DEV Community: sonpiaz

Free AI models just got really good. Here's how to use 23 of them with one API key.

sonpiaz — Sun, 05 Apr 2026 07:16:06 +0000

This week three free models dropped that compete with paid ones:

Qwen 3.6 Plus scores 61.6 on Terminal-Bench. Claude 4.5 Opus scores 59.3. Qwen is free. 1M context window.
Gemma 4 31B is Google's first Apache 2.0 model. It understands images and text.
GPT-OSS 120B is OpenAI's first open source model.

A year ago you'd pay $20-100/month for this quality. Now it's free.

The catch: each model needs its own account. Want to try all three? That's three signups, three API keys, three different ways to call them.

One key for everything

I found Kyma API. You sign up once, get one key, and use it for 23 models.

Here's what the code looks like:

from openai import OpenAI

client = OpenAI(
    base_url="https://kymaapi.com/v1",
    api_key="ky-your-key"
)

# Use Qwen 3.6 Plus
response = client.chat.completions.create(
    model="qwen-3.6-plus",
    messages=[{"role": "user", "content": "Review this code"}]
)

# Want Gemma 4 instead? Change one word.
response = client.chat.completions.create(
    model="gemma-4-31b",
    messages=[{"role": "user", "content": "What's in this image?"}]
)

If your code works with OpenAI, it works with this. Same format. No new library.

Which model for what

What you want to do	Use this	Why
Write code	Qwen 3 32B or Kimi K2	Top coding benchmarks
Analyze a long document	Gemini 2.5 Flash	1M token context
General questions	Llama 3.3 70B	Best all-rounder
Get the smartest answer	Qwen 3 235B	Largest free model available
Get the fastest answer	Llama 3.1 8B	About 100ms
Understand images	Gemma 4 31B	Multimodal
Try the latest trending	Qwen 3.6 Plus	Just released this week

How it works in practice

I signed up. Took about 30 seconds. Got a key starting with ky-.

Tested a few models. Switching between them is changing one word. Response times were fast, around 100-300ms for most models.

The free tier has rate limits but it's enough for side projects and testing. If you need more there's a pay-as-you-go option priced close to cost.

hidrix-tools: 18 MCP tools that give AI agents web search, social media scraping, and content analysis

sonpiaz — Sat, 04 Apr 2026 20:48:55 +0000

I open-sourced hidrix-tools yesterday — an MCP server that gives AI coding agents the ability to search the web, scrape social media, and analyze content.

The problem

If you use Claude Code, Cursor, pi, or any CLI agent — they can read files and write code. But ask them to search X/Twitter, pull Reddit threads, or scrape a Facebook group? They can't.

I kept copy-pasting research into conversations. It was slow and broke the flow.

What I built

An MCP tool server with 18 tools. Plug it into any agent that supports MCP — no vendor lock-in.

Your AI Agent                    hidrix-tools                     Internet
┌──────────┐     MCP protocol    ┌──────────────┐                ┌─────────┐
│ Claude   │ ◄─────────────────► │ web_search   │ ◄────────────► │ Brave   │
│ Cursor   │                     │ x_search     │                │ X/Twitter│
│ pi       │                     │ reddit_search│                │ Reddit  │
│ Codex    │                     │ facebook_    │                │ Facebook│
│ ...      │                     │   scraper    │                │ YouTube │
└──────────┘                     └──────────────┘                └─────────┘

Tools included

Search: web_search (Brave), x_search, x_thread_reader, x_user_posts, reddit_search, reddit_thread_reader, reddit_subreddit_top, youtube_search, tiktok_search

Scrape: web_fetch (URL → markdown), facebook_scraper (groups, pages, keyword search, Meta Ad Library)

LinkedIn: linkedin_search, linkedin_profile (no login needed)

Analyze: content_scorer (engagement ranking + time-decay), content_analyzer (topic clusters, patterns, trends)

Storage: SQLite — dedup, run history, query saved posts

What I shipped in the last 24 hours

11 commits since the initial release:

LinkedIn tools — search posts and profiles without login
Facebook scraper pipeline — groups, pages, search, Meta Ads
YouTube transcripts + channel tools
SQLite storage with dedup
Follow system — track people/keywords across platforms
Knowledge compiler — auto-builds an Obsidian wiki from collected data
Business intelligence — competitor tracking, market signals

The feature I use most: Knowledge compiler

My agent follows 10+ people on X. Every day it:

Collects their new posts
Stores them in SQLite (dedup)
Extracts key insights
Compiles wiki articles in my Obsidian vault

I wake up to an updated knowledge base every morning. Zero manual work.

