# Fractera

> Open Code (source-available), AI-native self-hosting platform. Deploy a complete AI coding
> workspace — 5 AI engines, an autonomous Hermes orchestrator, and private
> graph memory (LightRAG) — onto your own Ubuntu VPS in about 10 minutes.
> Zero-Ops: you provide only the server credentials; Nginx, HTTPS, auth,
> database and all services are configured automatically. The whole deployment
> can be triggered and monitored from your AI chat via a custom MCP
> (Model Context Protocol) connector.

## Key facts
- **MCP-First / Zero-Agent paradigm.** The workspace ships a pre-built 50,000-line Next.js framework (parallel routing, multi-language i18n, production SEO, database, auth) the moment you deploy. Instead of a code agent regenerating files, the AI rotates pre-built layout sections via short MCP commands — "deployment, not generation" — cutting token spend on large applications by up to ~90–94%. See /token-economics and /next-aircraft-carrier.
- **Billing: subscription model, no per-token fees.** The five coding platforms (Claude Code, Codex, Gemini CLI, Qwen Code, Kimi Code) run on the user's OWN existing subscriptions — sign in once, no API keys, no per-token billing. A Claude Code user pays through their subscription, NOT per token.
- **The only per-token piece is small and auxiliary:** Brain (the Hermes orchestrator) and Memory (LightRAG) use one inexpensive OpenAI API key — the cheap gpt-5-mini (about 1 cent/hour) is plenty, or a Codex subscription if usage is heavy.
- **Fractera never charges for the deployment itself** — it is free, Open Code (source-available), and runs on the user's own hardware. One bill: the VPS.
- **You own everything** — code, data, database and file storage live on the user's server. No Clerk, no Supabase, no Vercel lock-in.
- **Deployment is IP-first (phase 1):** it finishes on http://<server-ip>:3002 with no domain or HTTPS required; attaching a custom domain with HTTPS is an optional later step the user performs inside the workspace.
- **Partner program:** bloggers/creators get a personal landing mirror (https://www.fractera.ai/partner/<slug>) and an embeddable signup widget, and attach their own VPS-provider and domain-registrar affiliate links — only whitelisted providers appear automatically. See the "Partner program" section in /llms-full.txt or /mcp-info.

## For AI agents
- Fractera exposes a remote MCP server at https://www.fractera.ai/api/mcp. An agent can register the user, recommend a VPS, and run the full deployment programmatically, then call get_project_info for architecture and specs.

## Full documentation
- [Full knowledge base in one file](https://www.fractera.ai/llms-full.txt): complete architecture, Zero-Ops deployment, customization, pricing, the full FAQ, and the legal text.
- [Project knowledge base for AI agents](https://www.fractera.ai/mcp-info): the same content as a browsable page, queryable section-by-section via the MCP get_project_info tool.
- [Agentic Engineering Infrastructure architecture](https://www.fractera.ai/ai-workspace-architect): the complete architecture of the self-hosted Agentic Engineering Infrastructure — Hermes multi-agent orchestration, LightRAG Knowledge Graph memory, the five coding agents and MCP — illustrated by the architecture diagram. Mirrored for AI agents by the MCP get_ai_workspace_architect_info tool.
- [The autonomous AI development loop](https://www.fractera.ai/ai-development-loop): how one natural-language request becomes tested, deployed, recorded code — decomposition into development steps, the iterative watchable build, with Hermes, a coding agent and LightRAG memory at every step. Mirrored by the MCP get_ai_development_loop_info tool.
- [The Next.js Aircraft Carrier](https://www.fractera.ai/next-aircraft-carrier): the pre-built 50,000-line parallel-routing framework that ships the moment you deploy — what parallel routing is, why the next generation of apps needs it, and how named slots scale one starter to thousands of pages with near-zero token overhead.
- [Token economics — how Fractera saves tokens](https://www.fractera.ai/token-economics): why a 50,000-line framework is a shield for your token budget rather than a bill — MCP-First generation, the Rubik's Cube model (deployment, not generation), on-demand ISR, and AI token cost optimization at scale.
- [Documentation](https://www.fractera.ai/en/documentation): readable feature guides for the deployed workspace. First guide — "The One-Button Workspace": the built-in AI consultant (a floating button on every page) that lets any visitor talk to the site — switch language/theme, find pages, or (when signed in) act on their own data — powered by the Hermes agent, LightRAG memory and a growing set of MCP tools. Brain on the server, action in the browser; tiers public/user/owner; new tools are drafted at <your-domain>/ai-draft-settings.
