Sednium Technologies is a small software studio based in West Bengal, India. We build web applications, native Android apps, design systems, and the infrastructure underneath them. Much of what we make runs on-device, so editors, launchers, chat clients, and AI tools keep working without a network connection. The rest is lightweight web infrastructure.

The people who write the code also run it in production and answer the bug reports, which keeps the feature list honest.

• Website: https://sednium.com
• Org repos: https://github.com/Sednium-Technologies
• Contact: support@sednium.com (support, bugs), bhoid@sednium.com (product, commissions), loid@sednium.com (engineering)

• 31 October 2026: Kryptonide v4.0, shipping on the native Kotlin and Jetpack Compose client and the legacy web-runtime channel.
• December 2026: ONYXCHAT release for Android. Windows and Linux clients are in development.

A code editor and IDE for Android and the browser. The web stack runs locally through WebContainers, so HTML, JavaScript, React and Python projects execute without a server. Kotlin and Java are supported natively. The editor is Monaco, with a command palette, go-to-line, find and replace, and global search.

Projects sync two ways with Google Drive, and Git commit/push/pull runs on isomorphic-git so it works offline. Larry AI is the bundled agent: bring your own API key, and it plans against the whole project tree then applies small line-level edits instead of rewriting files.

v4.0 is scheduled for 31 October 2026 and ships on two channels: the native Kotlin and Jetpack Compose client, and the legacy web-runtime build (Monaco with Capacitor and WebContainers) that current projects are on. The legacy channel keeps receiving fixes after 4.0 so nobody has to rewrite a project to keep working.

• Website: https://kryptonide.sednium.com
• Research: https://www.sednium.com/research/kryptonide
• Source: Sednium-Technologies/krypton-ide · MIT · Releases
• Stack: TypeScript, Java, Vite, Capacitor, Monaco

Runs Minecraft: Java Edition on Android, from 1.7.2 through 1.21.4 and current snapshots. Mod loaders supported are Fabric, Forge, NeoForge and Quilt, with modpack browsing for CurseForge and Modrinth inside the app.

Different Minecraft versions need different Java versions, so OpenJDK 8, 11, 17, 21 and 26 are downloaded on demand into a sandbox per version. That keeps the JDKs out of the APK and avoids the "one runtime for every version" problem. On startup the launcher reads the device's OpenGL ES and Vulkan capabilities and recommends a renderer (GL4ES 1.1.6, Vulkan Zink, or VirGL) rather than forcing one. The game JVM runs in a separate process from the UI, so a crash in the game does not take the launcher with it.

Microsoft accounts work for Realms and public servers. Offline profiles work for LAN and singleplayer. Skins and capes are managed locally, including HD textures, with a 3D preview before applying.

• Website: https://blade-launcher.sednium.com · Docs
• Source: Sednium-Technologies/Blade-Launcher · GPL-3.0 · 292 MB universal APK (ARM64/ARMv7/x86/x86_64, JDKs downloaded on demand)
• Mod updates: MC Mod Updater for batch checking and dependency resolution
• Stack: Kotlin, Jetpack Compose (M3), C/C++ graphics translation layers

Minecraft is a registered trademark of Mojang Synergies AB. Blade Launcher is not affiliated with Mojang or Microsoft.

A browser tool for keeping a modpack current. You drop in your mods folder or a pile of .jar files and it works out what is outdated across Modrinth and CurseForge, what companion libraries are missing, and hands back one batch download. Nothing gets uploaded. Hashes are computed on your machine through the Web Cryptography API in a worker, and folder sync uses the File System Access API, so in Chrome it writes results straight back into .minecraft/mods.

Identification has two passes. A SHA-1 comparison against the Modrinth API and CurseForge is exact when the JAR is a stock download. When a file was renamed or built by hand, a semantic pass cleans the name (fabric-sodium-mc1.20.1-0.5.8.jar to sodium) and matches on project identity instead. Dependency checking runs against the whole pack rather than mod by mod, so Fabric API or Cloth Config that is already present does not show up as missing, and the one-click resolver only adds genuinely absent libraries to the queue. Alpha, beta and release channels are picked per run.

