NEXUS AI Labs — Research · Build · Ship

Research AI.
Build What's Next.

NEXUS AI Labs researches the convergence of blockchain and AI —
at the intersection of intelligence and trust,
we prove our research as real, working products.

Explore Solutions Meet the Core

Manifesto

If AI is intelligence,
blockchain is trust.

Blockchain brought ownership and verification to the digital world for the first time. AI layers intelligence and autonomy on top of it. Where the two converge — verifiable agents, on-chain autonomous execution, tokenized creations — the infrastructure of the next economy is taking shape.

NEXUS AI Labs researches AI built specifically for the blockchain+AI ecosystem. Rather than staying in theory, we prove every finding by shipping real products. When AI intelligence is layered on the trust blockchain has established, the full cycle of creation, operation, and monetization runs autonomously. The Nexus AI framework is the engine of that entire pipeline.

Core framework

Nexus AI

Apps shipped

+3 this year

6 / 5

Desktop / Web

Wails · Next.js

Tech stacks

polyglot

Products — flagship apps by NEXUS AI Labs

Built by us.
Proven by use.

These are the products NEXUS AI Labs concluded were needed — and built. Each app owns one stage of a real work pipeline and is running right now.

Bob

Davinci

01Step 00 — The Engine Behind Everything

Davinci

Declare a goal, and the loop ships the product

Nexus AI Framework IDE — autonomous development powered by Loop Engineering

Every Nexus Labs product was built inside Davinci. Declare a goal and the loop autonomously repeats planning, design, development, verification, and improvement. A 7-stage pipeline from Idea to Report runs as a PDCA cycle, 3-tier verification gates detect quality convergence automatically, and Wave orchestration accelerates everything with parallel execution. Monaco editor, PTY terminal, multi-backend AI, one-click deployment — an autonomous platform that goes beyond code suggestions, running the full course from planning to production.

7-stage loop pipeline — Idea → Plan → Design → Work → Verify → Improve → Report
3-tier verification gates + automatic convergence detection, 5-tier complexity classification
Wave orchestration parallel execution — 3.5× faster than sequential

Explore Davinci HomepagemacOS desktop (Tauri 2)

Text to 3D

02Forge — Text Becomes 3D

Text to 3D

One sentence becomes game-ready 3D

Text to 3D — realtime 3D generation engine, about a minute in the browser

One line of description yields a finished 3D asset — lighting, textures, and cinematic framing included — in about a minute, right in the browser. No mesh library, no downloads, no manual modeling. An LLM writes materials, lathe profiles, building macros, and even camera composition as a compact JSON DSL; every spec passes schema validation with auto-repair, then Three.js assembles PBR materials, seamless procedural textures, and lights. A VLM art director scores each render and rewrites the spec, raising quality every pass. One click exports a GLB with lights included, ready for Blender, Unity, Unreal, or three.js. 60 FPS even on the largest 117-part scene — every number here was measured on the running engine.

Self-correcting 5-stage pipeline — prompt → LLM parametric spec → validate + auto-repair → procedural build → VLM refinement
Assets are data — 4.7 KB average JSON spec: diff, version, copy, remix
117-part scene at 60 FPS + 145 ms GLB export — straight into Blender · Unity · Unreal · three.js

Explore Text to 3D Live GalleryWeb (realtime in-browser rendering)

AIVE

03Ad — Where Game Ads Get Made

AIVE

Upload two images, and a game ad is done

Story · cut plan · video generation · cut editing — the game-ad AI studio

AIVE is a web studio that lets game companies make their own ads and trailers. It was born from a practical need: keep shipping ad creatives without a video editing team. Write a short story and upload character images, and a cut plan starting from CUT 01 [00:00 ~ 00:02] — Hook fills itself in with Beat, Visual, Camera, and Audio per cut; that plan is handed to Seedance 2.0 and rendered into a video of up to 30 seconds and 20 cuts. Export in 9:16, 16:9, or 1:1 and publish straight to anywhere from TikTok to Steam.

Game trailers up to 30 seconds from two images (Seedance 2.0)
Automatic cut plans — Beat · Visual · Camera · Audio per cut, up to 20 cuts
4 input modes — story · URL · existing video · all of the above

Explore AIVE HomepageWeb (Next.js 16)

ImageToApp

04Step 00.5 — The App Factory

ImageToApp

One sentence of an idea becomes a working app

An AI app factory that turns a one-line idea into a spec, UI mockups, and a working app

"Build me an app" — one sentence is all it takes. The moment you type it, a 7-stage pipeline across three phases — PLAN, DRAW, BUILD — finishes planning, design, development, and verification on its own. It expands your one-liner into an app name, target audience, and core features, analyzes competitors to a real spec-document level, then an image AI draws high-resolution UI mockups for every screen in parallel. The mockups are decomposed into layout trees that become blueprints for code, producing an app where every button actually works. Runtime errors are fixed the moment they are detected, and a request like "switch to dark mode" is applied instantly. More than 36 apps have been born this way already.

