Skip to content
DeukPack brand concept - a dog carrying data packs

DeukPack — A robust contract between AI Agents and Unity

Server flexibility
Stable client performance

Dynamic JS/AI Flexibility × High-Performance In-place Reuse Performance

A hybrid schema engine for AI agent automation.

In an era where AI holds data sovereignty, DeukPack consolidates fragmented IDLs into a single 'semantic spine'. In Node.js servers, slice and dice JSON objects freely. DeukPack uses schemas (.deuk) as filters to pack only the necessary fields into a compact binary.

On the Unity client, this binary is instantly reassembled via In-place Unpack, significantly reducing garbage (GC) by reusing instances and containers. An efficient asymmetric compromise of flexibility and performance—this is the core architecture of DeukPack.

Zero-Alloc C# Parser Dynamic JS/TS Masking AI-Native JSON Wire MCP-based AI Guardrail
Use the core now Products overview DeukPack kits lineup

Language switcher (한국어 ↔ English)

  • Korean: Site root. English: /en/… prefix.
  • Switcher: Top-right language menu.
  • GitHub README: README.md · README.ko.md.

🚨 [MUST READ] Core Architecture Consolidation & Migration Notice (v1.9.0)

The core module structures for each language have been completely optimized and unified into DeukPackCodec!

  • Ultra-Simple Unified API: No more verbose factories. All languages now intuitively use the struct-based Hero.Pack() and Hero.Unpack() (2-Method) syntax.
  • ⚠️ Caution (C# / Unity Users): If you have been manually copying .cs runtime files without using UPM, you MUST delete all existing runtime folders before copying the new code to prevent name collisions.

⚡ What it looks like (At a glance)

1. Import OpenAPI (or write .deuk IDL) 

// Feed your existing OpenAPI (Swagger) to the CLI, OR write a clean IDL:
namespace Dto

struct Hero {
    1> int32 id
    2> string name
    3> float hp
}

2. Server (JS/TS): Direct POJO Serialization 

// pack() → binary (default). pack(obj, 'json') → JSON string.
const payload = Dto.Hero.pack({ id: 1, name: "Arthur" });
network.send(payload);

3. Client (C# Unity): In-place Read & Override Write 

Dto.Hero cachedHero = new Dto.Hero(); // Allocated ONCE, never again

void OnNetworkMessage(byte[] inputData) {
    // Optimized Unpack — overwrites existing instances and lists to minimize GC
    cachedHero.Unpack(inputData);
    Debug.Log($"Hero: {cachedHero.name}, HP: {cachedHero.hp}");

    // Pack to binary (default) and send — no new allocations
    cachedHero.hp = 99f;
    byte[] outputData = cachedHero.Pack();
}

Product lineup

Product One-line Definition Intro Doc
DeukPack Core · Engine Mixed-IDL Hybrid Serializer: Next-gen engine that accepts any IDL and extracts both AI semantic meta and multi-lang code. CORE_ENGINE →
DeukPack Protocol Messaging Runtime: Real-time communication engine providing high-performance serialization and AI guardrails. PROTOCOL →
DeukPack Excel Add-in Schema-Driven Editor: Excel tool for validating and syncing planning data based on schemas. EXCEL_ADDIN →
DeukPack Pipeline AI-Native Pipeline: Automated toolchain from builds to custom code generation in one command. PIPELINE →
DeukPackMcp Universal AI Hub (Coming Soon): Unified bridging of .proto, .deuk, OpenAPI and more as AI skills. MCP_HUB →
DeukNavigation Unity/server shared Deuk wire (.dpk) based Recast/Detour NavMesh solution. NAVIGATION →
Extension products DB migrator, EF support, Google Sheets, Unreal/Elixir, and more. EXTENSIONS →

→ See the Architecture diagram for data/schema flow between products. Details in Products overview and each product page. Stack-specific ready-to-use examples in DeukPack kits lineup.

About this page

deukpack.app is the canonical source for DeukPack brand, products, tutorials, and reference.

The free, open-source core is available on npm and GitHub (joygram/DeukPack).

For hands-on labs or the "Ruins of the Beginning" saga journey, go to kits.deukpack.app. The repo README, this site, and the kits site show the same product at different depths. See Documentation index for the relationship table.

To use the core now, click "Use the core now" above or go to Core · engine. That page links to install/CLI, this site's install/quickstart/IDL/C#/C++/pipeline/kits lineup, and GitHub (README, releases).

