut cannot leave the installation unusable.
Flags and overrides
Flags pass through the pipe with bash -s --:
curl -fsSL https://raw.githubusercontent.com/smixs/iva-agent/main/install.sh | bash -s -- --skip-setup
| Option | Effect |
|---|---|
--skip-setup | install everything, don’t run the wizard |
--non-interactive | no questions at all — defaults only, wizard skipped |
-h, --help | show the built-in help and exit |
REPO_URL=… | install from a fork (default https://github.com/smixs/iva-agent.git) |
BRANCH=… | install and keep updating from this branch (default main) |
INSTALL_DIR=… | where the code goes (default ~/iva) |
The last three are environment variables, read by the script at startup.
If the wizard didn’t run
Skipped setup, or no terminal at install time:
cd ~/iva && npm run setup
Then repeat the regular install command:
curl -fsSL https://raw.githubusercontent.com/smixs/iva-agent/main/install.sh | bash
The installer continues the first installation while autostart is still pending. It reuses completed stages, opens the wizard with your saved values as defaults, enables autostart and sends the confirmation in Telegram. On Linux it also enables linger so Iva keeps running after you leave the SSH session. Once autostart is configured, later runs use the updater.
Next steps
- Every
.envvariable, defaults and warnings — configuration.md - The
ivaserver CLI and Telegram commands — cli.md - Transport, timers, webhook mode and operations — deploy.md