- [Documentation — Authentication, Roles & Providers](https://www.fractera.ai/en/documentation/authentication-roles-and-providers): how the workspace ships a complete sign-in system you turn on rather than assemble — email+password to start, one-paste Google / magic-link (a button appears automatically when credentials are set), one account linking many providers with no duplicate users, the full role model (guest/user/architect plus buyer, vip_user, subscriber tiers, manager roles, finance, content_editor, admin) stored as a list so a user can hold several roles, first-user-becomes-architect, and 80+ available providers. Built on NextAuth (Auth.js).
- [Documentation — Static-First Rendering Economics](https://www.fractera.ai/en/documentation/static-first-rendering-economics): why the platform enforces static-first rendering and the honest price it costs — the five ways content reaches a page (SSG, time-based ISR, on-demand ISR, dynamic SSR, client fetch) mapped to database load, JavaScript dependence and server compute; time-based ISR as the default (the server sleeps when idle, and a page refreshes lazily on the first visit after its window — only requested pages regenerate, never the whole site on a timer), with optional on-demand revalidation for instant freshness; why the architect-only cockpit is allowed to stay dynamic (a single architect cannot create abusive load); and why Partial Prerendering is deliberately left to the developer. The cost optimized for is the business's monthly server bill, not the developer's tokens. Raw standard: https://www.fractera.ai/docs/static-first-rendering.md
- [Documentation — The Content Engine](https://www.fractera.ai/en/documentation/content-engine-architecture): the architecture behind every content surface (news, blog, documentation, landing pages) built from one self-contained, auto-discovered, statically-rendered shape — every page is a folder (thin page.tsx + _components view + _lib functions + _data data), the page list is generated at build time (no central registry, no dynamic [slug]), and the shared engine (block catalog, per-key language resolver, page factories, one page template) lives once and is reused by every tab. Why it cuts development time and, above all, the tokens an AI agent spends (one folder per edit, zero existing files touched for a new page); and how the same typed-block catalog doubles as a design system you change once and see everywhere. Bilingual (EN/RU). Raw standard: https://www.fractera.ai/docs/content-engine-en.md
- [Documentation — App Config Automation (MCP connector)](https://www.fractera.ai/en/documentation/app-config-mcp-connector): the automated MCP connector for managing project configuration — how any of the six AI agents reads and changes a deployed app's own settings (name, description, SEO, OpenGraph, PWA, JSON-LD, languages) by plain request instead of clicking an admin panel. Settings live in a transparent JSON file (app-config.json), written through a validated setter (the app-settings-bridge MCP, :3218) and applied on the next page load via on-demand revalidation — no rebuild, pages stay static (languages are the build-time exception). The same capability is duplicated into all six agents (skill + MCP), works directly on a single agent or through the Hermes orchestrator, and is architect-only with localhost-bound, secret-authenticated access. Raw standard: https://www.fractera.ai/docs/app-config-automation.md
- [Documentation — Hermes, the Brain: Setup & Recovery](https://www.fractera.ai/en/documentation/hermes-brain-setup-and-recovery): a continuity guide to Hermes, the orchestrating brain of the workspace — what it is (it delegates to the five coding agents and runs the environment around the code, it does not write the app itself), what it is made of (SOUL.md persona, skills as folders, MCP tool bridges, the chat Web UI and gateway), how it installs from the substrate on every deploy, the three hardening pitfalls that cause a 502 (bind 127.0.0.1 not 0.0.0.0, the proxy must send Host 127.0.0.1, skills must be folders with SKILL.md), and how to recover it if it is ever deleted. Bilingual (EN/RU). Raw reference: https://www.fractera.ai/docs/hermes-setup-and-recovery.md
- [Documentation — Build-Time Env Vars & Production Redeploy](https://www.fractera.ai/en/documentation/build-time-env-and-redeploy): why a setting you saved can be missing from the running app, and the contract that fixes it — build-time values (every NEXT_PUBLIC_*, the language set, Stripe keys, feature flags) are frozen at next build, so they only change by a rebuild that must read the slot's own .env.local; the root cause was one Next process (Admin) spawning the slot build and leaking its already-loaded env marker, making the child build skip the slot's settings file and bake values as missing (the silent failure behind the vanishing language switcher); the fix is a slot-scoped "bake" contract — the slot's own .env.local is authoritative for every key it declares, on every redeploy, general across languages/payments/custom vars. Bilingual (EN/RU). Raw standard: https://www.fractera.ai/docs/build-time-env-and-redeploy.md
