Lilo is a personal AI assistant that can help you with everything in life. It can track and execute your TODOs, remembers things, writes documents, builds apps, conducts research as well as remind you about things. You can reach it from anywhere: WhatsApp, Telegram, desktop, mobile or email.

Features · Quick start · Configuration · Workspace apps · External messaging · Mobile app · Security · Deployment

• Node.js 20+
• pnpm 10+
• An API key for at least one of: OpenAI, Anthropic, OpenRouter

git clone https://github.com/abi/lilo.git
cd lilo
pnpm install

Create a .env.local at the repo root:

# Required
LILO_WORKSPACE_DIR=./workspace          # where the agent's files live
LILO_SESSIONS_DIR=./.lilo-sessions      # persistent chat session storage

# At least one chat model
OPENAI_API_KEY=sk-...                   # enables GPT 5.5
ANTHROPIC_API_KEY=sk-ant-...            # enables Claude Opus 4.7
OPENROUTER_API_KEY=sk-or-...            # enables OpenRouter-routed models

# Recommended
LILO_AUTH_PASSWORD=choose-a-strong-password   # locks down the whole app

See Configuration for the full list.

pnpm run dev   # backend (http://localhost:8787) + frontend (http://localhost:5800) + typechecks

Open http://localhost:5800. If you set LILO_AUTH_PASSWORD, you'll get a login screen on first visit.

Your workspace should be auto-bootstrapped from the bundled
[workspace-template/](./workspace-template), so you'll immediately have a
Desktop, TODO list, Calories tracker, and a handful of other apps to play with.

All env vars are read from (in order of precedence):

1. Shell-exported variables
2. .env.local at the repo root
3. .env at the repo root

At least one is required to actually use Lilo.

If OPENROUTER_API_KEY is set, Lilo can route supported models through OpenRouter. Native provider keys take priority: for example, if OPENAI_API_KEY is set, GPT models use OpenAI directly; if it is missing but OPENROUTER_API_KEY is set, GPT models route through OpenRouter instead.

Limit the chat dropdown/API to specific models with a comma-separated allowlist.

LILO_CHAT_MODEL_ALLOWLIST=gpt-5.5,gpt-5.4-mini

Supported allowlist IDs: claude-opus-4-7, gpt-5.5, gpt-5.4-mini, and moonshotai/kimi-k2.6.

Each one unlocks a corresponding agent tool. Missing keys just disable the tool — the agent keeps working without them.

Point LILO_WORKSPACE_GIT_URL at a git repo to make LILO_WORKSPACE_DIR git-backed on boot. If the workspace is not already a git repo, Lilo initializes one and sets origin to this URL; if it is already a repo, Lilo keeps origin in sync with this value. This keeps your workspace (apps, data, and memories) versioned and portable across hosts. The frontend shows a manual "Sync" button that runs pull/rebase and push; it expects workspace changes to be committed first. Hide it with VITE_ENABLE_WORKSPACE_SYNC=false when you aren't using this flow.

Set at build time — Vite only inlines VITE_* vars.

The most distinctive thing about Lilo: the agent builds its own apps. Each app lives as a directory of HTML + assets under $LILO_WORKSPACE_DIR/, runs in a sandboxed iframe, and can read/write its own files, open chats, and make HTTP calls through a built-in window.lilo API. Ask the agent "build me a habit tracker" and it scaffolds one — no build step.

→ Full guide: docs/workspace-apps.md (directory layout, the window.lilo API surface, and the in-viewer element picker).

Lilo can be an email/SMS/Telegram chatbot. Each channel is an opt-in plugin — leave its env vars unset and it's disabled.

RESEND_API_KEY=re_...
RESEND_WEBHOOK_SECRET=whsec_...
LILO_EMAIL_AGENT_ADDRESS=hi@yourdomain.com        # your bot's inbound address
LILO_EMAIL_REPLY_FROM="Lilo <lilo@yourdomain.com>"
LILO_EMAIL_ALLOWED_SENDERS=you@yours.com,partner@theirs.com   # allowlist
LILO_PUBLIC_APP_URL=https://your-lilo.example.com # optional; adds chat permalinks to email replies

1. Set up a receiving domain in Resend.
2. Create a webhook pointing to https://your-lilo/api/inbound-email with the email.received event.
3. Send Lilo an email; it replies in-thread.

