DEV Community: Dmytro Bilukha

Real-World Agentic E-Commerce: Rebuilding My Grocery Cart with Claude and MCP

Dmytro Bilukha — Thu, 24 Sep 2026 08:03:50 +0000

Most agentic AI demos today focus on synthetic benchmarks, toy calculators, or coding assistants editing small functions.

What does consumer AI automation look like when plugged into real production systems?

I built silpo-agent-cli, a terminal CLI and Claude Code agent skill that interfaces with the official Model Context Protocol (MCP) server of Silpo, one of Ukraine's largest supermarket networks.

The Problem It Solves

Every week, grocery shopping involves 80% repeat purchases: coffee, milk, bread, produce, water. Re-adding these items in a web or mobile GUI wastes 15–20 minutes of repetitive tapping.

By exposing grocery basket operations to an agent via MCP, we can automate:

Re-aggregating regular cadence purchases.
Checking for active sales and discounts on staple goods.
Batch updating the real remote cart.

Safety & Guardrails By Design

Automating purchases must balance speed with financial safety. silpo-agent-cli enforces a strict guardrail:

The CLI stops at "Cart is Ready".

It fills your authentic shopping basket, verifies stock, and prints the summary. The final checkout step, delivery address selection, and payment card confirmation remain strictly in the human user's hands.

Quick Installation

pip install silpo-agent-cli
# or via uvx without installation:
uvx silpo-agent-cli reorder

AI Agent Skill Integration

Point Claude Code at the bundled skill definition:

"Reorder my usual groceries, replace oat milk with brand on sale if available, keep budget under 1500 UAH."

Claude inspects history, balances costs, and fills your basket.

GitHub Repository — github.com/MIt9/silpo-agent-cli
PyPI Package — silpo-agent-cli
Video Walkthrough — YouTube

Stream High-Resolution Stock Video Directly Into AI Agent Pipelines with pexels-cli

Dmytro Bilukha — Thu, 24 Sep 2026 08:01:15 +0000

Automated media production pipelines often require sourcing hundreds of relevant B-roll clips and stock images. Traditional SDKs return bloated JSON payloads containing hundreds of metadata keys that waste context windows when fed into LLMs or classifiers.

pexels-cli (px) is an open-source, high-performance CLI tool for Pexels built with uv and Python 3.10+.

Key Innovations for AI Pipelines

Classifier-Ready --state Mode: Emits condensed, single-line JSONL candidates specifically structured for non-generative classification engines (e.g. Laya).
Multi-Query Deduplication: Execute multiple search terms in parallel while automatically discarding duplicate video/photo IDs across result sets.
Payload Pruning (--fields): Whitelist only necessary keys (id, url, duration, width, height) to maximize token efficiency.
Asynchronous Parallel Downloader: Efficient streaming downloader with Rich terminal bars.

Quick Start

uv tool install pexels-ai-cli
# or
pipx install pexels-ai-cli

Example: Pipeline Chaining

# Search, deduplicate, format for decision model, and pipe
px videos --queries "tokyo night,shibuya crossing" --per-page 5 --state --dedupe \
  | laya-cli classify --questions-inline '{"relevant": {"type": "noul", "prompt": "Contains neon lighting?"}}' \
  | jq -r 'select(.predictions.relevant > 0.8) | .state'

GitHub Repository — github.com/MIt9/pexels-cli

Connect Claude to Perplexity AI Pro with Zero Search API Fees

Dmytro Bilukha — Thu, 24 Sep 2026 07:57:58 +0000

Modern LLMs struggle with recent software releases, live news, and undocumented breaking changes. While solutions like Tavily, Serper, or Google Custom Search exist, they introduce recurring per-call fees and lack the deep multi-source synthesis of specialized search engines.

If you already have a Perplexity Pro account, you can bridge it directly into Claude Desktop or Claude Code.

I built perplexity-pro-mcp — a lightweight MCP server that queries Perplexity Pro via your session credentials.

What it Provides

Tool	Capability
`perplexity_search`	Real-time web querying with synthesized answers, sources, and citations
`perplexity_follow_up`	Contextual multi-turn question answering inside the same research thread
`perplexity_threads`	Manage and resume prior search sessions
`perplexity_export`	Export structured research reports to Markdown/PDF

It also exposes Perplexity's built-in focus modes:

internet: Broad web search
academic: ArXiv, PubMed, and peer-reviewed journals
reddit: Community sentiment, troubleshooting discussions, and raw developer advice
youtube: Video transcripts and summaries

2-Minute Setup

1. Retrieve Session Cookie

Log into perplexity.ai, open DevTools (F12), head to the Network tab, click any request to perplexity.ai, and copy your full Cookie header string.