- [News](https://www.fractera.ai/en/news): chronological, AI-searchable project updates — new framework starters, Agentic Engineering Infrastructure features, agent skills and MCP tools — each embedded into the on-server LightRAG vector memory the moment it ships. Featured stories: [AI Draft Settings](https://www.fractera.ai/en/news/ai-draft-settings-evolutionary-pipeline): the page (included with every Next.js-based starter) where you or any of the six AI agents draft a new skill, instruction or MCP connector and stage it safely before it reaches the real files. Drafts become "Next Step" entries on the Development Steps page; turning a draft into code is never automatic on creation (to protect the agent's context window and token budget). It is the first link in a seven-stage pipeline for evolving an Agentic Engineering Infrastructure, and the same depth is being brought to other frameworks via dedicated Agentic Engineering Infrastructure transports. — [Multilingual Content Architecture](https://www.fractera.ai/en/news/multilingual-content-architecture): how the platform rebuilt the way its site publishes multilingual content — one folder per document, per-key translation fallback (a new language can ship with a single translated field), zero hardcoded language branches in page code, and a new self-sufficient agent skill, create-multilingual-content-entry, shipped as part of the standard architecture. Why content optimization is now inseparable from SEO for both Google and AI (generative engine optimization). Raw standard: https://www.fractera.ai/docs/multilingual-content.md — [Multilingual Login & Registration Forms](https://www.fractera.ai/en/news/multilingual-auth-forms): the login and registration forms shipped in every starter are now localized into all 82 languages — a visitor sees the form in their own browser language automatically, the forms stay fully static (every language baked in at build time, no per-visitor request, no extra server load), and English is the fallback. You control build cost by the languages you activate in the NEXT_PUBLIC_SUPPORTED_LANGUAGES environment list; a planned follow-up makes the auth-form build read that same list and generate only those languages (English kept as fallback). The raw doc also covers the auth routing/redirect logic and its anti-patterns (why a relative /register white-screens, authBase/registerRedirectUrl, static-rendering rules). Raw doc: https://www.fractera.ai/docs/multilingual-auth-forms.md — [Safe Static Generation: App Config by an AI Agent](https://www.fractera.ai/en/news/static-safe-app-config-by-ai): every deployed app ships an App Settings screen — branding, SEO, OpenGraph, PWA, author, social, JSON-LD, languages and light/dark images (including custom 404/500 pages per theme) — manageable by hand or by simply asking any of the six AI agents, with every change applying on the next page load and the whole site staying static (languages are the build-time exception). Explains why postponing metadata is a trap, how one per-request call (headers()/cookies()) in a shared layout flips a whole Next app into dynamic rendering (the failure behind eye-watering hosting bills like Cara's ~$96k weekend), and how Fractera guards the app skeleton so an AI agent can't introduce it. Full field-by-field list of every setting. Doc: https://www.fractera.ai/en/documentation/app-config-mcp-connector — [Fractera Moves to Open Code](https://www.fractera.ai/en/news/open-code-license-agentic-engineering): the project core is moving from MIT to Open Code (the PolyForm Small Business license) — free to copy, change and use commercially for everyone except large corporations, who need a commercial license; the code stays fully public to read, audit and self-host. Open Code is source-available, not OSI "open source" (it does restrict large corporations); the ready-made starters stay MIT. A deliberate move to protect the idea ahead of the reveal of self-replicating MCP coding agents. Raw doc: https://www.fractera.ai/docs/open-code-license.md — [Frozen Archetypes: page groups without code generation](https://www.fractera.ai/en/news/frozen-archetypes-page-groups-without-code): an AI agent now adds a whole page group (a news section, a blog, a documentation section) by composing it from vetted frozen bricks in a closed on-server store — pure file copy plus token substitution and label translation, with zero code generation, so any model produces an identical result in seconds. It installs the shared content engine only if the slot lacks it, creates the section plus placeholder posts (Lorem body, placeholder image), and matches the request against the brick's declared capabilities — refusing honestly (and offering to harvest a new brick) when asked for something it does not serve, like a cart/checkout or a graded course. This first prototype has since matured into