Not a demo — just a cron job + MCP tools.

Install (3 commands)

git clone https://github.com/sonpiaz/hidrix-tools.git ~/.hidrix-tools
cd ~/.hidrix-tools && bun install && cp .env.example .env
# Add your API keys to .env

Connect to Claude Code

{
  "mcpServers": {
    "hidrix-tools": {
      "command": "bun",
      "args": ["run", "~/.hidrix-tools/server.ts"]
    }
  }
}

Works with Claude Code, Cursor, OpenClaw, pi, Codex, Hermes — any MCP client.

Adding your own tool

Tools are auto-discovered. Copy the template, edit, restart:

cp -r tools/_template tools/my-tool
# Edit tools/my-tool/index.ts
bun run server.ts
# [hidrix-tools] ✓ my_tool

I built 50 AI skills for affiliate marketing — here is the social intelligence layer

sonpiaz — Fri, 03 Apr 2026 21:27:15 +0000

The Problem

Most AI-written affiliate content is generic. The AI picks a format by default, writes from training data, and hopes it converts. No engagement signals. No real sources. No data on what's actually working.

I wanted to fix that.

What I Built

affiliate-skills is a collection of 50 SKILL.md files — each one tells any AI exactly how to execute an affiliate marketing task. Input/output schemas, step-by-step workflows, error handling, and chaining metadata.

The v1.2 update adds a social intelligence layer: 5 new skills that bring real engagement data into the content creation pipeline.

The 5 New Skills

1. trending-content-scout (S1 Research)

Scans YouTube, TikTok, X, and Reddit for top-performing content by engagement score.

What it returns:

Top 20 content sorted by engagement
Winning formats (comparison: 45%, tutorial: 25%, review: 20%)
Best hooks ("I replaced my $5K video team" → engagement: 42.3)
Content gaps (nobody comparing HeyGen vs Synthesia on TikTok)
Engagement benchmark (median views, top 10% threshold)

Engagement Score Formula:

engagement_score = (likes × 2 + comments × 3 + shares × 5) / max(views, 1) × 1000

Shares weighted 5x (strongest viral signal). Comments 3x (high-intent). Likes 2x (low-effort positive).

2. content-angle-ranker (S1 Research)

Generates 8-12 content angle candidates and scores each on 4 dimensions:

angle_score = (platform_fit × 0.25) + (competition_level × 0.30) + 
              (engagement_prediction × 0.30) + (creator_fit × 0.15)

Output: ranked list with a clear #1 recommendation and ready-to-use parameters for the next skill.

3. traffic-analyzer (S1 Research)

Evaluates any website's traffic health before you commit to promoting their program. Calculates a Traffic Health Score 0-100, compares multiple domains side-by-side.

4. content-research-brief (S2 Content)

Collects 5-10 real source articles, auto-tags them (AI, Funding, SaaS, Tools, Trends), extracts key stats and quotes, and generates 3+ unique content angles — each backed by a different primary source.

5. infographic-generator (S2 Content)

Generates complete infographic specs: layout, data points, copy, color scheme. Platform-optimized (LinkedIn 1080×1350, IG square, Twitter landscape). Output as structured spec or self-contained HTML/CSS.

The Flywheel

Every skill knows what comes next. Data flows forward through the funnel and back through analytics:

S1 RESEARCH → S2 CONTENT → S3 BLOG & SEO → S4 OFFERS & LANDING
     ↑                                              ↓
     │                                        S5 DISTRIBUTION
     │                                              ↓
     └──────── S6 ANALYTICS ◀────────────────────────┘
                    ↓
              S7 AUTOMATION → SCALE

Closed loop. S6 feeds back to S1. Your performance data refines the next scout run.

Data-driven. By the time you write, you already know what format, hook, and platform to use.

Research-backed. Content skills don't write from thin air. content-research-brief collects real articles first.

API-Optional Design

All skills work with just web_search and web_fetch. No API keys required.

But if you have APIs, the data gets better:

social_data_config:
  youtube:
    provider: "youtube-data-api"
    api_key: "AIzaSy..."
  tiktok:
    provider: "rapidapi"
    api_key: "YOUR_KEY"
    host: "tiktok-api23.p.rapidapi.com"

Supported: YouTube Data API v3, RapidAPI, SerpAPI, Apify, or any custom API with field mapping.

