TeleCrow guide

Connect ChatGPT or Claude to Telegram — No Code, Your Own API Key (2026)

To connect ChatGPT or Claude to Telegram without code: create a bot with @BotFather, paste its token into TeleCrow's Personal AI template, then add your own OpenAI or Anthropic API key and pick a model. TeleCrow hosts the bot for you — no Zapier/Make glue, no server, and no per-message markup; the AI usage bills at your provider's raw rates.

The real setup: a BotFather token plus your API key, live in minutes

The Create Bot wizard is two steps. Step one, pick the Personal AI template — it's first in the template list ("Your own AI chatbot — add an API key, pick a model, and it's live."). Step two, paste your bot token from @BotFather; if you've never made one, an in-app BotFather walkthrough shows you exactly which taps to make. TeleCrow verifies the token live against Telegram and shows your bot's name and username with a "Token verified" badge, and the token is stored encrypted. Click Create and TeleCrow provisions the bot in its own isolated Docker container — hosting is handled, but the AI stays switched off until you save a provider key.

Opening the bot lands you on a single "AI Brain" tab with three numbered steps. Step 1, Connect: choose OpenAI, Anthropic, Google Gemini, or OpenRouter, pick a model from the dropdown, and paste your API key (a "Need a key?" link points to the provider's key page, e.g. platform.openai.com or console.anthropic.com). On save, TeleCrow live-tests the key and model against the real provider API with a tiny one-token call — a bad key or a model that doesn't belong to that provider is rejected with a clear error before anything is stored, while a momentary provider rate-limit won't block the save.

Step 2, Personality: write the system prompt that sets your bot's tone, role, and rules — up to 8,000 characters. Step 3, Knowledge, is optional (more on it below). A collapsible Advanced block exposes the tunables: temperature 0–2 (default 0.7), max reply tokens 64–4,096 (default 800), and a context window of 0–40 remembered turns per chat (default 10; set 0 for a stateless bot). Hit save and the bot's container is rebuilt automatically with the new config baked in — changes go live within seconds, with nothing to redeploy. From there the bot answers any Telegram message through your configured model, and /reset clears a chat's conversation memory.

The cost math: your own key vs per-message SaaS vs self-hosting

Bring-your-own-key is the whole economic point. TeleCrow never meters your AI usage and never bills per message — there's no platform AI credit system at all; your OpenAI or Anthropic account is charged directly at provider list rates. In 2026 those rates are cheap at the volumes a chat bot generates: with budget models like gpt-5-mini, gpt-4o-mini, claude-haiku-4-5, or gemini-2.5-flash, a typical exchange under the default 800-token reply cap costs a small fraction of a cent. Even flagship models (gpt-5, claude-opus-4-8) cost more per token but still typically price a single reply in the low cents — check your provider's pricing page for current rates.

Compare that with per-message AI-chatbot SaaS, where plans commonly run $19–$99+ per month with caps of a few thousand messages and overage fees on top — you're paying an intermediary's markup on every model call. And compare it with the Zapier/Make/n8n route, where the automation platform's own task pricing stacks on top of the AI bill and you still have to wire the flow together yourself.

The DIY alternative — cloning a Python script and running it on a VPS — avoids markup too, but now you own a server ($5–10/month), process supervision, dependency updates, and keeping an API key safe on a box you administer. TeleCrow's trade: hosting, isolation, and the config UI are handled, and the plan gating is simply bot count — the free plan runs 1 bot (with "Powered by" branding appended to messages), Basic at $3.99/month runs 1, Pro at $6.99 runs 2, and Ultra at $11.99 runs 4. There is no AI-specific paywall: the Personal AI template works on every plan, including free.

Choosing your model: GPT, Claude, Gemini, or the OpenRouter catalog

The model dropdown is served live from the backend and covers the current lineups. On an OpenAI key that includes gpt-5, gpt-5-mini, gpt-5-nano, gpt-4.1, gpt-4o, and the reasoning models (o3, o4-mini, o1) among others. On an Anthropic key you get Claude Opus 4.8, Opus 4.7, Sonnet 4.6, and Haiku 4.5. A Google key unlocks Gemini 2.5 Pro and Flash plus the 2.0 and 1.5 Flash tiers.

OpenRouter is the wildcard option: one key routes to many labs' models, including openrouter/auto (which picks for you), GPT-5, Claude Sonnet, Gemini 2.5, Llama 3.3 70B and 3.1 405B, DeepSeek V3 and R1, Mistral Large, and Qwen 2.5 72B. And the dropdown isn't a hard allowlist — a "Custom model…" option lets you type any model id your provider serves; TeleCrow warns on ids it doesn't recognize but accepts them, so a brand-new release can be used the day your provider ships it.

Practical guidance: for a high-volume Q&A or concierge bot, the mini/haiku/flash tiers are usually indistinguishable from flagships and an order of magnitude cheaper; reach for gpt-5 or Claude Opus when the bot needs to reason through hard, multi-step questions. You can switch model or provider any time from the dashboard — the container rebuilds in seconds. Just note the key must match the provider: the save-time verification rejects, say, a Claude model on an OpenAI key.