Replies set Reply-To: LILO_EMAIL_AGENT_ADDRESS, so the recipient's reply round-trips back into the same inbox.

TWILIO_ACCOUNT_SID=AC...
TWILIO_AUTH_TOKEN=...
LILO_WHATSAPP_AGENT_NUMBER=whatsapp:+15555550123
LILO_WHATSAPP_ALLOWED_SENDERS=whatsapp:+15555550124

Point a Twilio WhatsApp webhook at https://your-lilo/api/inbound-whatsapp.

TELEGRAM_BOT_TOKEN=123456:ABC-...
TELEGRAM_WEBHOOK_SECRET=long-random-secret
LILO_TELEGRAM_ALLOWED_USER_IDS=123456789

Point your Telegram bot webhook at https://your-lilo/api/inbound-telegram and pass the same secret as Telegram's secret_token:

curl -X POST "https://api.telegram.org/bot$TELEGRAM_BOT_TOKEN/setWebhook" \
  -d "url=https://your-lilo/api/inbound-telegram" \
  -d "secret_token=$TELEGRAM_WEBHOOK_SECRET"

LILO_TELEGRAM_ALLOWED_USER_IDS must contain numeric Telegram user IDs, not group chat IDs. To find yours, message your bot and inspect getUpdates, or use a Telegram ID helper bot such as @userinfobot.

Each contact gets their own persistent chat, so the agent remembers your conversation across messages.

Lilo works well as an iPhone home-screen app. It still needs your Lilo backend to be running; this is an app-like Safari wrapper, not an offline native app.

1. Deploy Lilo somewhere your phone can reach, ideally over HTTPS. Railway's generated public domain works well.
2. On your iPhone, open the Lilo URL in Safari. Use Safari specifically; other iOS browsers do not expose the same Add to Home Screen flow.
3. Log in if you configured LILO_AUTH_PASSWORD.
4. Tap the Share button, then tap Add to Home Screen.
5. Name it Lilo, tap Add, and launch it from your home screen.

Lilo is built for a single, self-hosted user — not multi-tenant SaaS.

• Set LILO_AUTH_PASSWORD before exposing an instance to the internet. It gates the web app, all REST/SSE/WebSocket endpoints, and workspace apps. Without it, everything is open.
• The agent has shell, filesystem, and network access. Prompt injection is a real risk — anything it reads (web pages, emails, PDFs) can try to hijack it. Keep messaging allowlists (LILO_EMAIL_ALLOWED_SENDERS, LILO_WHATSAPP_ALLOWED_SENDERS, LILO_TELEGRAM_ALLOWED_USER_IDS) tight, and keep credentials in env vars, not the workspace.
• Webhooks (/api/inbound-*) skip the password gate on purpose and use provider-signed verification instead.
• Report security issues privately through GitHub Security Advisories. Please don't open public issues for sensitive reports.

./scripts/setup-railway.sh

The interactive setup script will:

• Link or create a Railway project named lilo
• Prompt for the GitHub repo to connect
• Set core env vars
• Mount a persistent volume at /data for LILO_WORKSPACE_DIR + LILO_SESSIONS_DIR
• Generate a public domain on port 8080

You still configure the optional message-channel keys (Resend / Twilio /
Telegram / Firecrawl / etc.) from the Railway dashboard.

pnpm run dev            # run backend + frontend + live typechecks in parallel
pnpm run dev:backend    # port 8787
pnpm run dev:frontend   # port 5800
pnpm run dev:template   # like `dev` but with LILO_WORKSPACE_DIR pointed at the bundled template (useful for trying out the default apps without polluting your own workspace)
pnpm run build          # build both packages
pnpm run lint           # oxlint across the repo
pnpm run format         # oxfmt across the repo

This is an open source project. Issues and PRs welcome — keep components small and one-file-per-concern (see AGENTS.md).

Join us on Discord to ask questions, share workspace apps you've built, or follow development.

MIT