2. Configure Claude Code

claude mcp add perplexity -e PERPLEXITY_COOKIE="<YOUR_COOKIE>" -- npx -y perplexity-pro-mcp

Or Claude Desktop (`claude_desktop_config.json`)

{
  "mcpServers": {
    "perplexity": {
      "command": "npx",
      "args": ["-y", "perplexity-pro-mcp"],
      "env": {
        "PERPLEXITY_COOKIE": "<YOUR_COOKIE>"
      }
    }
  }
}

Example Interaction

Prompt to Claude: "What are the latest breaking changes in Next.js 15 canary releases and how do I adapt my Server Actions?"

Claude executes perplexity_search(query="Next.js 15 canary breaking changes server actions", mode="copilot"). Instead of hallucinatory advice, Claude returns verified code snippets accompanied by direct source citations.

Check out the repo on GitHub: github.com/MIt9/perplexity-pro-mcp

Manage ElevenLabs ElevenReader From Claude Desktop and Cursor via MCP

Dmytro Bilukha — Thu, 24 Sep 2026 07:55:30 +0000

ElevenLabs' ElevenReader is one of the most natural-sounding text-to-speech platforms for converting documents, research PDFs, and EPUBs into spoken audio.

However, moving documents from research workflows or local code repositories into ElevenReader traditionally requires manual browser drag-and-drops.

I built mcp-elevenreader, an open-source MCP server that provides full conversational control over your ElevenReader library.

Features

Document Ingestion — Upload PDFs, EPUBs, or plain text from your workspace directly to ElevenReader.
Library Tracking — Query reading progress, bookmarks, and completed reads.
Voice Selection & Customization — Configure narrator voices and collections via chat.
Zero-Install Execution (Powered by uvx) no virtualenv setup required.

Setup Instructions

1. Retrieve Refresh Token

Log into elevenreader.io, open DevTools (F12), switch to Console, and run:

JSON.parse(localStorage.getItem(Object.keys(localStorage).find(k => k.startsWith('firebase:authUser:')))).stsTokenManager.refreshToken

2. Configure Claude Code

claude mcp add elevenreader -e ELEVEN_REFRESH_TOKEN="<your-token>" -- uvx mcp-elevenreader

Or Claude Desktop (`claude_desktop_config.json`)

{
  "mcpServers": {
    "elevenreader": {
      "command": "uvx",
      "args": ["mcp-elevenreader"],
      "env": {
        "ELEVEN_REFRESH_TOKEN": "<your-token>"
      }
    }
  }
}

Sample Prompt

"Claude, upload the file ./papers/deepseek-r1.pdf to my ElevenReader library and let me know when it's ready for listening."

Check out the code on GitHub: github.com/MIt9/mcp-elevenreader

Turn Claude into an Autonomous AI Librarian with mcp-book-lover

Dmytro Bilukha — Thu, 24 Sep 2026 07:52:40 +0000

Managing an extensive digital library often means context switching between Goodreads for tracking, Calibre for conversion, and various web repositories for discovery.

mcp-book-lover is an open-source Model Context Protocol server that unifies library management, multi-source book search, format conversion, and recommendation engines inside Claude or Cursor.

Core Architecture & Capabilities

With over 35 tools, mcp-book-lover handles the full reading lifecycle:

Multi-Source Discovery: Concurrent searching across Open Library, Google Books, Project Gutenberg, and regional digital archives.
Lossless Conversion: Handles .epub, .fb2, .pdf, .txt, and compressed archives (.fb2.zip), preserving table of contents, inline graphics, and metadata.
Series & Cycle Awareness: Understands chronological reading orders and tracks volume progression.
Reading Analytics: Tracks annual reading challenges, average ratings, and personalized recommendations.

Quick Start

No API keys required. Configure Claude Code or Claude Desktop with uvx:

Claude Code

claude mcp add book-lover -- uvx mcp-book-lover

Claude Desktop (`claude_desktop_config.json`)

{
  "mcpServers": {
    "book-lover": {
      "command": "uvx",
      "args": ["mcp-book-lover"]
    }
  }
}

Example Interaction

Prompt: "What should I read after finishing the Three-Body Problem trilogy? Search Open Library for hard sci-fi recommendations and show available formats."

Claude queries your review history, searches candidate titles, checks format compatibility, and organizes recommendations directly inside your chat.