Profiles make a pack reusable. Save, clone and switch between configurations, send a link like https://mcmods.sednium.com/?profile=... that loads that exact setup in someone else's session, import standard JSON manifests, or copy a formatted mod list for Discord, Markdown or plain text. An updated pack exports as a single zip bundle.

It pairs with Blade Launcher, where the same loaders and version ranges matter, but it works standalone with any desktop setup.

• Website: https://mcmods.sednium.com
• Source: CoderBhoid/Minecraft-Mods-Updater · MIT declared in the README, LICENSE file to be added
• Stack: TypeScript, React, Vite, Tailwind CSS, Vercel

An Android messenger with no phone number and no server-side message history. The relay forwards an encrypted payload over Firebase Cloud Messaging and stores nothing, so there is no message table to hand over under compulsion.

Identity is an 8-character hex Routing ID derived from a Google account. Adding someone means scanning a QR code or pasting their ID, which removes contact syncing and SIM-swap attacks from the model.

Encryption is layered rather than invented. The inner layer is the standard libsignal Double Ratchet for forward secrecy and post-compromise security. The outer layer wraps the handshake with an AES-256-GCM key derived by HKDF-SHA256 from the classic ECDH secret concatenated with an ML-KEM-768 shared secret (BouncyCastle), which is what makes recorded traffic useless later. Each message gets a fresh 96-bit nonce from SecureRandom. Chat history is encrypted with AES-256 on device and synced to the hidden appData folder of your own Google Drive.

The client is closed source. The architecture, including the six-part S.H.I.E.L.D. design that Loid specified, is written up in full at sednium.com/research/onyxchat and sednium.com/research/shield so it can be reviewed without the code.

• Website: https://onyxchat.sednium.com
• Research: https://www.sednium.com/research/onyxchat
• Status: releasing December 2026. Windows and Linux clients are in development. Architecture notes are public.

An OpenAI-compatible gateway for running several models at once. Up to five models from fifteen providers share one endpoint, and a request starts a cycle: a leader model is elected from live benchmark scores (60% reasoning, 40% coding), it pulls current package documentation and CVE data, the other models review and argue about the result, and one merged answer comes back.

The operational parts are the reason people keep it around. Requests cascade to the next provider on a 429 or a 500. Multiple keys per provider rotate round-robin. Tool calls are checked against an RBAC policy before routing, so an agent cannot run arbitrary CLI commands or read outside allowed directories. Repeated identical tool calls trip a circuit breaker (default 3) to stop an agent burning credits in a loop. Frequently repeated queries are answered from a local semantic cache.

Provider keys never touch a Sednium Technologies server. They are encrypted into a rocky_vault.json file in your Google Drive appData folder and decrypted client-side. From the vault you can mint scoped keys for your own apps that talk to the gateway.

• Website: https://rosette.sednium.com
• Research: https://www.sednium.com/research/rosette
• Source: CoderBhoid/rosette-gateway

An AI client that runs models on the phone. Inference goes through a native llama.cpp JNI binding for GGUF weights, with Google LiteRT as a second engine, so there is no Python runtime or Node daemon in the app and no root needed. A RAM watchdog reads ActivityManager.MemoryInfo before loading a model and tells you when a quantisation is too large for the device instead of letting Android kill it.

Quantised weights can be searched and streamed from a built-in Hugging Face browser into Documents/Oorty/models/. Conversations are dual-written as plain Markdown with YAML frontmatter into Documents/Oorty/chats/, which you can open in Obsidian, Logseq, VS Code or commit to Git. Nothing is locked in a private database.

It also speaks MCP, so multi-server tool loops run with JSON schema validation, approval prompts before side effects, and local SQLite queries. Cloud models (Gemini, Claude, OpenAI, Groq, Mistral, NVIDIA NIM) route through the same UI with your own keys. No telemetry, and airplane mode changes nothing except cloud availability.