Feed it your docs: the built-in knowledge base (RAG)

Step 3 of the AI Brain turns the bot from a generic chatbot into one that answers from your material. Three source types are supported: pasted text (up to 50,000 characters per source), website URLs (TeleCrow fetches the page server-side and extracts the text), and PDF uploads up to 15 MB — the PDF is parsed in memory and then discarded, with only the extracted text kept. Content is chunked into 2,000-character pieces with 200 characters of overlap, capped at 400 stored chunks per source, and up to 150 chunks are baked into the running bot.

Retrieval quality depends on your provider, and TeleCrow is upfront about it. With an OpenAI key (text-embedding-3-small) or a Google key (text-embedding-004), the bot uses Smart search — semantic vector retrieval that matches meaning, not just words. Anthropic and OpenRouter don't offer embeddings APIs, so those keys transparently fall back to keyword search; the UI explains this and suggests OpenAI or Google if Smart search matters to you. Switch providers and stale chunks are re-embedded automatically.

This is retrieval, not training: your documents are looked up per question, and no model is fine-tuned on your data. One nice extra — the same knowledge base is also available on TeleCrow's Support bot type, so a support bot can deflect FAQs from the same docs.

What it honestly doesn't do

The Personal AI bot is deliberately a focused tool: a single-model, text-in/text-out assistant living inside Telegram. It answers messages through your chosen model with your prompt, memory, and knowledge — and that's the scope. If your project needs the items below, this isn't the right fit yet, and it's better you know before you build.

  • No voice — no speech-to-text or text-to-speech; and no image generation or image understanding. Text in, text out.
  • No tool-calling, multi-agent workflows, or autonomous actions — one message in, one completion out.
  • No fine-tuning or training on your data — the knowledge base is retrieval (RAG) only, with hard caps (50k chars per text source, 15 MB per PDF, 150 chunks live in the bot).
  • Telegram-only — there's no website chat widget or web embed, and no public API for driving the bot programmatically.
  • PDF is the only file upload format — no DOCX, CSV, or TXT files; paste the text or add a URL instead.
  • It's a 1:1 personal assistant template with no commerce — no products, orders, or checkout on this bot type — and no conversation-transcript viewer in the dashboard.

Frequently asked questions

Is my OpenAI or Anthropic API key safe with TeleCrow?

The key is encrypted at rest with AES-256-GCM in a dedicated column and is never returned by any API — the dashboard only ever receives a "key stored" boolean and shows a masked placeholder. It's live-verified against your provider once before saving, and submitting an empty key clears it and disables the bot immediately. TeleCrow never supplies or resells keys; usage bills straight to your provider account.

Which AI models can I connect to Telegram this way?

OpenAI (including gpt-5, gpt-5-mini, gpt-4.1, gpt-4o, o3, o4-mini), Anthropic Claude (Opus 4.8, Opus 4.7, Sonnet 4.6, Haiku 4.5), Google Gemini (2.5 Pro/Flash down to 1.5), or an OpenRouter key for Llama, DeepSeek, Mistral, Qwen, and more. A "Custom model" option also accepts any model id your provider serves, so new releases work immediately.

How much does a ChatGPT or Claude Telegram bot cost to run?

TeleCrow adds no per-message AI billing — you pay your provider directly. With budget models (gpt-5-mini, claude-haiku-4-5, gemini-2.5-flash) a typical reply costs a small fraction of a cent, so light use is pennies per month. TeleCrow's side is just the plan: free runs 1 bot with "Powered by" branding, Basic $3.99, Pro $6.99 (2 bots), Ultra $11.99 (4 bots).

Do I need to know how to code?

No. The whole flow is paste-and-click: a two-step wizard (pick the Personal AI template, paste your BotFather token — with an in-app walkthrough if you've never made a bot), then a three-step AI Brain panel for provider, key, system prompt, and knowledge. TeleCrow hosts the bot in its own isolated container and rebuilds it automatically on every save — no webhooks, servers, or scripts.

Can the bot answer questions from my own documents or website?

Yes. Add pasted text (up to 50,000 characters per source), website URLs, or PDF uploads (up to 15 MB) to the knowledge base and the bot answers from them. With an OpenAI or Google key it uses Smart semantic search; Anthropic and OpenRouter keys fall back to keyword search because those providers don't offer embeddings APIs. It's retrieval only — no model is trained on your data.

What's the difference vs just creating a bot with @BotFather?

@BotFather only registers the bot and hands you a token — the bot does nothing until software answers its messages. Connecting that token to ChatGPT or Claude yourself means writing and hosting that software. TeleCrow is the missing half: paste the token, add your AI key, and it runs the bot for you, with the prompt, memory, model choice, and knowledge base managed from a dashboard.

Can it generate images, handle voice messages, or take actions?

No. The Personal AI bot is strictly text-in, text-out: no image generation or vision, no speech-to-text or text-to-speech, and no tool-calling or autonomous actions — each message gets one completion from your chosen model. If you need those capabilities you'd have to build a custom bot; TeleCrow's version is optimized for being a fast, cheap, no-code chat assistant.

Build, host and monetize a Telegram bot with TeleCrow

Sell digital products, gate paid channels and accept crypto — no code, on the free plan.

Start free on TeleCrow