🤖 The Best Telegram Game is now Agentic Friendly

I've updated the code of the Complete Crypto Telegram Mini App Game. I updated all the packages to the latest versions, fixed some minor bugs in the game component (components/game/SuikaGame.tsx), and - THE COOLEST FEATURE - added a special script to compact all the code into a single file you can feed directly to any AI.

Just run:
npx tsx ai-scripts/export-for-ai.ts full
npx tsx ai-scripts/export-for-ai.ts server
npx tsx ai-scripts/export-for-ai.ts frontend

It strips out license headers to save tokens, tags every file as 🔧 server / 🎨 frontend / 📦 shared, and gives you one clean text file you can paste straight into ChatGPT, Claude, or any other AI.

Why does this matter? Because now you can:
→ Paste the full codebase and ask an AI to add new features for you
→ Export only the backend, and have it build new API routes that fit your existing auth, rate limiting, and database schema
→ Export only the frontend, and have it create new pages and components that match your design system
→ Debug issues by giving the AI the exact context it needs - nothing more, nothing less

This turns a production-ready Telegram Mini App into your personal launchpad. You don't need to understand every line. Just tell the AI what you want, give it the right context, and ship.

The full source code is available on my website

➡️ https://nikandr.com/en/product/viral-telegram-app-mastery/preview

NIKANDR'S APPS | YOUTUBE

💪 BTW I can change the game in the best Telegram game just in one minute

The full source code of the Complete Crypto Telegram Mini App Game has a modular structure so you can easily unplug and plug a different game and everything will keep working perfectly. You literally need to change just one file - SuikaGame.tsx - to swap the entire game concept.

Everything else stays untouched: the TON & Telegram Stars payments, the referral system, the daily bonuses, the leaderboard, the anti-cheat, the rate limiting - all of it just works with your new game out of the box.

The game component receives only 3 things: sessionId, userId, and an onGameEnd(points) callback. Build any game you want - a clicker, a puzzle, a runner - as long as it calls onGameEnd with a score, the entire backend handles the rest.

That's the power of clean architecture.

Full source code → https://nikandr.com/en/product/viral-telegram-app-mastery/preview

NIKANDR'S APPS | YOUTUBE

😮 Demand is CRAZY. Am I missing something?!

Over the last 3 months I've received 100+ DMs - most of them from people asking me to help build Web3 loot box games on the TON blockchain.

I've worked on 6 projects so far, helping with features like USDT top-ups on TON, smart contract mechanics for loot boxes, withdraw flows, and more.

The wild part - the demand right now is even bigger than during the Hamster Kombat TON bull run. Am I missing something?!

Anyway, through all of this I've built a solid codebase (one that broke my Claude Max limits a few times and cost me $500 more than expected). I'm putting the final touches on the full source code right now.

So instead of pinging everyone who's DMed me over the last 3 months - I'll be releasing it soon.

Don't tell anyone. I don't want this to get too hyped 👀

NIKANDR'S APPS | YOUTUBE

🥚 I just vibe-coded the egg game from "I Think You Should Leave" and plugged it into my Telegram Mini App in 5 minutes

Saw the Feed Eggs bit from Tim Robinson's show, thought it was hilarious, and decided to turn it into a playable game inside my Telegram app.

Here's what I did:
- Asked AI to convert the HTML game into a React component
- Made sure it matches the existing game interface (sessionId, userId, onGameEnd callback)
- Swapped one file. Done.

Everything else kept working instantly:
→ TON & Telegram Stars payments
→ Referral system
→ Daily bonuses
→ Leaderboard
→ Anti-cheat & rate limiting

No backend changes. No config tweaks. No debugging. One file swap and the entire app now runs a completely different game.

That's not a hack - that's clean architecture doing its job.

Whether you want a Suika-style merge game, a clicker, a puzzle, or a ridiculous egg-feeding simulator from a sketch comedy show - the codebase doesn't care. Build your game, call onGameEnd(points), and everything else handles itself.

🍆You can play this Egg Game now → https://t.me/ViralTgGameBot/app

Full source code → https://nikandr.com/en/product/viral-telegram-app-mastery/preview

NIKANDR'S APPS | YOUTUBE

👔 I just built a Doodle Jump clone with Trump and plugged it into my Telegram Mini App

Took an open-source Doodle Jump game, asked AI to turn it into a React component, and swapped one file. Done.