Feedback Protocol

When a skill underperforms, it auto-generates a skill_feedback block:

skill_feedback:
  skill: "trending-content-scout"
  status: "partial_failure"
  issue_type: "data_quality"
  step_failed: "Step 2"
  description: "TikTok returned 0 results"
  severity: "low"
  suggestion: "Add alternative search queries"

9 auto-detection triggers. Issue taxonomy: data_quality, hallucination, chain_break, schema_mismatch, etc.

Try It

Any AI, no install:

Paste into Claude, ChatGPT, or Gemini:

"Scout trending content about AI video tools on YouTube and TikTok. 
Show me top content by engagement, winning formats, and content gaps."

Claude Code / Pi:

git clone https://github.com/Affitor/affiliate-skills.git ~/.claude/skills/affiliate-skills
cd ~/.claude/skills/affiliate-skills && ./setup

Full repo: github.com/Affitor/affiliate-skills

50 skills. 8 stages. MIT license. Works with any AI.

What does your content research workflow look like? Would love to hear how others approach data-driven content creation.

I Built 45 AI Agent Skills for Affiliate Marketing — Here's How They Work

sonpiaz — Mon, 16 Mar 2026 04:56:22 +0000

The Problem

AI is great at writing copy. But ask it to "find the best affiliate program for AI video tools" and you get generic, outdated advice from its training data.

What if your AI could query a real-time database, compare commission rates, calculate potential earnings, then write optimized content — all in one conversation?

What I Built

affiliate-skills — 45 open-source AI agent skills that cover the entire affiliate marketing funnel:

Stage	Skills	What they do
S1 Research	6	Find and evaluate affiliate programs
S2 Content	5	Write viral social media posts
S3 Blog & SEO	7	Build optimized blog content
S4 Landing Pages	8	Create high-converting HTML pages
S5 Distribution	4	Deploy and schedule content
S6 Analytics	5	Track performance and optimize
S7 Automation	5	Scale what works
S8 Meta	5	Plan, audit, improve the system

The Flywheel

The key insight: skills chain together. S6 Analytics feeds data back to S1 Research, creating a closed optimization loop. Each skill has typed input/output schemas for agent interop.

How It Works

Each skill is a SKILL.md file following the agentskills.io open standard. The AI reads the file and gains structured knowledge about a specific workflow.

# Install (Claude Code)
npx skills add Affitor/affiliate-skills

# Or paste the bootstrap prompt into any AI

Try It

Paste this into any AI:

Search the Affitor affiliate directory for AI video tools.
Use this API: GET https://list.affitor.com/api/v1/programs?q=AI+video&sort=top&limit=5
Show me the results in a table with: Name, Commission, Cookie Duration, Stars.
Then recommend the best one and explain why.

Architecture

Standard: agentskills.io (Anthropic, Linux Foundation AAIF)
Compatibility: Claude Code, ChatGPT, Gemini CLI, Cursor, Windsurf, OpenClaw
Data source: list.affitor.com REST API
License: MIT

How I Built a Game Bot That Uses LLM Vision for $0.01/Hour (Not $1/Hour)

sonpiaz — Sun, 15 Mar 2026 17:57:34 +0000

Most tutorials about using LLMs for game automation show you how to send screenshots to GPT-4 or Claude every few seconds. It works — but at $1+/hour, you'll burn through hundreds of dollars a month.

I built a framework that gets the same results for $0.01/hour. Here's how.

The Expensive Way

Every 3 seconds:
  1. Take screenshot
  2. Send to Claude: "What should I do?"
  3. Parse response
  4. Execute action
  5. Repeat

Cost: ~1,200 API calls/hour × $0.001 = $1.20/hour

This is how most LLM game agents work. It's simple, but the cost adds up fast — especially if you want to run 24/7.

The Cheap Way: 5-Layer Hybrid Architecture

The insight is that 98% of game decisions don't need AI. You can handle them locally, for free:

Layer 5: STRATEGIC AI ──── LLM reviews strategy every 30 min ($0.004)
Layer 4: WORKFLOWS ─────── Scripted tap sequences (FREE)
Layer 3: WORLD MODEL ───── Tracks timers and predicts events (FREE)
Layer 2: STATE MACHINE ─── Detects screens and handles popups (FREE)
Layer 1: PERCEPTION ────── Pixel analysis + OCR (FREE)

Layer 1: Perception (FREE, <100ms)

Instead of asking an AI "what screen is this?", I check pixel colors at known positions:

Bottom navigation bar is dark with colored icons? → Home screen
Dark overlay with bright center? → Popup
Bright panel in lower half? → Building menu

This takes 30-50ms vs 15-20 seconds for full OCR.

For the 20% of cases where pixels aren't enough, PaddleOCR runs locally to read text. Still free — no API calls.

