The easiest way to build a game faster — and your best assistant once you do. Sentari AI Assistant doesn't wrap the engine from the outside; it works inside it. It reads your scene and viewport, reasons about what it sees, then authors real assets, Blueprint graphs, materials, PCG worlds and Gameplay Ability System data through a toolset it has already verified against the engine.
Each capability is a family of tools that Sentari has resolved against the engine's real API and runtime-tested. A few representative tool names are shown in each — the actual roster is 265+ and growing.
Spawn, move, align, distribute and snap actors; assign meshes and materials; trace and validate placement. Recipes scatter natural, terrain-following foliage as efficient instanced meshes.
Build whole graphs atomically — nodes, wires, pin defaults, compile and rollback as one unit. Pre-flight validates every name against reflection before a node is created.
Author master materials from a declarative node spec, instance them with parameter overrides, and audit exactly which params exist before you touch them.
Node-level PCG primitives — add/wire nodes, edit settings, read generated points. Turing-complete composition plus proven scatter recipes for landscape and spline, at World-Partition scale.
Read abilities, effects and attribute sets; inspect ExecCalc captures; then drive the live game — activate abilities by tag, fire cues, read cooldowns and live attributes.
An optional vision model reads the editor viewport or full desktop — lighting, layout, on-screen errors — and answers questions grounded in what it actually sees. Capability-checked per model.
Bring your own keys and generate in-editor: text/image-to-3D meshes (Meshy, Tripo), voice, SFX and music (ElevenLabs), and images (FLUX via fal.ai) — then drop the result straight into your scene.
Play the game, then command it — teleport actors, set properties, move AI on the NavMesh, inspect live state. Changes are transient, the asset stays clean.
+ Niagara VFX authoring & recolour · 800+ Epic editor actions · performance & asset audit · 12-language UI · in-panel CLI launcher for Claude Code & Codex.
Deep, verified tooling doesn't have to be expensive or complicated. Sentari pairs node-level depth with a local-first, budget-aware design — so the easy stuff costs nothing and the hard stuff costs little.
Every change is verified — compile-clean or PIE-tested — before Sentari reports success, and every action is undoable. Powerful enough for production, safe enough to trust blindly.
verify · undo · safeLM Studio and Ollama are built in, not bolted on. Native tool-calling, VRAM-aware keep-alive and per-model capability probing — your local GPU is a first-class citizen.
ollama · lm studioStart free on an open-source Llama, Qwen or Gemma. Reach any frontier model — Claude, GLM, Gemini — through OpenRouter or a direct API key. Same panel, your choice of brain.
byom · openrouterKeep the high-volume grind on free local weights; spend cloud only on reasoning that earns it. One-time license, no subscription, no credits to burn. The money stays in your pocket.
no subscriptionRun fully offline — your project never has to leave the machine. The same local-first architecture that makes it private is exactly what makes it affordable.
offline modeAsk in plain language; get real assets. From your first scattered forest to a balanced combat ability, Sentari is the fastest path from idea to thing-that-works-in-engine.
your best assistantSplit the workload like a smart team does: free local models carry the volume, and your paid cloud model is reserved for the heavy lifting. You stay fast, and your bill stays small.
High-volume, low-risk work runs on a local Llama, Qwen or Gemma — zero per-prompt cost, fully offline.
Complex reasoning, vision and long planning go to a frontier model — invoked only when the local lane can't carry it.
The same dockable panel shrinks to a slim chat rail that tucks beside your viewport, then expands into a full bench — chat, tool calls, 3D preview and logs — the moment you need the whole workspace. Switch on demand; your layout persists.
Letting an AI touch a real project takes more than a good demo. Three things hold true of every action Sentari takes — not as policy you have to trust, but as behaviour you can check.
Work entirely offline on a local model, or bring your own cloud key. No account, no telemetry, no middleman. Your project and your prompts go exactly where you send them and nowhere else.
A change is reported as done only after it compiles or passes a check, and it lands on the editor's undo stack like any other edit. An honest “that did not apply” is the plugin working, not failing.
Give it a free hand where you trust it and a hard boundary where you don't — per tool family, per asset folder. Guests you connect can be held to read-only. You draw the lines.
Sentari plans with the model, then calls verified tools that write into your project. Here's what a scatter request looks like end to end.
› you "scatter oak trees across the south hill, denser near the river" › sentari plans 1. read landscape bounds → south hill region detected 2. scatter_on_landscape → template: SM_Oak, density tuned by slope 3. instanced + random yaw → non-grid, terrain-following, 2 draw calls 4. generate + verify → 625 instances spawned, pivots on surface ✓ done instances: 625 · actors: 0 · seed: 1234 undo available · non-destructive · assets written to current level
The most underestimated thing Sentari does. Attach a reference image, describe the outcome you want, and it calls the real lighting tools in the same turn it looks at your picture — no values read off the photo by hand, no plugin-specific syntax to learn.
“Set the atmosphere and lighting like in the picture I gave you.”
The reference photo rides along with the message. The reply — from a local Qwen in LM Studio, not a cloud model — reports exactly what it changed: sun to a −5° horizon angle at 2700 K and 5 lux, orange height fog at density 0.06, a warm skylight at 2.5× for bounce, and a post-process pass with bloom 1.5, saturation 1.3, contrast 1.2 and −0.3 exposure compensation to keep the sun bright without blowing it out. The viewport behind it is already the result.
The same scene from the player's point of view. Nothing here was hand-tuned — every value came from the model reading the photograph, and every one of them is an ordinary editor change you can undo.
The captures above are real, straight from the editor. The clips below come from the same sessions — the remaining tiles fill in as the reel is cut.
Add Sentari_AI_Assistant from Fab to your project and enable it. UE 5.7 & 5.8.
Local (Ollama / LM Studio) or cloud, or any model via OpenRouter. The bridge listens on :8765.
The dockable panel connects automatically — compact or extended. Start talking to your scene.
Every action is verified and undoable — compile-clean or PIE-tested — before Sentari reports success.
One-time license on Fab — no subscription, no credits. Drop your email to be notified at launch, or tell us what you're building and we'll show you what the assistant can do for it.