And here's the thing - this plug-in game architecture is a goldmine for anyone who wants to capitalize on trending events. Trump does something wild? Swap in a Trump game. World Cup starts? Soccer clicker. New meme blows up? Build a game around it in an hour and plug it in. Your payments, referrals, leaderboard, and anti-cheat keep running like nothing changed.

One file swap. Same story as before:
→ TON & Telegram Stars payments
→ Referral system
→ Daily bonuses
→ Leaderboard
→ Anti-cheat & rate limiting

All untouched. All still working.

Three games now live in the same codebase - Suika merge game, Feed Eggs, and Trump Jump. Switching between them is one line change in one file. The game component just receives sessionId, userId, and calls onGameEnd(points). The backend doesn't know or care what game is running.

That's three completely different game genres - physics puzzle, clicker, and platformer - running on the same payments, referrals, and anti-cheat system without a single backend change.

👾 You can play this game now → https://t.me/ViralTgGameBot/app

Full source code → https://nikandr.com/en/product/viral-telegram-app-mastery/preview

NIKANDR'S APPS | YOUTUBE

🙄 Your landing page SUCKS if it doesn't have physics

In the AI era, anyone can generate a stunning landing page in minutes. Which means "stunning" is now the baseline - not the differentiator.

To actually stand out you need something people want to TOUCH.

So I built a landing page with a real-time cloth physics simulation - a security scan report you can grab, drag, and crumple. WebGL, Verlet integration, 1,250 particles, 15 constraint iterations per frame. All in a single HTML file.

The receipt isn't decoration. It IS the hook. People stop scrolling when something moves like it belongs in the real world.

Here's what powers it:
→ Canvas 2D generates the receipt texture
→ WebGL renders a 25×50 particle mesh
→ Verlet physics with structural, shear & bending constraints
→ Pointer raycasting for grab interaction
→ Wind simulation for subtle idle motion
→ Transparent compositing so it floats over the page

The whole landing page is ~400 lines of CSS + ~300 lines of JS. Zero dependencies. No Three.js. No libraries. Just raw WebGL.

I published the full code on CodePen: https://codepen.io/Nikandr-Surkov/pen/GgjjMNJ

Steal it. Fork it. Put physics in your landing page.

Your competitors won't.

NIKANDR'S APPS | YOUTUBE

⛔️ Nobody remembers a landing page they scrolled through

I put an entire SaaS landing page inside a macOS desktop simulation.

Each section opens as a draggable window from the dock. Boot screen, traffic lights, minimize, maximize, resize, terminal with easter eggs - all working.

No one scrolls past this. They play with it.

Single HTML file. No frameworks.

CodePen: https://codepen.io/editor/Nikandr-Surkov/pen/019cd880-4938-7aee-b77b-67f5cc108cc8

NIKANDR'S APPS | YOUTUBE

🤯 What if your landing page had liquid glass floating over it?

Metaball droplets with real physics — they merge, split, refract the text underneath, and respond to your cursor.

WebGL shader with Fresnel, chromatic aberration, soft shadows. Canvas 2D texture underneath. One HTML file.

Drag them around. Click to spawn more. Watch two droplets merge into one.

CodePen: https://codepen.io/editor/Nikandr-Surkov/pen/019cddc3-76f7-70a7-821c-3eec39238918

Landing pages should feel alive.

NIKANDR'S APPS | YOUTUBE

😮 I Ran a 3B AI Model in JavaScript — No Cloud, No Python, Just WebGPU

Something is shifting in AI and most people are missing it.

While everyone is focused on bigger models and more expensive GPU clusters, Microsoft went the opposite direction. They released BitNet b1.58 — an LLM where every weight is just -1, 0, or +1. Not compressed from a full-precision model. Trained from scratch with ternary weights baked into the architecture. The result: 0.4GB of memory instead of the usual 1.4–4.8GB, up to 6x faster inference on CPUs, and 82% lower energy consumption. Their framework can even run a 100B parameter model on a single CPU at human reading speed.

And it's not just Microsoft. TII released Falcon-Edge — their own native ternary models at 1B and 3B scale. Alibaba launched Qwen 3.5 Small, which takes a different approach with sparse MoE architecture — the 2B variant runs on an iPhone in airplane mode, no server needed. The trend is the same: serious AI leaving the cloud.