IDEA EXPANSION — auto-expands a one-liner into a spec and market-research report
MOCKUP GROUNDING — decomposes image-AI mockups into layout trees that ground the code
WORKING CODE — generates finished apps where buttons, inputs, and saving actually work

Explore ImageToApp HomepagemacOS desktop (Apple Silicon)

Bob

05Step 01 — Meetings · Managing

Bob

Perfect meeting notes, with not a single character leaving your device

100% on-device AI meeting copilot — transcription, speaker ID, summaries

When a meeting starts, Bob transcribes in real time, tells speakers apart, and when it ends, the minutes and a one-line summary are already done. FunASR SenseVoice speech recognition, CAM++ speaker embeddings, Qwen3 4B summarization on llama.cpp — all of this AI runs inside your Mac with no cloud. The only kind of meeting AI you can use without worrying about security.

Real-time transcription + automatic speaker separation, re-identification via saved profiles
FunASR SenseVoice speech recognition · CAM++ speaker embeddings
Local Qwen3 4B summarization and suggestion synthesis on llama.cpp

Explore Bob HomepagemacOS desktop (Wails v2)

Kiroro

06Step 02 — Market Research · Growth

Kiroro

While the marketer sleeps, growth keeps going

24/7 autonomous performance marketing run by 27 AI agents

Just enter a product URL and a goal. 27 specialized AI agents covering market analysis, positioning, SEO, community, and X outreach build the strategy, create the content, publish it, and measure results with GA4 to optimize themselves. The Autopilot Engine runs this full cycle as an unmanned 24-hour loop — an entire marketing team inside one SaaS.

27 agents including Product analysis · Positioning · SEO · Community · X Outreach
Autopilot Engine — unmanned operations loop on a virtual day (default 60s/tick)
Pipeline graph (@xyflow) · performance chart (recharts) visualization

Explore Kiroro HomepageWeb SaaS (Next.js 16)

PerfectPixel

07Step 03 — Asset Production

PerfectPixel

One sentence becomes a living, moving pixel character

One line of text → 100+ motion sprites — the AI pixel-art studio

The moment you type “a knight in blue armor,” a character that walks, runs, and swings a sword is ready. It auto-generates an 8-direction sprite set with 100+ motions, refines them through a self-correcting quality loop and real pixelization, and exports formats you can import straight into a game engine. A studio that completes your game's visuals without an art team.

Base character + 100+ motions auto-generated from one line of text
8-direction sprite sets with engine-friendly export
Self-correcting quality loop and real pixelization

Explore PerfectPixel HomepagemacOS desktop (Wails v2)

wasd

08Step 04 — Game Development

wasd

One line of description, a playable game 60 seconds later

Text-to-game engine — write it and it's built, say it and it's fixed

Type “drone racing through a neon city” and within 60 seconds a real game you control with arrow keys opens up. A hybrid architecture combining a GameSpec JSON schema, genre templates, and constrained LLM code generation means the output runs reliably every time — and saying “make the enemies faster” in chat applies it instantly. Finished games are shared immediately with a single link.

Platformer · top-down · tower defense · puzzle · strategy — 2D Phaser templates
FPS · racing — 3D Three.js templates
Chat-based iteration (iterate via chat)

Explore wasdWeb (Next.js 16)

Game by Labs

Battle Stage

The show must never end.

On-chain tactical deck-builder TRPG — the first game built with the wasd engine

Built to prove the wasd engine could ship a real game. Step into the role of a circus manager, build your hero card lineup, and conquer five stages corrupted by the Blight.

On-chain Tactical TRPG

wasd Engine

5 Act + Otherside

Play Now View Game Details

Battle Stage

Title Screen

Overworld

Battle

Gallery

RED STUDIO

09Step 05 — Video Production

RED STUDIO

Where a prompt becomes a film — sound and all

LTX-2.3 × H100 — the AI video studio you command from your desktop

Describe a scene and LTX-2.3 on a RunPod H100 80GB turns it into 24fps video. Gemma-3-12B enriches your prompt cinematically, while the pipeline handles spatial upscaling and audio synthesis on its own. From scenario writing to batch generation and gallery management — command GPU-cluster-grade video generation from a single Mac desktop app.

Prompt → 5-second @24fps video generation (up to 768x512)
Gemma-3-12B prompt-enrichment pipeline
Spatial upscaling + audio synthesis

Explore RED STUDIO HomepagemacOS desktop (Wails v2)

StoryMaker

10Step 06 — ONE Story Creation Pipeline

StoryMaker

Write a novel, get a webtoon

Writing · characters · worldbuilding · webtoon rendering — the all-in-one AI creation studio

ONE Story is NEXUS's web novel, webtoon, and AI chat service. StoryMaker was built to implement and validate the AI creation pipeline that ONE Story needs — internally first. Co-write a 20-episode, 90,000-character web novel with AI, then turn it into a webtoon with one button. Each episode is auto-analyzed into an 18–50 panel storyboard, then rendered with speech bubbles, narration, and sound-effect overlays into a vertical-scroll webtoon.

