things i've built

on-device LLM inference for iOS/iPadOS/visionOS — llama.cpp/Metal, MLX, and Apple Foundation Models behind one protocol. 24 modules, streaming, KV-cache snapshots, MCP client, 14 built-in tools.

the android sibling — AICore/Gemini Nano for zero-download inference plus an ExecuTorch runtime for our own weights, behind one InferenceBackend. scaffolded blind before the pixel even arrived.

chatgpt-style web ui for LM Studio. streaming, markdown, model auto-detect, keyboard everything. the whole thing is vanilla js components — no framework of the month.

openwebui-style ui but for embedding models. paste text → vectors. playground with cosine-similarity heatmaps, batch embedding, history — local via LM Studio or cloud via 4 providers.

a mitmproxy that TLS-terminates only Google's model host and reroutes gpt-oss-120b requests to a local LM Studio. slick, brittle, definitely against the tos. full revert path included.

a premium AI product shell — one surface, multiple serving pools, explicit routing, unified usage budgets. tiers, lanes, modes; a calm dark-first chat/code/agent/browser experience.

a real arch linux + xfce desktop in a sandbox, driven by claude over one multiplexed ssh. screenshots in, mouse/keyboard/shell out. 40 tools. your own personal operator, but self-hosted.

production-minded starter for browser-use + computer-use agents. dockerized xfce/noVNC desktop as the only computer boundary, per-session workspaces, replayable runs, policy layer. the model never touches the host.

a research agent that drives a real chromium like a person — parallel tab swarm, universal model gateway (anthropic → LM studio), and writes a source-cited markdown report to your desktop.

hardened mcp server for local LLMs. web search through your own searxng, but every filesystem/exec tool runs in a read-only, no-network, non-root docker sandbox. strict allowlist, path-confined workspace.

an MCP server that lets claude code summon subagents on literally any non-anthropic model opencode can reach — gpt, gemini, free models. parallel second opinions, with a cost/token footer so you know what you just burned.

a dirt-simple local MCP message bus so claude code and gemini cli can leave each other notes. stdio, sqlite, zero network, zero auth. two AIs politely passing the keyboard.

pits claude against ARC-AGI-3 using the official claude code harness — an MCP server exposing the prize games as tools. 64×64 color grids, compact hex frames, the model plays until it wins or game-overs.

a pocket ecosystem that evolves in your terminal. creatures design themselves by dying less than their neighbors — five genes, mutation, selection. predators, grazers, and arms races aren't coded, they just happen. one file, no deps.

a docker dev box with four AI CLIs (codex, gemini, qwen, claude) preinstalled. persisted home volume, auth-callback bridge for browser logins, preview watcher. an IDE-less sandbox where agents live.

a local-first bloons td 6 AI pilot. a melonloader mod in the game speaks a tiny action protocol; an outside bridge owns the AI auth and returns validated JSON. codex/opencode/local modes, bad output becomes noop, not game chaos.

apocalypse web search. while online, hoard search results + readable pages into a local index. offline, hybrid search (FTS5 + embeddings fused by RRF) returns your cache with a grounded AI overview from LM Studio. plus ZIM archives for the big picture.

a personal single-user discord selfbot. ~70 commands from tic-tac-toe to zalgo text, an `ai` command that routes cloud↔local, local embeddings, a snipe cache, afk, and a private loopback dashboard. everything risky defaults off.

a fork of the archived discord.js-selfbot-v13 with 49 new client methods — platform spoofing, quests, account standing, soundboard send-side, a voice watchdog, and an anti-detection client-properties scraper.

a terminal UI that finds live blooket games by scanning 000000–999999. launches a real chrome to pass cloudflare's JS challenge, then hammers the games API with concurrent workers. space to pause, c to clear.

a meteor client addon for 2b2t that hunts player bases. torch density, chest clusters, underground air pockets, nether highways — every heuristic scores each chunk, renders boxes, and logs it. god help whoever you find.

a native macOS trading helper for pet simulator 99. catalog browser, W/F/L calculator, and gemini-backed buy/sell calls. inventory via screen capture + OCR, read-only, and a recommender that knows RAP is manipulable.

two agents, kai and albert, learn tag from scratch via self-play PPO with zero semantic labels — just observations and rewards. a local web dashboard with replays, checkpoints, and an export button.

a tiny macOS app that spams one key — but only while a target app is frontmost. real CGEvent key codes so games actually pick them up. parallel threads, a menu-bar hotkey, and a nanosecond-danger mode that's labeled dangerous for a reason.

an original deltarune chapter-1 expansion: an optional 1,550 hp observatory boss (eidolux), two new enemies, three bullet phases, full mercy + combat routes, original art. safety model: the steam install stays read-only, only a cloned copy gets patched.

an iOS terminal for iPhone 11-class devices, built with no SDK headers — hand-declared runtime, POSIX, sockets, and Network.framework, packaged into an IPA with a dylib of unix tools. plus a dropbear ssh server, because why not.

self-hosted rustdesk rendezvous + relay in a compose file. for the friend who refuses to let teamviewer exist. includes the whole "tell your friends how to connect" script.

a research project for running steam windows games on macOS without app-store builds. wine/proton runtime management, DXVK/MoltenVK, a steam compatibility-tool shim, and an honest "why not a kext" doc. among us runs. sometimes.