So I decided to test the limits. Could I run one of these models in the most unlikely runtime — JavaScript?

Turns out, yes. TII's Falcon-E 3B Instruct running locally via WebGPU compute shaders in Node.js. About 80 lines of code. A ~1GB model download from Hugging Face. After that — fully offline, fully private, zero cost per inference.

No Python. No CUDA. No Docker. No API key. No monthly bill.

Your documents, conversations, and data never leave your machine. That's not a feature toggle — it's the architecture.

Code: https://github.com/nikandr-surkov/bitnet-webgpu-chat.git

NIKANDR'S APPS | YOUTUBE

🦞 I made a 3D lobster walk across a landing page using only CSS

No WebGL. No canvas. No Three.js. Just divs.

24 voxel cubes, each with 6 CSS faces. Walking bob, claw snaps, antenna sway, alternating leg cycle. All keyframe animations.

color-mix() generates light/shadow faces from one color. One JS function stamps out every cube. The rest is pure CSS 3D transforms.

Full OpenClaw landing page built around it. One HTML file. Zero libraries.

CodePen: https://codepen.io/editor/Nikandr-Surkov/pen/019ce825-7c45-795e-a1bf-5edffc130ef5

CSS can do more than you think.

NIKANDR'S APPS | YOUTUBE

🚨 BREAKING: Make any AI understand an entire GitHub repo - with up to 95% fewer tokens

I built a free tool that converts any GitHub repository into a single AI-ready text file in seconds.

Just paste a repo URL → get the full file tree, source code, and a Symbol Index - all formatted for LLMs.

What is Symbol Index?

This is the killer feature. Instead of dumping raw code into OpenClaw/ChatGPT/Claude, GitExtract intelligently analyzes your code and extracts every function, class, method, constant, and type - across 15 languages.

You get a structured map of the entire codebase.

This alone can save you massive token usage. Feed the Symbol Index first → AI understands the architecture → then only feed the specific files it needs. No more pasting 200 files into context.

What you get:
📁 Directory tree
📝 Full source code (concatenated, formatted)
🧠 Symbol Index (functions, classes, methods, constants, types)
📊 Token count estimate

Supports 15 languages: Python, JavaScript, TypeScript, Go, Java, Rust, C, C++, C#, Ruby, PHP, Dart, Elixir, SQL, and TSX.

How to use:
Web UI — paste a URL at gitextract.com
JSON API — GET /api/v1/ingest?url=user/repo

No signup. No API key. No rate limit headaches (we clone repos directly, no GitHub API dependency).

It's completely free. I built this because I was tired of manually copying files into OpenClaw/Claude/ChatGPT context windows.

Try it → https://gitextract.com

NIKANDR'S APPS | YOUTUBE

⚡️ GitExtract Update: Structured JSON Export is here

You asked for it — now GitExtract returns every file as clean, structured JSON.

Instead of one massive text blob, you get this:

[
{"path": "src/app(.)py", "content": "..."},
{"path": "src/utils(.)py", "content": "..."}
]

Each file is a separate object. Your agent can loop through them, pick specific files, or process them one by one. No more parsing text walls.

Why this matters for AI agents:

Before: you had to regex-split the text dump to find individual files.
Now: just iterate over a JSON array. Every file has its path and content ready to go.

This is huge for anyone building AI pipelines, RAG systems, or coding agents that need to process repos file-by-file.

How to get it:

API → GET /api/v1/ingest?url=user/repo&include_json=true
Web → new "Structured JSON" tab on every result page
Download → one click, get the full .json file

Still free. Still no signup. Still no API key.

Try it → https://gitextract.com

NIKANDR'S APPS | YOUTUBE

😰 I canceled my ChatGPT subscription

I still generate images, more than ever. I just stopped paying $20/month for them.

GPT Image 2 (released April 21) runs on every ChatGPT plan AND in the OpenAI API. Same model, two ways to access it.

