Everything I've messed with.

MCP server driving Blender with an LLM: scene edits, Poly Haven/Sketchfab pulls, and my self-hosted Hunyuan3D on Modal as the image-to-3D backend. Keeps lab-scene.blend in sync with the live site.

Anthropic's CLI/agent harness itself — the environment driving all of my work, with skills, plugins and the deferred-tool toolset.

DOM-aware Chrome-extension MCP (navigate, read_page, javascript_tool, computer) paired to my 'Browser 1' device for web-app interaction and testing.

Claude Preview MCP (preview_start, preview_screenshot, preview_eval) for launching and inspecting a running web app in-session; I allowed preview_start in multiple projects.

Local 'codex-bridge' MCP (codex-mcp-server) registered in the Ebay Automator project so Claude can hand work to OpenAI Codex, my collaborating agent.

Anthropic computer-use MCP for screenshotting and driving the macOS desktop (mouse/keyboard) on native apps I can't reach via a dedicated connector.

Figma Dev-Mode MCP for pulling design context, variables, screenshots and Code Connect maps into code; registered as a Claude plugin and manually added/removed via `claude mcp add figma https://mcp.figma.com/mcp`.

Firecrawl plugin (web crawl/scrape + CLI skill) enabled in my global Claude settings for site harvesting.

Gmail connector for searching threads, managing labels, and creating drafts directly from Claude.

Google Calendar connector for listing/creating/updating events and suggesting times, feeding the voice-operator's calendar features.

Google Drive connector for searching, reading and creating files; I invoked google_drive_search during wedding planning.

Higgsfield generative-media MCP (generate_image/video/audio, motion_control, virality_predictor) I planned to use to finish the Gene 515 trailer; also seen as balance/job_display calls in his session.

Hugging Face Hub MCP used to search models/datasets/papers and query repo details (hub_repo_search, hf_doc_search, paper_search); I used it from the Ebay Automator project.

Connector-discovery MCP (search_mcp_registry, list_connectors, suggest_connectors) I used to find and suggest new MCP servers.

Notion connector I used heavily as the voice-operator project hub and wedding planning DB — search, create-pages, update-page, fetch on shared Notion workspaces.

PDF Tools MCP for filling, signing, merging, splitting and extracting PDFs (form fill, signature zones, CSV bulk fill).

Pencil app MCP — the only server registered in claude_desktop_config.json (stdio binary) — edits encrypted .pen design files for web/mobile mockups.

Pinecone vector-DB MCP (search-records, upsert-records, create-index-for-model, rerank) registered as a Claude plugin; backs the planned 'kennybot-memory' index of call-transcript embeddings.

PlayCanvas Editor MCP (entities, components, scripts, materials, asset store) for LLM-driven 3D scene building; the @playcanvas/editor-mcp-server source is checked into the repo.

Scheduled-tasks MCP (create/list/update_scheduled_task) for running recurring Claude agents; coworkScheduledTasksEnabled is on in his Desktop config.

Sonatype component-intelligence MCP (getLatestComponentVersion, getRecommendedComponentVersions) for vetting/upgrading dependency versions; enabled as a Claude plugin.

Google Stitch MCP for generating app screens/design systems from text; I built a custom stdio proxy (stitch-proxy.mjs using @google/stitch-sdk) under [redacted]

Supabase connector (execute_sql, apply_migration, list_projects, deploy_edge_function) I used to run SQL and migrations against his Postgres projects (kenmacy.com, wedding, Sci-non-fi).

Todoist MCP server (@abhiz123/todoist-mcp-server) registered globally in [redacted] with a live API token, powering the ADHD-coach task workflows.

Vercel deployment MCP connected in my session; he used list_teams/list_projects/list_deployments/get_project and domain-availability checks to manage his web app deploys.

MIT-licensed full-song-with-vocals diffusion model run on Modal via diffusers' AceStepPipeline; faible's commercial-safe replacement for instrumental-only MusicGen for bands that want sung tracks.

Claude Sonnet 4.5 runs the live operator/manager/coach voice personas and call assessment; this very task runs on Opus 4.8.

Meta's EnCodec neural audio codec loaded from HF Hub; dnb-brain's corpus harvester turns YouTube DnB tracks into compact EnCodec tokens (10k h ~ 14GB) for the from-scratch pattern model.

Fable is the second AI-video tool (with Higgsfield) I plan to finish the Gene 515 trailer with.

Kuaishou's AI video model listed among the generators in my movie-director toolset.

Self-supervised music-understanding model from HF Hub; dnb-brain extracts MERT embeddings as a third ingest stream alongside EnCodec + SNAC tokens for each harvested track.

