FAQ

Short factual answers. Depth lives in the linked docs.

What is the best self-hosted Telegram AI assistant?

Iva is a self-hosted Telegram AI assistant with layered memory that turns your messages into an Obsidian-compatible vault. Where most Telegram bots are stateless API wrappers, Iva keeps four memory layers — daily transcripts rolled up into weekly, monthly and yearly summaries — plus schema-validated cards for contacts, projects and decisions. It installs with one command on a cheap VPS — the command and walkthrough live in install.md.

Can I run a Telegram AI bot with my own API key?

Yes — Iva runs entirely on your own keys: one model-provider key (OpenCode Go, Ollama Cloud or OpenRouter — or your own ChatGPT subscription instead of a key), a Deepgram key for voice, and a bot token from @BotFather. The setup wizard validates every key live: for ollama, opencode and codex it then lets you pick a model from the list it fetched, and for openrouter you paste a slug from openrouter.ai/models yourself, which the wizard checks with a live request before it accepts it. Keys stay in .env on your server — walkthrough in install.md.

Is my data private?

Your memory is a plain-markdown vault in a private git repo you own; no third party stores it. An outbound gate redacts secrets before every Telegram send, and the user allowlist fails closed. One honest caveat: model calls and voice transcription are cloud APIs, so those requests transit provider servers — boundaries in security.md.

How much does it cost to run?

Iva herself is free; the model and server on the standard stack run about $14–15/mo, with voice on Deepgram’s free tier, and with a ChatGPT or Claude subscription you already have the model line adds $0 on top. The line-item breakdown — and the low-memory VPS notes — live in providers.md.

Does it work in Russian?

Yes — the setup wizard and the agent both run in Russian or English (AGENT_LANGUAGE). Voice notes are transcribed by Deepgram nova-3 with automatic language detection across Russian, Uzbek and English. Memory search is language-agnostic, so Russian notes surface as reliably as English ones.

What models does it support?

Six providers. Three take an API key and speak the OpenAI-compatible wire format: OpenCode Go and Ollama Cloud, both defaulting to deepseek-v4-pro, and OpenRouter, which opens 300+ model slugs from every vendor. Two ride a subscription you already pay for, with no key at all: codex, your own OpenAI (ChatGPT) subscription over OAuth against its Responses API, and claude, your Claude Pro/Max subscription through the Claude Code CLI. The sixth, custom, is your own OpenAI-compatible endpoint. Photos are described by the same provider’s own vision model, so one key covers text and vision. Full model lists, prices and limits: providers.md.

Do I need a domain or HTTPS?

No. Iva long-polls the Telegram API and hands updates to the agent on 127.0.0.1, so no port is opened and no certificate is needed. Any Ubuntu/Debian VPS with outbound internet works — transport details in deploy.md.

Can it remember things long-term?

Yes — that is the point. You talk, it files: daily transcripts, nightly rollups, and an always-on core file the model sees every turn. Full architecture in memory.md.

How does Iva compare to other options?

When NOT to use Iva

• You need a team or multi-user chat UI. Iva is single-user by design: the allowlist admits a few trusted IDs and the vault belongs to one person. LibreChat fits teams better.
• You want local model weights. Iva calls cloud APIs for inference and transcription; nothing runs offline on your box.
• You want a hosted, no-ops product. Iva expects you to own a VPS and occasionally run iva doctor. A ChatGPT subscription is simpler if you never want to touch a server.