GitHub: github.com/MIt9/mcp-book-lover

Sub-35ms Typed AI Decisions Without Token Generation Or Hallucinations

Dmytro Bilukha — Thu, 24 Sep 2026 07:49:20 +0000

Modern AI pipelines often burn compute using 8B+ parameter generative models just to answer questions like:

"Is this support ticket urgent?"
"Does this comment violate moderation policies?"
"Should this request route to the billing or tech support department?"

Autoregressive generation for classification introduces high latency (waiting for token-by-token decoding), unpredictable JSON formatting errors, and unnecessary costs.

laya-cli is a high-performance CLI and daemon wrapper around Laya — providing typed decisions (choice, score, noul) in a single forward pass without autoregressive token generation.

Key Capabilities

Deterministic Typed Outputs in One Pass:
- choice: Discrete labels with calibrated probabilities.
- score: Ordinal expected values (e.g. priority 1 to 5).
- noul: Probability that an assertion is true.
Speed: ~33ms for single predictions, ~7ms per item when batched.
Resident Daemon: Run laya-cli serve to keep weights resident in memory, eliminating cold starts for subsequent CLI or sub-process calls.
Unix Pipeline Compatibility: Native streaming support for JSON and JSONL feeds.

Quick Installation

uv tool install laya-cli
# or: pip install laya-cli

Usage Examples

Interactive Categorization

laya-cli predict "User reports database timeout after running migrations" --preset triage

Streaming Pipeline

cat incoming_events.jsonl | laya-cli classify --preset moderation --format jsonl >> audited.jsonl

Links & Community

GitHub — github.com/MIt9/laya-cli
PyPI — pypi.org/project/laya-cli

Build an Automated Multi-Modal Media Pipeline with One Zero-Dependency CLI

Dmytro Bilukha — Thu, 24 Sep 2026 07:45:59 +0000

The modern generative media ecosystem is intensely fragmented. If an autonomous agent or background worker needs to create a complete promotional clip, it typically needs:

An image generator (Flux, GPT Image, Midjourney API)
An image-to-video / text-to-video model (Seedance, Wan, Kling)
A voiceover or sound synthesizer (ElevenLabs, Suno)

Handling multiple API keys, diverse request structures, and inconsistent polling loops creates enormous boilerplate.

I built kie-media-cli to solve this: a unified, zero-dependency Node.js CLI tool for the KIE API (kie.ai).

Core Principles

Zero Runtime Dependencies: Written in clean, modern Node.js using native fetch and standard library modules. No dependency vulnerabilities or slow install steps.
Unified Ergonomics: Consistent commands whether you are rendering an image, generating a video, or tracking job status.
First-Class Agent Support: Pairs with Claude Code skills, letting AI coding agents invoke CLI commands reliably with strict JSON parsing.

Usage & Setup

Interactive Configuration

npx -y kie-media-cli setup

The wizard guides you through setting your API key, checking your current credit balance, and securing your config in ~/.kie-media/config.json.

Generating Media

# Generate image
kie generate image --prompt "Cyberpunk landscape, cinematic" --output render.png

# Generate video
kie generate video --prompt "Slow pan across mountain peaks" --output scene.mp4

# Check credit usage & job logs
kie history --limit 5

Agent Integration with Claude Code

To let Claude generate multimedia assets autonomously:

npx -y skills add MIt9/kie-skills/kie-generate

Claude can now inspect model options, format prompts, monitor render status, and integrate outputs into your repositories.

GitHub Repository — github.com/MIt9/kie-media-cli
NPM Package — kie-media-cli

Generate 3D Meshes (GLB, OBJ, STL) Directly From the Terminal and AI Agents

Dmytro Bilukha — Thu, 24 Sep 2026 07:42:41 +0000

Integrating generative 3D meshes into web applications (Three.js, Babylon.js) or game engines usually involves manual asset downloads from web dashboards.

hi3d-cli (hi3d) is a lightweight, zero-dependency CLI tool and agent skill for generating textured 3D models via the Hi3D API.

Supported Output Formats

Web & AR — .glb, .usdz
Game Development — .obj, .fbx
3D Printing — .stl, .3mf
Render Maps — .exr, .png, .bmp

30-Second Setup

npx -y hi3d-cli setup

The interactive wizard configures your Access & Secret keys from platform.hi3d.ai into ~/.hi3d/config.json.

Generating a 3D Mesh

# Generate 3D model from image reference
hi3d generate --image ./concept.png --format glb --output ./public/models/character.glb

Agent Integration

Pair with Claude Code via npx -y skills add MIt9/hi3d-skills/hi3d-generate to enable automated 3D modeling directly in your dev workflow.