Multi-LLM web novel writing (Claude and more via OpenRouter)
Automatic per-episode storyboard analysis — 18–50 panels
Speech bubble · narration · sound-effect overlay rendering

Explore StoryMaker HomepagemacOS desktop (Wails v2)

ONEchain Skills

11Chain — ONEchain by sentence

ONEchain Skills

One sentence, and your agent acts on-chain

A suite of 9 natural-language AI agent skills that drive ONEchain

ONEchain Skills turns nine skills — from token launches to NFT trading, in-game checkouts, on-chain reads, staking, prediction, and quests — into single natural-language sentences. Say 'buy 0.1 ONE of RUBYx' and the agent runs a ONEgametoken AMM swap; say 'show NFT collections' and it reads the ONEnft indexer. A single one-skills-suite bootstrap installs every skill into local agent runtimes like Claude Code, Cursor, and Codex CLI. Read-only skills work with no credentials; fund-moving write skills execute through explicit local or gateway signing.

ONEforge · ONEgametoken — token launches, bonding-curve trades, and GameToken AMM swaps/LP by sentence
ONEnft · ONEshop — MarketplaceV1 NFT trading and in-game package checkouts in natural language
ONEexplorer — read blocks, transactions, addresses, and contracts with no credentials

Explore ONEchain Skills Skills siteWeb · Local agent runtimes

The Core — research infrastructure built in-house at NEXUS AI Labs

The brain of the Labs,
Nexus AI framework

For AI agents to run autonomously on the trust blockchain has established, they need an orchestration layer to command them. Nexus AI is that core — an in-house framework that coordinates multiple AI backends in a single workbench and continuously turns AI judgment into proven, deployable products. It was recently rebuilt end to end around Loop Engineering — the discipline of designing not a single prompt, but the entire cycle an AI runs by itself until it reaches its goal.

da:ideada:planda:workda:verifyproducts proven

Loop Engineering autonomous loops

The agent loop — perceive → reason → plan → act → observe — keeps running on its own until it reaches a verifiable goal (Done when). A goal evaluator and stop hook judge every iteration, so completion is proven by passing verification, not by the AI merely claiming it's done.

Loop guardrails & file-based memory

Six guardrails — max iterations, token/cost budgets, stall detection, circuit breakers, completion criteria, and human checkpoints — block runaway loops, goal drift, and silent failures. Each cycle clears its context and restores state from file-based memory, so even overnight long-running loops converge without wavering.

Multi-AI backend research environment

Drives Claude Code · Codex · Gemini CLI as native structured streams and experiments with 70+ LLMs via a direct OpenRouter connection. Every event is normalized into a common StreamEvent, enabling instant cross-model comparison and analysis.

Wave DAG parallel execution

Decomposes work into a DAG, runs it in parallel across isolated git worktree sessions, and converges through wave-level merges and Gate approvals. Validate research hypotheses simultaneously — one person delivering team-level throughput.

da:* research-to-product pipeline

Runs idea exploration (idea) → design (plan) → implementation (work) → verification (verify) as an automated skill chain. One command triggers the full cycle where research becomes code and code becomes a verified product.

24/7 autonomous operations infrastructure

Three remote-control channels — E2EE mobile bridge · Telegram · SSH — plus rotation and quota management from a Go account-pool server. 'Execution is everything in the end' — the reason the Labs never stops running.

Tauri v2React 19RustGoPythonSQLite

In Action — the Nexus AI framework on screen

Key features, screen by screen

The Nexus AI framework isn't a slide deck — it's a workbench that actually runs. From the Loop Engineering dashboard to the 24/7 remote operations panel, see the key features exactly as they appear on screen.

01 — Loop Engineering

A loop that runs itself until the goal is reached

The perceive → reason → plan → act → observe pipeline runs with a live log while the Goal Evaluator rules PASS or RETRY on every iteration. A Done-when checklist, a Loop 7/50 iteration badge, the stop-hook status, and token-budget gauges — every state of the loop is controlled from one screen.

02 — Multi-AI Backend

Claude · Codex · Gemini in one workbench

Switch between three AI backends by tab, give them the same task, and swap among 70+ LLMs instantly with the OpenRouter model picker. Per-backend quota gauges and the account-pool indicator in the status bar show remaining capacity in real time.

03 — Wave DAG

Parallel like a team, safe behind the Gate

The work DAG unfolds across Wave 1 → 2 → 3 columns while agent sessions isolated in git worktrees execute simultaneously. When each wave finishes, the Gate Approval panel shows build, test, and lint checks before you approve the merge.

04 — 24/7 Remote Ops

At 3 AM, the lab is still running

Direct agents from anywhere over the E2EE (X25519 · ChaCha20) mobile bridge and a Telegram command feed, while the overnight session monitor tracks per-account cost and quota rotation — 'All systems running, 03:42 AM'.

Library