Source, issues, and contributions: GitHub repository. Linking and citation are free for both commercial and non-commercial use.

🚀 How is it different from existing tech? (Key Features)

These are the three main features of DeukPack compared to Protobuf, FlatBuffers, and JSON.

  1. Zero-Copy Override without Allocation
  2. Legacy limits: Protobuf or Thrift requires allocating (new) a separate DTO class every time you want to mask or override specific fields for different users.
  3. DeukPack advantage: With just ONE generated C# object, a single Write(oprot, overrides) function generates variable payloads without GC (GC).
  4. Direct POJO Processing without Classes
  5. Legacy limits: Even in JS environments, they force heavy class wrappers like new Message() or complex offset Builder patterns like FlatBuffers.
  6. DeukPack advantage: Simple JavaScript objects ({ id: 1 }) thrown by AI are immediately packed into byte arrays. Flexible in the JS environment, statically typed on the client.
  7. 'AI Semantic Spine' beyond a simple spec sheet
  8. Legacy limits: Traditional IDLs are basic byte conversion mappings for communication.
  9. DeukPack advantage: The .deuk file natively supports extends (multi-level inheritance) and tablelink (DB/Meta table references). It acts as the most condensed Knowledge Graph (RAG context) that allows AI (LLMs) to grasp the ERD relationship of an entire project at a glance.

🛡️ Security & Reliability (OOM Defense / Anti-DDoS)

As a fundamental infrastructure component that constantly parses external network byte payloads, DeukPack implements defense mechanisms against network-layer parsing vulnerabilities (OOM, Buffer Flooding, Infinite Recursion).

  • Universal OOM (Out-of-Memory) Defense (v1.7.0+): For all supported engines (JS, C#, C++, Java, Elixir), DeukPack enforces validation boundaries during streaming—MAX_SAFE_LENGTH (10MB) and MAX_ELEMENT_COUNT (1,000,000) limits. Even if malicious clients transmit artificially pumped list headers, packets are discarded (Fail-Fast) before any memory allocation takes place, preventing process crashing.
  • Progressive Chunk Validation (JSON Flood Defense): Completely replacing legacy ReadToEnd() methods, stream endpoints now implement length pre-evaluations, neutralizing both giant string bombs and multi-bracket JSON stack flooding vulnerabilities in Node.js and Java backends.
  • Continuous DDoS Fuzzer Suite: Integrated seamlessly into the CI pipeline, the test-fuzz-oom.js automation injects parsers with 2GB+ abnormal buffers and unbounded structural trees to verify resilience.

Supported Protocols & Languages (v1.9.0)

Language / Platform Pack (.dpk) TBinary TCompact TJSON JSON (Wire) YAML / CSV Protobuf OpenAPI MCP In-place / JIT
TypeScript / JS ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ (v1.5) ✅ (v1.6)
C# (.NET / Unity) ✅ ✅ ✅ ✅ ✅ ✅ (v1.2.7) ✅ ✅ - ✅
C++ ✅ ✅ ✅ (v1.5) - ✅ (v1.5) 🚧 ✅ - - ✅ (v1.4.2)
Java ✅ ✅ ✅ (v1.5) ✅ (v1.5) ✅ 🚧 - - - 🚧
Elixir (BEAM) ✅ (v1.7) - - - - - - - - ✅ (BEAM)
Excel (Add-in) ✅ - - - - - - - - -
Python ✅ (Pure/Rust) 🚧 - - ✅ - 🚧 - - ✅ (v1.9)
  • ✅: Full Support / Production Ready
  • ⚠️: Preview / Partial Support
  • 🚧: Pilot / Development in Progress
  • -: Not Currently Supported

⚡ Performance & Benchmarks

Notice: Benchmark Data Under Revision We have identified an inconsistency between the highly-optimized protocol implementations and the aggregated result table. The performance metrics matrix and cross-language benchmarking details are currently being revised and will be republished once the evaluation is complete.

To ensure objective performance testing, we have adopted industry-standard benchmarks such as BenchmarkDotNet, mitata, and pytest-benchmark. Achieving 60-100% reduction in memory allocation and 2.5x to 5x faster parsing speeds compared to legacy industry formats. If you find any errors or have suggestions regarding the benchmark scenarios, please let us know.

Next steps