And why combining cloud and local models makes more sense than a pure SaaS solution
When you operate a network of dozens or hundreds of content websites, content is your fuel. Without it, sites stagnate, search engines reduce indexing, and organic traffic declines. Finding and paying copywriters for one hundred articles per day quickly becomes either a logistical burden or a financial absurdity.
That is why we built Article Orchestrator for WordPress, a local system that combines the power of cloud-based AI models with a locally running model on a standard computer. No SaaS. No recurring license fees. Full control over data. The only expense is API usage for external language models.
Why not just use ChatGPT?
Manual work with ChatGPT works as long as you manage two websites and have time. The moment you need to systematically supply dozens of websites with unique content, you hit a ceiling. Copying prompts, manually publishing into WordPress, no topic deduplication, no clear overview of what was generated and when. Most importantly, you rely on one model, one provider, one pricing structure.
The core idea behind Article Orchestrator is simple. Not every task requires the most expensive model. Writing articles, where language quality matters, is handled by GPT-5.1, Claude Opus 4.6, or any other flagship model via API. Filtering RSS feeds, deduplicating topics, quality checks, and categorization are handled by a local 12B model running in LM Studio on your own machine. In our case, an older MacBook M2 with 64 GB RAM and 12 cores. No cost. Hundreds of queries per day, zero additional expense.
In practice, paid API calls represent roughly one quarter of total AI calls. The remaining workload runs locally.
Agents with Their Own Identity
The system does not rely on a single monolithic prompt. Instead, it operates through agents, autonomous units where each has its own identity, personality (we call it a “soul”), tools, and memory.
For example, the RSS Scanner agent regularly processes configured feeds, extracts relevant headlines, and uses the local model to reformulate them into original topics for your website. The Topic Generator creates ideas from scratch based on the website’s focus and gaps in existing content. The Competitor Analyzer monitors competing websites and proposes topics you are missing. There is also a Custom Agent, essentially a blank template where you define your own logic, such as retrieving Google Trends data or pulling information from other sources.
Each agent operates on its own cron schedule. One may run every morning at six, another three times per day, another only on business days. Each retains memory of previous runs.
Multi-Provider Architecture: No Vendor Lock-In
One website can generate articles via OpenAI, another via Claude, and a third via a local model. Through the API, we continuously update the list of available models to ensure access to current and suitable options. When OpenAI releases a new model, it becomes immediately available in the system.
We have prepared a detailed case study that explains the system architecture, the agent framework, operational economics, and three practical deployment scenarios, from content site networks to corporate blogs and multilingual e-commerce content.
An English white paper with technical implementation details, cost comparisons, and compliance considerations is also available.
Interested in implementing this for your website network? Contact [email protected].
