Agentic Engineering Infrastructure: Self-Hosted & Token-Free

There is a widespread belief that an orchestration layer like Hermes is wasteful — that it is better to talk directly to a coding model and skip the middleman. We disagree completely, and here is exactly why.

The real cost driver in AI-assisted development is not the number of requests — it is context window inflation. Every message in a direct coding session appends to the running context. After two or three hours on a typical project, the context is carrying the full transcript of the session. Most providers charge for both input and output tokens, and because every new message pays for everything before it, costs grow faster than linearly. Real-world data shows a Claude Code session on a mid-size project in hour three costs five to ten times what it cost in the first fifteen minutes. When you forget to manage the context window yourself — and everyone forgets eventually — that cost compounds.

Hermes is the external observer that prevents this from happening. It runs on inexpensive models (gpt-5-mini, qwen-turbo, and similar) that cost twenty to fifty times less than frontier coding models. These models do not need sophisticated reasoning — they read a spec, extract what matters, route a task, and immediately clear the context before the next call. That is exactly the work they are suited for.

• ✓LightRAG eliminates session amnesia — without persistent memory, every new session spends the first 10–30% of its tokens re-explaining the project from scratch. LightRAG stores architecture, patterns, anti-patterns, and decisions, so the actual coding prompt can be laser-focused from the first word.
• ✓Precise task specs from context — instead of a vague "build a dashboard", Hermes reads LightRAG and generates: "In app/dashboard/page.tsx, add a component using the Card pattern from components/ui/card.tsx, following the auth pattern from lib/auth.ts." This alone eliminates several back-and-forth exchanges per task.
• ✓Immediate context clearing — after each task completes, Hermes clears the coding model's context. The next task starts with a clean slate, not carrying the entire previous conversation.
• ✓Intelligent routing between models — simple tasks (documentation, refactoring, file renaming) go to cheap or free models. Complex architectural decisions go to frontier models. You pay frontier prices only for frontier work.
• ✓Pattern and anti-pattern tracking — Hermes knows what approaches have worked in your project and which ones caused problems. It does not let the coding model re-discover the same dead ends.

This is only the surface. The deeper value is in what Hermes prevents: redundant searches for components that already exist, re-solving problems that were solved last week, sending the same context to five models in parallel when one focused call would do. Every one of those is a token cost that never appears on your invoice because it never happened.

We believe the future of efficient AI-assisted development is not about squeezing more out of a single model — it is about a system that knows your project well enough to ask the right model the right question at the right moment, and stop before it becomes expensive. That system is Hermes, and it is at the core of what Fractera is building.

Regular vibe coding puts all the heavy lifting on the AI: design the architecture, write boilerplate, locate the right component, recall what was decided last session. Every token spent on that overhead is a token not spent on your actual feature.

Fractera eliminates that overhead at every layer:

• ✓Production-ready starters — Auth, database, file storage, and advanced routing ship fully pre-configured. The AI skips months of scaffolding and goes straight to your feature from day one.
• ✓Component highlighting — Click any element on your live site to jump directly to its source. No tokens wasted asking "where is the navbar?".
• ✓LightRAG, your company brain — A persistent vector store that remembers your entire codebase, every architectural decision, and domain knowledge. Every AI message arrives with full context.
• ✓AI-optimized skills & instructions — Pre-built prompts and workflows designed for non-professional developers. The right approach on the first try.
• ✓Cross-platform orchestration — LightRAG coordinates all five coding platforms so they share context. Switching from Claude Code to Gemini CLI doesn't mean losing the thread.

The result: tasks that take 10–20 back-and-forth messages in a vanilla AI chat typically resolve in 2–3 focused exchanges inside Fractera — token spend drops 70–90% compared to classic vibe coding.

This is Fractera Pro. And once you've tried it, there's no going back to plain vibe coding.

Modern web development has a tempting formula: combine a handful of "free" services — Clerk for auth, Supabase for your database, a dozen other cloud APIs. It looks great on day one.

The problems surface when your prototype becomes a real business. Each service quietly updates its pricing. What starts as free becomes a paid tier that scales with traffic.

Worse: ten billing cycles to track. Miss one and your project goes dark. No warning, no grace period. Partners who switched to Fractera often share this story: a live project lost because one service expired. By the time they figured it out, the reputation damage was done.

Fractera is built around a different principle: everything your business needs — authentication, databases, media storage — lives on your server, not spread across a dozen cloud dashboards.

• ✓One subscription, one server — cost does not scale with your users.
• ✓If you pause your business, your data does not disappear. Back it up and restore when you're ready.
• ✓Your application code lives on GitHub — recovery is always possible, even if dependencies have aged.
• ✓Built-in AI systems can help rebuild the project even when some packages are outdated.

One server. One bill. Maximum resilience — whether you are running one project or ten.