• Website: https://oorty.sednium.com
• Source: https://github.com/Sednium-Technologies/OORTY
• Stack: Kotlin, Jetpack Compose, llama.cpp, LiteRT, MCP

A headless icon API. Any icon from 200,000+ (Material Symbols, Material Design, Lucide, Simple Icons, FontAwesome and thirteen other sets) comes back as SVG from a single URL:

https://sednicon.sednium.com/api/render?q=rocket&color=ff6600&size=32&anim=spin

There is no npm package, no build step and no bundler configuration. It is an <img src>. You can also describe an icon in English and get a generated SVG back, using your own provider key, which is kept on your device and never stored on a server. Runs on Vercel Edge Functions, under 50ms, free, no rate limits for normal use.

• Website: https://sednicon.sednium.com · Library · Generate
• Docs: https://sednicon.sednium.com/docs · Status
• Source: AnkushDas4/SEDNICON · MIT (declared in README)

A reader for ten channels (Headlines, Technology, World, Business, Politics, Science, Health, Entertainment, Sports, India) plus your own RSS feeds. Article HTML is fetched, passed through Mozilla Readability and DOMPurify, and rendered with reading-time estimates and a progress bar. The audio reader uses the browser's own speech engine, so nothing is recorded or uploaded. We also serve cached RSS 2.0 endpoints for feed readers and launcher widgets.

No accounts, no analytics, no ad scripts. Theme, typography, bookmarks and custom feeds live in localStorage. It is a PWA, so the shell and saved stories work offline.

• Website: https://news.sednium.com
• Research: https://www.sednium.com/research/sedniumnews
• Source: CoderBhoid/Sednium-News
• Stack: TypeScript, Vite, Tailwind, Vercel serverless functions

An open file format for agent memory, which solves the practical problem of an agent that has to restart without re-reading forty thousand tokens of history.

State is saved as signed, block-partitioned snapshots. Retrieval is lazy: an unencrypted but HMAC-verified header index is read first, and only the blocks you ask for are decrypted. Old context degrades on a schedule (under an hour stays FP16_RAW and lossless, an hour to a day goes to INT8, anything older collapses to INT2 semantic summaries), which makes ancient sessions roughly sixteen times cheaper to store and load. Reference counts can override decay, so a variable that is still in use keeps full precision. A background optimizer decays inactive blocks and migrates older files. Encryption is AEAD, encrypt-then-MAC, with a post-quantum layer on the key wrapping.

Bindings exist for Python 3.10+ and C++17, including save_active_context.py, recall_context.py and autonomous_optimizer.py, plus direct hooks into vLLM PagedAttention blocks and llama_state_get_data() / llama_state_set_data() in llama.cpp.

• Website: https://ai.sednium.com
• Research: https://www.sednium.com/research/aif
• Source: CoderBhoid/ai-format · MIT

The tool that makes our product videos. A Node and Express service drives Remotion: a prompt goes through Gemini when a key is configured, and falls back to a local rules engine (LocalAI.js) that uses compromise for semantic routing over named style templates. The chosen template renders through Remotion with a Three.js scene to an MP4 in out/. Static files are served from public/, and background caching is there to stop assets flickering mid render. It is internal tooling and stays deliberately unpolished. The repo is public because the render setup is fiddly and anyone wiring Remotion to a Node server will hit the same problems. Research: https://www.sednium.com/research/charon Source: Sednium-Technologies/charon · ISC per package.json, no LICENSE file in the repo yet

Built on-device. Shipped in production. Supported by the people who wrote it. Website · GitHub · Research Support · Commissions · Engineering

Made in West Bengal, India · © 2026 Sednium Technologies. All rights reserved. Individual projects are licensed as stated in their repositories. Minecraft is a trademark of Mojang Synergies AB. Sednium Technologies is not affiliated with Mojang or Microsoft.