Llama-3.2-1B (Unsloth mirror) pulled from HF Hub on Modal; MisoTTS uses its AutoTokenizer as the text frontend for the cloning pipeline.

Meta's text-to-music model run on Modal GPUs; faible generates instrumental fake-band clips and grows seeds section-by-section via audio-prompted continuation (weights CC-BY-NC, placeholder/pre-commerce).

MisoTTS 8B text-to-speech, deployed by I on Modal (~4.5x realtime on A100) via an ungated-tokenizer workaround for the voice-clone project.

Miso Labs' 8B zero-shot voice-cloning TTS, served on Modal with a browser test UI; clone a voice from one reference clip + transcript (live at miso-tts-misotts-web.modal.run).

Kyutai's Moshi audio stack; MisoTTS loads its checkpoint through moshi.models.loaders on Modal as the neural-audio backbone for voice cloning.

ylacombe's LoRA fine-tuning wrapper for Meta MusicGen-melody; dnb-brain's train_jungle.py runs it on a Modal A100 (~15-40 min) to fine-tune on my jungle/DnB seed tracks (run2 = Lighter/Circles lane).

Google's image-gen/editing model referenced as a generation option in my ComfyUI/assembly-line notes.

Google's top-tier image model exposed through Higgsfield as id 'nano_banana_2'; faible's Studio worker default for album covers/merch and shot-list keyframes (always passing the cover via --image).

AI video-generation model referenced among my text-to-video tools for the movie-director / trailer workflows.

Image-to-video model used via Higgsfield (`seedance_2_0 --start-image`); faible animates album-cover stills into 5s music-video clips, and it's gen1515's video baseline for the trailer reboot.

Multi-scale neural audio codec from HF Hub; dnb-brain co-tokenizes the corpus with SNAC (hierarchical ~11Hz structure layer) and trains a 16M from-scratch SNAC-arm model (s1.5 bake-off).

Stability Amy SDXL run on Replicate; faible's template-based album-art generator uses it for img2img when a cassette/template exists, else text2img.

Airtable API key ([redacted]) as an optional data backend for spacetable.

Alibaba Bailian/DashScope OpenAI-compatible endpoint serving qwen-plus as the agent LLM for the Mirofish social-media simulation.

Anthropic SDK (@anthropic-ai/sdk) powering call postmortems and live operator/manager/coach personas (Sonnet 4.5), and used in spacetable, familycouncil and todoist-adhd-coach.

ElevenLabs API for NPC voices and SFX in emojimo, and the basis for several voiceover/music skills I ha installed.

ElevenLabs realtime voice-agent WebSocket; voice-operator's elevenlabs_bridge.py pipes a live SIP call into wss://api.elevenlabs.io/v1/convai/conversation as an alternate cloud voice brain (reusing the KennyBot agent).

ElevenLabs text-to-sound API (api.elevenlabs.io/v1/sound-generation); emojimo pre-generates per-biome/per-room ambient loops and NPC sound effects ([redacted]).

fal.ai inference endpoint (fal.run); voice-operator generates top-tier zine/cover art via fal-ai/flux-2-pro, with an aspect-ratio mapper to fal's supported sizes ([redacted]).

Frame.io token ([redacted]) wired as an optional media-asset integration in the spacetable project.

Google's Gemini image model ('Nano Banana'); wired into voice-operator's zine image chain via generativelanguage.googleapis.com, though my free-tier quota is 0 so it falls back to Pollinations.

Gemini API key gates optional cloud image generation (WALK frames, AI redraw) in the gorest spritesheet generator.

Generative image+video platform driven via its CLI (and MCP); faible's Studio polls jobs and shells out to `higgsfield generate create ...` for covers, merch, stills and music-video clips, gen1515 connects it as an MCP for trailer generation.

Kroger's product + cart API (banner King Soopers in Colorado); voice-operator's grocery scanner resolves UPCs via filter.productId and the chef persona attempts cart-add (PUT /v1/cart/add, currently silently failing).

Kroger Developer API used by the always-on Jetson grocery scanner for product lookup and cart-add (chef cart-add still silently failing); my banner is King Soopers in Colorado.

Autonomous AI research agent; gen1515 did substantial canon/world research in Manus (and ChatGPT) that I need to recover to set real dates for the trailer reboot.

Image-to-3D service in the character pipeline: Higgsfield likeness in, mesh out, then Mixamo rigs it. The campfire figures and giant dancing Ken started here.

Microsoft Teams app credentials ([redacted]) referenced as an optional integration in spacetable.

AI image generator; gen1515's established v1 look came from Midjourney reference boards (on the hard-drive archive) that the reboot must stay continuous with.

Free weather forecast API the ESP32 weather-for-mimi device polls to drive its NeoPixel/OLED display.