You use your own subscriptions — that is the whole point. The five coding platforms (Claude Code, Codex, Gemini CLI, Qwen Code, Kimi Code) run on the subscriptions you already have: you sign in once with your normal account, like a local CLI. No API keys to manage, no per-token billing surprises. This is where the heavy AI work happens, and it costs you nothing beyond the subscription you already pay the AI vendor.

The only per-token piece is small and auxiliary: Brain (Hermes) and Memory (LightRAG) use one inexpensive OpenAI API key — the cheap gpt-5-mini runs at roughly a cent per hour and is plenty to get going. If your Brain usage grows heavy, you can switch it to a Codex subscription instead.

• ✓Coding platforms (Claude Code, Codex, Gemini, Qwen, Kimi) — your existing subscriptions, no per-token billing.
• ✓Brain + Memory — one small OpenAI key (gpt-5-mini, ~1¢/hour), or a Codex subscription if heavy.
• ✓So a Claude Code user pays through their subscription, not per token.

For full AI-coding workloads, the recommended minimum is 6 cores and 8 GB RAM. Storage depends on your project — 75 GB is a solid baseline.

Once active AI development wraps up, you can downgrade to 2 cores / 4 GB RAM — typically €1–2 per month.

The complete cycle: from an empty VPS to a live product on the marketplace. Under 30 minutes to launch, everything after that runs at your development pace.

1. 1Buy a VPS from Contabo (our recommended provider). Receive your IP address, login, and password.
2. 2On the "Your Server" tab, enter those credentials and launch the deployment. Watch real-time events for a few minutes. Two emails arrive: one when installation starts, another when it completes.
3. 3Your dashboard sits in the top-right corner — click your name after signing in. Inside you will find direct links to your live site and admin panel. The same links arrive in your completion email.
4. 4Sign up in the admin panel with a real email and password (save them carefully). The first user becomes the administrator — only admins can change settings. You can promote more admins later.
5. 5Have your own domain? Open "Settings", enter the domain name, copy the A-record. Set it at your DNS provider — and it works immediately.
6. 6Authorize one or more coding platforms (Claude Code, Codex, Gemini CLI, Qwen Code, Kimi Code). Codex is recommended as the channel for working through Hermes. Authorization is standard browser-based, just like in a local CLI. One-time setup.
7. 7Open Hermes and configure it your way. Add a messenger channel, or use the chat directly in the admin panel.
8. 8Tell Hermes what to change or what to generate. For example: "Let Claude do this and Codex do that" — both run in parallel through Hermes orchestration.
9. 9Upload documents (requirements, business description, any reference material) to your Company Brain — powered by LightRAG — manually, or ask Hermes to automate your data ingestion.
10. 10Want to accelerate? Ask Hermes to find a ready-made solution from the Fractera skills marketplace.
11. 11Your domain is live from minute one. After every development cycle, the site updates automatically — no manual deploy step.
12. 12Finished a project? Ask Hermes to turn it into a skill and publish it on the marketplace. Set a price, sell it, earn.

This loop repeats as many times as you want. Every product stays yours — on your server, under your domain, your code, your data, your subscriptions.

When you start a business for the first time, almost everything around you is unfamiliar. Registering the company, opening a bank account, standing up a website and social profiles, streamlining production, finding traffic, balancing the unit economics, and finally scaling. An AI agent helps at every step — but it is not a single "do it all" button: it is a road you cannot walk in a day, or even a month.

Along that road the AI keeps proposing solutions. Some turn out right, some wrong; some save you money, some waste it, and once in a while income appears out of nowhere. Until now, this priceless zero-to-product experience lived only in your head. Now it can be digitized.

And this is nothing like reading a book or running through a checklist. The AI that solved every task alongside you, listened to your conversations, prepared your presentations — can assemble not a step-by-step manual but a living knowledge base. Not yes/no, but the complex conditional logic of decisions that modern AI can analyze and carry over — even to a different region, with different inputs and a different outcome.

So whoever is lucky enough to buy your digitized experience walks the same road faster and with far fewer mistakes. Admit it — you would want to be in their place too: to save years of your life and not pay for someone else's mistakes with your own money.

• ✓More than a book — a book gives one route; the loop remembers the forks.
• ✓More than a checklist — a checklist answers yes/no; the loop reasons in context.
• ✓More than a course — a course teaches in general; the loop knows your exact path.

This is what we call the product loop — and it is the product we want to build in the near future.

Already have experience, or a company of your own? Want to digitize your current business — or your next one? Become a project sponsor: you join a private Telegram group where you can talk to the founder directly and be first to shape what our core service becomes — the product loop and its skill marketplace.

Stay with us — we will build it together.

Yes. Connect your existing GitHub repository to your Fractera workspace and start coding with AI immediately. Depending on your project's complexity, some initial migration steps may be needed — Fractera's built-in AI assistants can guide you through.

Yes — once a project is complete and no longer needs active AI-assisted development, you can export it to Vercel.

Important: moving to Vercel means leaving the Fractera environment behind. The AI coding workspace, terminal access, LightRAG memory, and all five development platforms only run on your own server.