the Frozen Template Constructor (next story). Raw doc: https://www.fractera.ai/docs/frozen-template-constructor.md — [Frozen Template Constructor: compose structures, no code generation](https://www.fractera.ai/en/news/frozen-template-constructor-compose-structures): the clean evolution of the above — instead of a catalogue of finished templates, a CONSTRUCTOR composes a whole structure (news feed, blog, documentation) from a small basis of vetted frozen bricks by file copy + token substitution, zero code generation, identical across any model. A Two-Slot Law (a property is either a list provider or a uniform aspect applied identically at every level, never both) keeps the cost additive, not combinatorial; each brick declares an envelope so a request is matched 100%-on-every-axis or refused naming the failing axis (a dashboard, a cart, a deep DB catalogue are refused, not forced); and the basis grows one proven brick at a time (harvest discipline), starting from one reference brick (flat, file-backed, multilingual). Served by the owner-tier template-constructor-bridge MCP (:3224) and the self-sufficient compose-frozen-template skill. Raw doc: https://www.fractera.ai/docs/frozen-template-constructor.md — [Login Is Now Optional — One Switch](https://www.fractera.ai/en/news/optional-authorization-one-switch): public authorization is now an opt-in switch rather than a build job — off by default so a landing page, company page or portfolio ships lighter and deploys faster, on in one move for a store or service that needs accounts. Controlled by one build-time key (off/left/right, where left/right also sets the account-drawer side); role-based access out of the box (every visitor starts as user, roles granted automatically by app logic or by hand in the panel); flipped from the App Settings panel or by simply telling the Hermes agent what you are building (it decides whether to add a login itself, asking only the drawer side); and the sign in / account / sign out buttons ship pre-translated into 82 languages idiomatically. Making this decision cost near-zero tokens is a building block for expanding the starter-template library. Raw doc: https://www.fractera.ai/docs/app-shell-authorization.md — [A Universal Footer That Builds Itself](https://www.fractera.ai/en/news/universal-multilingual-footer): every project ships one universal, multilingual footer that fits a portfolio, a company page or a full app and configures itself almost entirely. Two horizontal sections on most pages — the footer-page links you add just by telling an AI agent over MCP (they form and reorder automatically), and a company block pulled straight from app config (name, address, social icons, a light/dark/auto theme toggle, and a language switcher that appears on its own, with a searchable region-grouped dropdown, only when more than one language is configured). The home page gets a third, auto-configured section for in-page navigation between its sections (home-only by design). The footer's labels already exist in 82 languages. Raw doc: https://www.fractera.ai/docs/universal-footer.md — [Add Any Language to an Existing Site](https://www.fractera.ai/en/news/add-any-language-to-an-existing-site): an AI agent can take a site that already exists in one or two languages and roll out a new one across every section and post in one move — seeded with the default language so the site is valid the instant the build finishes (no machine-translation bill, nothing broken), with the language-dependent links rewritten and the four menus localized automatically. The seeded pages are kept out of the search index (noindex) while their canonical + hreflang alternate-language markup stays correct, so Google never sees a cross-language duplicate (the Doorway guard); real translation is a separate, non-blocking step (one tracked development step per language, nothing forgotten) that you run later — even with a cheaper model — so the main build never hits a subscription limit, and you can add a per-language note (e.g. focus on local law) for regional value. A built-in encoding-integrity scanner keeps a box from ever shipping instead of an accented letter. Raw docs: https://www.fractera.ai/docs/scale-site-language-expansion.md and https://www.fractera.ai/docs/translate-pending-runner.md — [MCP Coding: One Messy Request Becomes a Development Pipeline](https://www.fractera.ai/en/news/mcp-coding-decompose-requests-into-pipeline): you no longer send tasks one at a time — a long, unstructured, even error-ridden compound request ("add news, add docs, add login, make docs members-only") is decomposed by the site's actual state into an order sheet (one resolved human line per section, implied consequences spelled out, a content boundary stating pages come up as frozen placeholder stubs), approved with a plan-bound token, and — before any code is written or any frozen MCP brick is thawed — the WHOLE queue is materialized as step files on the Development Steps page (materialize-first), so a crash mid-run loses