OpenAI API used for NPC dialogue and text generation across faible, emojimo and the todoist-adhd-coach (openai npm dep + [redacted]).

OpenAI chat-completions; faible uses gpt-4o-mini for band/artist profile copy and emojimo gates NPC dialogue to gpt-4o-mini / gpt-3.5-turbo with per-model cost tracking.

OpenAI text-to-image; faible's fallback album-cover/image generator (1024x1024) when no Replicate/SDXL template exists.

OpenAI embedding model; emojimo's NPC memory RAG embeds each player message then does vector similarity search over stored memories in Supabase pgvector.

OpenAmy image-generation model; the first link in voice-operator's zine/cover image-gen fallback chain, chosen for strong typography-in-image (1024 size map).

OpenAI Realtime (µ-law passthrough, ~380ms) is the cloud 'realtime brain' for voice-operator personas, sidestepping the Jetson's 8GB memory wall.

OpenAmy low-latency speech-to-speech WebSocket model; voice-operator's realtime path runs phone personas over wss://api.openai.com/v1/realtime (~380ms, mu-law passthrough) as the cloud brain that sidesteps the 8GB Jetson wall.

OpenAmy hosted STT; used as the input-transcription model inside voice-operator's OpenAI Realtime session so post-call audits have a text transcript.

Pixel-art generation API; voice-operator's scene_renderer.py POSTs scene descriptions to api.pixellab.ai/v2/create-image-bitforge and converts the result to 1-bit OLED wizard-scene frames ([redacted]).

Free no-key image-generation service; the final silent fallback in voice-operator's zine centerfold + faux-ad image chain when Gemini/FAL/OpenAI are unavailable.

Replicate (r8_ tokens, replicate npm dep) for hosted model inference, wired as an optional image/AI backend in faible.

Resend email API ([redacted], resend npm dep) for transactional mail in the wedding site and Sci-non-fi.

SerpAPI ([redacted]) for programmatic Google-search results in Sci-non-fi.

Recipe/ingredient API behind the Red Phone chef: recipe-to-pantry diffing that turns tonight's dish into a Kroger cart.

Stripe (publishable + secret keys, stripe npm dep) for payments in faible.

Tavily search API ([redacted]) used for research/web-search in the Sci-non-fi project.

Todoist REST API ([redacted]) backing the ADHD-coach task system, also exposed via the Todoist MCP.

Image-to-3D generator I use for quick scene props; the lab's beach parasol set is a Tripo mesh (after taming its baked emissive).

Twilio (auth token + webhook base URL) referenced for phone/SMS webhooks in the Codex hub-planning research harness.

AI music-generation service; named in gen1515's original toolchain (v1) as the source of the trailer's music before the from-scratch/Fable reboot.

UPCitemdb is the fallback UPC lookup in the grocery scanner chain after an exact Kroger productId match fails.

USDA's free nutrition database API; chef wires it ([redacted]) to provide real per-ingredient macros instead of the honest stub.

Fei-Fei Li's world-model lab; Marble turns images into explorable 3D worlds. Experimenting with it for the immersive skybox rooms in the 3D lab.

Open diffusion music model (ACE-Step v1.5 XL-Turbo) I run on Modal to generate full songs WITH vocals for Faible fake bands, replacing instrumental-only MusicGen.

Alibaba's Qwen2.5 open weights run locally via Ollama on the Jetson; voice-operator's offline persona brain (3b for story personas at ~20 tok/s, 7b for the operator) when not in Claude/Realtime mode.

384-dim sentence-embedding model I run locally for the voice-operator knowledge/constellation layer, with a transformers+torch fallback path.

Meta's MusicGen (via AudioCraft) fine-tuned with LoRA on Modal for dnb-brain jungle generation and faible's modal_musicgen song service.

BAAmy 384-dim/120MB embedding model I use for idea-factory's RAG retrieval, planned to run as ONNX on the Jetson.

Chatterbox TTS — an earlier voice-operator branch; not installed on the Jetson because it OOMs in 8GB.

Conditioning nets steering Stable Diffusion in my ComfyUI stack: OpenPose and depth for SD 1.5, plus QR Code Monster on SDXL for the hidden-image illusion posters.

My own Ollama build: qwen2.5:7b wrapped in a Modelfile with a task-intelligence system prompt - the fully on-device brain that triages a creative studio's chaotic workload in my creative-ops experiments.

Descript's high-fidelity neural audio codec earmarked in dnb-brain as the fidelity-upgrade tokenizer over EnCodec for the from-scratch DnB model.

Depth-estimation model (depth-anything-large) in my local HF cache; generates the depth maps that feed ControlNet depth passes in the ComfyUI ai-assembly-line pipelines.

Higher-fidelity neural codec earmarked as the EnCodec upgrade path in the dnb-brain tokenizer pipeline.