What $20/month for ChatGPT Plus buys you:
- ~50 image prompts per 3-hour rolling window
- Each image counts against the same 160-message cap as your text chats
- "Thinking mode": slower planning pipeline, Plus-only, genuinely useful for layout-heavy work
- The full bill, even on months you barely use it

What I built instead: a Telegram Mini App on top of GPT Image 2 (via Replicate or your own OpenAI key). Pay-per-image, the same per-image rate OpenAI bills the API:

- Low, 1:1 → $0.006 ($20 = ~3,300 images)
- Medium, 1:1 → $0.053 ($20 = ~370 images)
- High, 1:1 → $0.211 ($20 = ~95 images)

To even hit $20 here you'd have to burn through ~370 medium-quality images in a month. I rarely cross 50. No daily caps, no 3-hour throttling, no peak-time slowdowns, no message budget shared with text chats.

And the controls are nicer than ChatGPT's:
- Aspect ratio picker, quality picker with live $/image before you click
- Up to 8 reference images for image-to-image, addressed in the prompt as ref1...refN for style mixing, character continuity, edits
- Reusable prompt presets, synced across devices via Telegram CloudStorage
- Every result auto-archived to your bot chat. Full restore via "Use prompt"
- Owner-locked via HMAC + your Telegram user id, safe to host publicly

Open-source, self-hosted. Bring your own Replicate or OpenAI key.

Repo → https://github.com/nikandr-surkov/gpt-image-2-telegram-mini-app

NIKANDR'S APPS | YOUTUBE

🪙 Tolk NFT Blueprint released

A complete, production-ready NFT implementation on TON, written entirely in the new Tolk language.

This is what I wish I'd had when I started learning Tolk: a clone-and-ship reference repo, not another half-finished demo.

What's inside:

- NFT Collection + Item contracts in Tolk 1.2
- TypeScript wrappers with every Tolk quirk handled (address?, RemainingBitsAndRefs, dict encoding)
- 9 Sandbox + Jest tests - deploy, mint, batch, transfer, admin
- Blueprint scripts - TonConnect & mnemonic deploy / mint
- Pinata IPFS metadata recipe with the chicken-and-egg CID workaround
- TEP-62 / TEP-64 / TEP-66 compliant
- Verified on TON mainnet (deployed live)

➡️ Repo: github.com/nikandr-surkov/tolk-nft-blueprint

If it saves you time - drop a ⭐️ on GitHub.

NIKANDR'S APPS | YOUTUBE

💱 Toncoin is now Gram

Pavel Durov announced the rebrand on June 1, reviving the original 2018 name from Telegram's first whitepaper. The transition runs ~3 weeks, with no token swap required.

What changes for builders:

🏷 Token ticker: TON → GRAM. UI labels, payment displays, and exchange listings will roll over during the transition window.
🔗 The TON blockchain name stays the same. Contracts, addresses, RPC endpoints, SDKs — unchanged. This is a token-label change, not a chain migration.
⚡️ Telegram has also taken over as the network's largest validator (~2.2M staked) and cut fees roughly 6x. Cheaper on-chain ops for anyone routing payments or NFT mints through a mini app.

Action item this week: grep your mini app and bot UI for "Toncoin" and ticker references, and plan a copy swap to "Gram" / "GRAM" over the next few weeks. The chain stays TON; only the token name moves.

NIKANDR'S APPS | YOUTUBE

🧠 Claude Fable 5 is here

Anthropic released Claude Fable 5 yesterday - the first publicly available model from the new Mythos tier that sits above Opus. It's positioned for long-running, autonomous coding sessions where previous models lost the thread.

What stands out:

🟢 80.3% on SWE-Bench Pro (Opus 4.8: 69.2%, GPT-5.5: 58.6%, Gemini 3.1 Pro: 54.2%). On Cognition's FrontierCode it hits 29.3% vs Opus 4.8 at 13.4%.
🟢 Built for end-to-end work: planning, tool use, multi-hour runs, self-repair. Better than Sonnet or Opus at owning a whole feature from a single prompt.
🟢 $10/M input, $50/M output. Free in Claude apps until June 22, then paid usage credits.
🟢 Has cybersec/biology guardrails that route some requests to Opus 4.8 instead.

For us building Telegram bots and TON mini apps, the practical upshot: Fable 5 is now the strongest model for refactors, security audits, and "go through my whole repo and fix X" work. Next post: a ready-to-use prompt to audit your project with it.

NIKANDR'S APPS | YOUTUBE