GitHub Repository — github.com/MIt9/hi3d-cli
NPM — hi3d-cli

How I Built an MCP Server That Lets Claude Control ComfyUI

Dmytro Bilukha — Thu, 24 Sep 2026 07:34:19 +0000

ComfyUI is arguably the most powerful node-based runtime for generative diffusion models (SDXL, Flux, Stable Video Diffusion). However, working with ComfyUI usually involves either:

Dragging virtual cables across hundreds of spaghetti nodes in a browser canvas, or
Manually crafting fragile, deeply nested JSON workflow payloads.

With Anthropic’s Model Context Protocol (MCP), there is a better way: treat workflow construction as an agentic planning problem.

I built and open-sourced mcp-comfy-ui-builder. Here is how it works under the hood.

The Architecture

To make an LLM like Claude reliably build ComfyUI workflows without hallucinating invalid connections or crashing the GPU, the MCP server provides 3 core subsystems:

1. Embedded Knowledge Base & Pre-Flight Checks

Instead of blindly sending prompts, the agent first runs get_system_resources. This tool reads local GPU VRAM and memory, returning safe boundaries (max resolution, batch sizes).

The server bundles knowledge/base-nodes.json (62+ default nodes and compatibility rules). For custom nodes, npm run sync-nodes discovers and indexes them dynamically from your live runtime.

2. Declarative Workflow Construction

Rather than writing raw graph JSON, the model invokes high-level semantic tools:

build_workflow with pre-tuned templates (txt2img, txt2img_flux, img2img, inpainting, upscale)
add_node and connect_nodes for dynamic alterations
validate_workflow to ensure all type slots (MODEL, CLIP, VAE, LATENT) match before queueing

3. Real-Time WebSocket Execution

Standard HTTP polling against /prompt and /history adds 1–2 seconds of latency.

mcp-comfy-ui-builder maintains a persistent WebSocket connection to ComfyUI. During generation, the server receives sub-100ms progress events, showing node-by-node updates in Claude's UI before downloading outputs via list_outputs.

Quick Setup

1. Install & Build

npm install mcp-comfy-ui-builder
npm run build

2. Configure Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "comfy-ui-builder": {
      "command": "node",
      "args": ["/ABSOLUTE/PATH/TO/mcp-comfy-ui-builder/dist/mcp-server.js"],
      "env": {
        "COMFYUI_HOST": "http://127.0.0.1:8188"
      }
    }
  }
}

Now open Claude Desktop and say:

"Generate a futuristic cyberpunk skyline at sunset with neon reflections, upscale 2x."

DEV Community: Dmytro Bilukha

Real-World Agentic E-Commerce: Rebuilding My Grocery Cart with Claude and MCP

The Problem It Solves

Safety & Guardrails By Design

Quick Installation

AI Agent Skill Integration

Stream High-Resolution Stock Video Directly Into AI Agent Pipelines with pexels-cli

Key Innovations for AI Pipelines

Quick Start

Example: Pipeline Chaining

Connect Claude to Perplexity AI Pro with Zero Search API Fees

What it Provides

2-Minute Setup

1. Retrieve Session Cookie

2. Configure Claude Code

Or Claude Desktop (claude_desktop_config.json)

Example Interaction

Manage ElevenLabs ElevenReader From Claude Desktop and Cursor via MCP

Features

Setup Instructions

1. Retrieve Refresh Token

2. Configure Claude Code

Or Claude Desktop (claude_desktop_config.json)

Sample Prompt

Turn Claude into an Autonomous AI Librarian with mcp-book-lover

Core Architecture & Capabilities

Quick Start

Claude Code

Claude Desktop (claude_desktop_config.json)

Example Interaction

Sub-35ms Typed AI Decisions Without Token Generation Or Hallucinations

Key Capabilities

Quick Installation

Usage Examples

Interactive Categorization

Streaming Pipeline

Links & Community

Build an Automated Multi-Modal Media Pipeline with One Zero-Dependency CLI

Core Principles

Usage & Setup

Interactive Configuration

Generating Media

Agent Integration with Claude Code

Generate 3D Meshes (GLB, OBJ, STL) Directly From the Terminal and AI Agents

Supported Output Formats

30-Second Setup

Generating a 3D Mesh

Agent Integration

How I Built an MCP Server That Lets Claude Control ComfyUI

The Architecture

1. Embedded Knowledge Base & Pre-Flight Checks

2. Declarative Workflow Construction

3. Real-Time WebSocket Execution

Quick Setup

1. Install & Build

2. Configure Claude Desktop

Links

Or Claude Desktop (`claude_desktop_config.json`)

Or Claude Desktop (`claude_desktop_config.json`)

Claude Desktop (`claude_desktop_config.json`)