GENERAITR Engine

Node-graph tools like ComfyUI are paved roads: built for a human dragging a cable between two fixed points. GENERAITR Engine has no road to pave.

You can bolt an MCP wrapper onto ComfyUI, and many people have. But every agent request still gets forced through the same paved road underneath: the agent describes intent, and something has to translate that intent into a graph a node engine can run. The interaction paradigm is still a human-era one, just with an API in front of it. GENERAITR Engine takes the opposite approach. The MCP surface is the execution model itself. Workflows are declarative step lists, not graphs, so an agent can move freely and solve a problem the way it judges best, not the way a fixed graph happened to be wired.

A deliberate constraint keeps the design honest: the Engine core is deterministic and LLM-free. Agents live entirely outside it (that orchestration is the job of TAOS, covered separately). Engine just executes what an outside agent asks for, safely and predictably. That separation is what makes it trustworthy infrastructure rather than another opaque AI black box.

What it offers above "just call a model API" is the layer serious production needs. A per-model knowledge bank captures each model's prompt grammar, strong and weak use cases, and parameter semantics, like having an expert on call for every model. On top of that: translation from creative intent to concrete parameters, presets, declarative workflows without node graphs, upfront time-and-cost estimation before you commit to a generation, and guardrails, including EU AI Act risk classification and content policy checks, that run before a job is ever queued rather than after.

Behind one interface, Engine routes jobs across multiple backend families: paid model aggregators (fal.ai, Runware, Eden AI, AIMLapi) for closed-source models, and self-hosted open-weight models on scale-to-zero serverless GPU infrastructure at Verda, an EU-based provider. Three configurable warmup modes, cold, auto and always-warm, let each deployment trade cost against latency explicitly.

Engine is in active development, and the work is concrete infrastructure engineering. Recent proof-of-concept work has been measuring real-world GPU cold-start times to decide whether pure scale-to-zero is viable or a warm-pool design is needed, one of the steps toward a production job runner. It already powers generation for the GENERAITR SaaS, but is designed to run fully standalone, with no multi-tenant or billing layer required.

The larger bet: the primary caller of generation infrastructure is increasingly an autonomous agent, not a person with a mouse. Engine is an attempt to build the layer underneath the current tools, for that world, from first principles.