Also keep in mind that Vercel and cloud storage pricing can escalate quickly under real-world traffic. Migrating back to a self-hosted Fractera server is straightforward — your code is already on GitHub.

Yes — you can deploy a plain server with no AI at all. It is a common, deliberate choice, especially among experienced developers who want to offload the work of managing cloud resources and avoid DevOps overhead.

A server like this can be used at any time, synced with your home or work IDE (for example, Visual Studio Code), or simply to host and run your existing projects on infrastructure you control.

• ✓Optional authentication — turn it on whenever you need it.
• ✓Local database and object storage (for example, media) — included, with no cloud subscriptions.
• ✓Freedom from cloud lock-in — meet regulatory or internal-policy requirements and avoid surprise bills.

A plain server like this needs only a lighter, more economical VPS. During deployment, switch to custom mode and tell us which services you want.

Don't worry if you turn off something you might need later — you can always add any recommended or custom AI tool through the terminal built into your admin panel.

Not necessarily. If the Fractera connector is your only custom MCP, you can use it on the free tier and get the full deployment service — no subscription required.

The condition is that you have not added another primary or custom MCP. If you already use a different custom MCP, you have two options: remove it before adding the Fractera connector, or move to a paid Claude plan (for example, Claude Pro, around €20/month), which lets you keep several custom connectors at once. Note this is Anthropic's own plan for the connector — it is unrelated to Fractera, which is always free.

Yes — and for developers with an existing local setup, this is often the most natural workflow.

If you prefer working with hot reload in your IDE, you don't have to change your habits:

1. 1After setting up your Fractera project, connect it to a GitHub repository and push the initial codebase.
2. 2Pull the code to your local machine. Develop the way you always have — your IDE, hot reload.
3. 3When ready to ship, push your changes to GitHub.
4. 4Return to your Fractera workspace, pull from GitHub, and hit Deploy.

Your changes go live in production in minutes. In effect, your own server becomes a self-hosted alternative to Vercel: GitHub is the bridge between your local environment and production.

Your local environment will continue to use the database and file storage that live on your server — no cloud subscriptions required.

Yes — the Fractera Partner Program. The model is simple and provider-agnostic: you recommend Fractera deployment, your readers buy a VPS from any host with an affiliate program, the host pays you directly.

Fractera doesn't pick the provider for you and doesn't restrict your choice. You decide who to work with: Hetzner, DigitalOcean, Vultr, Hostinger, RuVDS, Timeweb, Contabo — dozens of hosts run affiliate programs. Direct programs (RuVDS, Timeweb, Hostinger) approve in a day and pay up to 30–40% of the customer's first payment. Affiliate-network programs (Contabo via CJ.com, others via Admitad) pay a fixed amount per new customer but take a few days to approve.

The Fractera cabinet works with any affiliate link — we provide the infrastructure (mirror page, embed widget, MCP, cabinet), the host pays the commission.

For questions: [email protected]

Your income from partnership is closer and simpler than you might imagine. One approval at any hosting provider, then passive payouts for as long as your content keeps bringing readers.

1. 1Pick a host with an affiliate program. Direct programs are faster: apply on the provider's own site (Hetzner, RuVDS, Timeweb, Hostinger and others each have their own affiliate cabinet). Affiliate networks (CJ.com for Contabo, Admitad for several providers) are slower but unlock a catalogue of 3000+ brands through a single registration.
2. 2Get your affiliate link once approved by the provider.
3. 3Open the Fractera partners page, create your partner cabinet, add your affiliate link (or multiple links from different providers — the cabinet supports several at once).
4. 4Receive your personal Fractera landing mirror, an embed widget for your blog, and a single MCP link for AI agents. Your links replace the defaults in the "Where to buy a VPS" block.
5. 5Recommend Fractera in your blog, YouTube channel, social posts — wherever developers and entrepreneurs gather.
6. 6Any visitor who follows your link and buys a server from a partner host triggers a payout. Amount depends on the provider: direct programs pay up to 30-40% of the first payment, networks pay a fixed amount (Contabo: 20-55 € depending on configuration). The host pays you directly — Fractera is never in the middle.

The math is simple: one approval at one provider, then passive payouts for as long as your article, post, or video keeps bringing traffic. Hosts compete hard for new customers — and you get rewarded for every one you bring in.

Yes — and this is a different kind of opportunity. We are looking for official regional partners to build localized versions of Fractera together: adapted for local regulations, delivered in the native language of the region, backed by people who understand the local market from the inside.

This is not a reseller arrangement. A regional partnership means your own branded product, your own customer base, your own revenue stream — on an open-source foundation that is already production-ready. The infrastructure is built. What a regional partner brings is market knowledge, local trust, and the ability to serve customers in ways a remote team cannot.

• ✓Regulatory alignment — help your region meet local data residency and compliance requirements.
• ✓Native-language support — technical support and documentation in the language of your customers.
• ✓Your own business — not a franchise, a co-venture with full transparency into the product.