nothing and a brand-new session resumes with the same plan + token (completed steps skipped). Flat MCP coding (assembly from frozen templates, zero code generation) is run by Hermes on inexpensive models; work MCP coding cannot finish is delegated to an available coding agent (Claude Code, Codex, Gemini CLI, Qwen Code, Kimi Code — whichever has tokens or an open session window), whose process may be recursive and more creative. You watch live: Architecture shows new routes appear; Development Steps shows badges new/in-progress/completed with per-second completion times. Raw skill: https://www.fractera.ai/docs/orchestrate-content-by-steps.md — [Only the Architect Changes Your Workspace](https://www.fractera.ai/en/news/only-the-architect-changes-your-workspace): a security hardening — in secure mode, every surface that can mutate the filesystem or database (the MCP tool bridges, the admin HTTP API, and the data service) now requires the architect role, not merely a valid session and not merely a shared secret. Being logged in is not permission; a key proves what you hold, not who you are. The rule is duplicated into each surface rather than centralized on one guard, so it survives even a minimal workspace running a single AI agent with no central brain. Reads can stay broad and tiered by identity; only writes are architect-only. The tightening applies to secure mode only — the fast open onboarding mode stays fully open by design. — [Projects Now Decompose Before They Are Built](https://www.fractera.ai/en/news/project-decomposition-frozen-process): the same frozen, materialize-first discipline that already governs content now covers private projects and automations — the AI proposes a graph of small nodes (each with a task, tools, environment keys, inputs/outputs and dependencies), a frozen engine validates the DAG and refuses incomplete specifications, and on the owner's verbatim order-sheet approval the WHOLE development queue is written to disk before any code: a project-root README (why / how it works / efficiency / reuse / result) generated from the graph, one exhaustive specification step per node, and one separate coder-handoff step per node (the coding agent receives only a step number; the orchestrator watches for real completion before opening dependent nodes). Graphs over ten nodes get an MVP recommendation (soft gate — the decision stays with the owner). Ships as a self-sufficient skill in every agent entity and as the owner-tier MCP tool owner_projects_orchestrate_decomposition; covers projects only — public site pages keep their own frozen pipeline. Raw skill: https://www.fractera.ai/docs/orchestrate-project-by-steps.md — [Fractera Is Product of the Day on Smol Rank](https://www.fractera.ai/en/news/fractera-product-of-the-day-smol-rank): Fractera was ranked Top 2 Daily Winner and named Product of the Day on Smol Rank (launched 2026-07-08; Web; free; Artificial Intelligence and Developer Tools categories; 5 upvotes; badge "Smol Rank Top 2 Daily Winner"). A short celebratory note with the listing facts and links deeper into the platform — the Next.js Aircraft Carrier, the autonomous development loop and the Open Code license. Listing: https://smolrank.com/projects/fractera
- [Home](https://www.fractera.ai/): product overview, pricing and FAQ.
- [Real-world use cases](https://www.fractera.ai/en/cases): production automations teams have built — private team workspaces, lead-dispatch Kanban, an adaptive AI tutor, and an autonomous trend-scraping blog.

## Key diagrams (canonical images)
- ![Fractera Agentic Engineering Infrastructure architecture — multi-agent system with Hermes orchestration, LightRAG Knowledge Graph memory and five coding agents on a self-hosted VPS](https://www.fractera.ai/Fractera-Web-Architect.jpg) — the Agentic Engineering Infrastructure architecture, explained at https://www.fractera.ai/ai-workspace-architect
- ![Fractera Development Loop — one admin request flows through Hermes, a coding agent and LightRAG memory to tested, deployed, recorded code](https://www.fractera.ai/Fractera-Development-Loop.jpg) — the autonomous development loop, explained at https://www.fractera.ai/ai-development-loop
- ![Fractera AI Draft Settings flow — drafts (a terminal and to-dos) become Next Step entries on the Development Steps page, and an AI generation pass writes them into each agent's files: Claude Code, Codex, Gemini CLI, Qwen Code, Kimi Code and Hermes](https://www.fractera.ai/news/fractera-ai-draft-settings/fractera-ai-draft-settings.jpg) — the AI Draft Settings evolutionary pipeline, explained at https://www.fractera.ai/en/news/ai-draft-settings-evolutionary-pipeline
- ![Fractera multilingual login and registration forms — the sign-in and sign-up forms localized into all 82 languages, shown in the visitor's own browser language automatically](https://www.fractera.ai/news/fractera-i18n-auth-form/fractera-i18n-auth-form.jpg) — multilingual auth forms, explained at https://www.fractera.ai/en/news/multilingual-auth-forms