Layer 2: State Machine (FREE)

A simple FSM tracks which screen the bot is on. Popups are managed with a stack — when one appears, the previous state is pushed. When dismissed, it pops back.

This handles the #1 frustration with game bots: random purchase popups interrupting your workflow.

Layer 3: World Model (FREE)

Instead of re-discovering game state every cycle, the bot remembers:

Which buildings are upgrading and when they finish
Whether research is active
Which troop training queues are idle

It predicts timer completions and acts at exactly the right moment — no polling.

Layer 4: Workflows (FREE)

Known tasks are scripted as deterministic tap sequences:

def upgrade_building(self, name):
    self.ensure_home()           # Navigate to home screen
    self.find_and_tap(name)      # Template match → cached pos → spiral search
    self.tap(720, 2100)          # Tap upgrade button
    self.tap(720, 1500)          # Confirm
    self.world.observe_upgrade_started(name)
    self.ensure_home()

Each step verifies the screen state before proceeding. If something unexpected happens, it falls back to Layer 5.

Layer 5: Strategic AI ($0.004 every 30 min)

Claude Sonnet reviews a screenshot and answers: "What's the most important thing to do right now?"

This catches things the scripted workflows can't:

Builders are idle but the bot didn't notice
A new event started that changes priorities
The current strategy needs adjustment

Cost: ~48 calls/day × $0.004 = $0.19/day.

Self-Improving: The Bot Gets Better Over Time

Two patterns borrowed from AI research:

Template Capture (Voyager-inspired): When the bot finds a building, it captures a screenshot snippet as a template. Next time, OpenCV template matching finds it instantly instead of searching.

Self-Reflection (CRADLE/Reflexion-inspired): Every success and failure is logged. Before retrying a task, the bot checks past failures to avoid repeating mistakes.

The Framework

I packaged this as an open-source framework where each game is a self-contained folder:

4x-game-agent/
├── agent/              # Core utilities (ADB, OCR, template matching, LLM)
├── games/
│   ├── kingshot/       # Reference implementation (fully working)
│   └── _template/      # Copy to add your game
└── docs/

The core is intentionally thin — just utilities. All game logic (screen states, workflows, coordinates, strategies) lives in the game folder. Wrong approach? Delete the folder and start over.

Getting Started

git clone https://github.com/sonpiaz/4x-game-agent.git
cd 4x-game-agent
pip install -e ".[all]"

# Test with Kingshot (or copy _template for your game)
cd games/kingshot
python main.py --test

Contributing

The most valuable contribution is adding a new game. The framework handles the hard parts (ADB, OCR, template matching, LLM integration). You just need to:

Map UI coordinates for your game
Write screen detection rules
Script the tap workflows
Define game knowledge (upgrade prerequisites, etc.)

GitHub: github.com/sonpiaz/4x-game-agent

Games we're looking for: Rise of Kingdoms, Evony, Lords Mobile, Call of Dragons, Whiteout Survival.

Have questions? Open an issue on GitHub or drop a comment below.

DEV Community: sonpiaz

Free AI models just got really good. Here's how to use 23 of them with one API key.

One key for everything

Which model for what

How it works in practice

Links

hidrix-tools: 18 MCP tools that give AI agents web search, social media scraping, and content analysis

The problem

What I built

Tools included

What I shipped in the last 24 hours

The feature I use most: Knowledge compiler

Install (3 commands)

Connect to Claude Code

Adding your own tool

Links

I built 50 AI skills for affiliate marketing — here is the social intelligence layer

The Problem

What I Built

The 5 New Skills

1. trending-content-scout (S1 Research)

2. content-angle-ranker (S1 Research)

3. traffic-analyzer (S1 Research)

4. content-research-brief (S2 Content)

5. infographic-generator (S2 Content)

The Flywheel

API-Optional Design

Feedback Protocol

Try It

I Built 45 AI Agent Skills for Affiliate Marketing — Here's How They Work

The Problem

What I Built

The Flywheel

How It Works

Try It

Architecture

Links

How I Built a Game Bot That Uses LLM Vision for $0.01/Hour (Not $1/Hour)

The Expensive Way

The Cheap Way: 5-Layer Hybrid Architecture

Layer 1: Perception (FREE, <100ms)

Layer 2: State Machine (FREE)

Layer 3: World Model (FREE)

Layer 4: Workflows (FREE)

Layer 5: Strategic AI ($0.004 every 30 min)

Self-Improving: The Bot Gets Better Over Time

The Framework

Getting Started

Contributing