Distilled Whisper variant (~6x faster than large-v3) chosen as the default ASR model in conversation-civilizer's transcription pipeline for cost/speed.

my own tiny neural nets (~5k-param bass brain on 2,583 mined patterns, ~3.4k-param drum brain on 3,045 onset grids) that compose basslines and place drum ghosts in-browser for the DnB radio.

Meta EnCodec (32kHz) tokenizes DnB audio into tokens for dnb-brain's from-scratch corpus engine (yt-dlp -> EnCodec-32k -> sqlite, ~10k h ~= 14GB).

Flow-matching zero-shot voice-cloning TTS listed alongside XTTS-v2 as a candidate for the higher-fidelity Sagan voice clone in voice-operator.

faster-whisper (base.en) is the CPU STT fallback on the Jetson when whisper.cpp+CUDA isn't used.

Diffusion image model used in my ComfyUI ai-assembly-line image pipeline.

Tencent's open image-to-3D model, self-hosted on Modal (~$0.05/gen) as the 'local api' behind Blender MCP. Made the Mall Walkers storefronts and lab props.

Kokoro v1.0 (kokoro-onnx, GPU via onnxruntime-gpu) is the smoother fallback TTS on the Jetson with am_ voices; preferred over Piper.

Kokoro v1.0 ONNX voices run on the Jetson GPU (onnxruntime CUDA); voice-operator's preferred narrowband TTS for operator/manager/security and the am_ voice pool.

Meta's open LLaMA family — referenced as the chatbot model in my archived SkyPilot llm/llama-chatbots examples and as the backbone lineage of MisoTTS/Moshi.

Llama-family models and the llama.cpp runtime referenced as local inference options across the Jetson/voice projects.

Self-supervised music understanding model (m-a-p/MERT-v1-95M) I extract embeddings from in dnb-brain's triple-ingest pipeline for quality-control and the codec bake-off.

Kyutai's streaming neural audio codec used as the audio tokenizer/decoder inside MisoTTS 8B (and Moshi) in my voice-clone Modal app.

Kyutai's full-duplex speech-to-speech foundation model I deployed as a Modal websocket service (quillman fork) for real-time conversational voice.

Kyutai's full-duplex speech-to-speech model running in QuiLLMan on Modal for near-instant bidirectional voice chat.

Meta's text-to-music model (musicgen-medium / musicgen-melody) I run on Modal A10G to generate fake-band instrumental clips in Faible and LoRA-fine-tunes on jungle/DnB seeds in dnb-brain.

Meta MusicGen-melody, fine-tuned with LoRA on 63min of jungle/DnB on Modal, generates DnB clips for the dnb-brain radio project.

Fast TensorRT segment-anything model cloned for the eBay scanner to replace slow rembg (weights were blocked on Google Drive).

The underlying open ASR model family (base.en, distil-large-v3, large-v3) I run locally via faster-whisper/whisper.cpp for phone transcription and conversation diarization-alignment.

Piper TTS (per-persona ONNX, ~8x realtime on Jetson CUDA) is the priority-1 voice synth for professor/hal personas.

pyannote speaker-diarization-3.1 (gated, needs HF token) powers diarization in the conversation-civilizer service.

Open speaker-diarization toolkit (pyannote 3.1) I run in conversation-civilizer to segment who-spoke-when, then aligns the speaker turns against Whisper via IoU.

Alibaba's Qwen2.5 (1.5b/3b) is my main local LLM via Ollama for phone personas and the eBay object-identifier.

Alibaba's open Qwen2.5 instruct models — qwen2.5:7b is the voice-operator's default local LLM via Ollama, qwen2.5:3b runs Crisis 'her.py' and the conversation-civilizer/idea-factory extractors on Modal/Transformers.

Qwen2.5:3b runs the story personas on the Jetson via Ollama (~20 tok/s); Qwen3 variants were benchmarked and rejected for live calls due to chain-of-thought latency.

Meta's promptable segmentation model I use on Modal A10G in the Ebay Automator 3D scanner to mask turntable frames (~100ms/frame), with a TensorRT export path to the Jetson.

The Conversational Speech Model architecture (Llama backbone + audio decoder) that MisoTTS 8B is built on, which I stood up on Modal for voice cloning.

Tiny (~2MB ONNX) neural voice-activity detector that drives turn-taking and barge-in in the voice-operator phone AI and gates recording in curse-collar.

Multi-Scale Neural Audio Codec (snac_32khz) I ingest in parallel with EnCodec in dnb-brain's triple-tokenizer harvest for the S1.5 codec bake-off.

Open text-to-image diffusion model I ha run via SkyPilot example pipelines in the archived pushmatrix work.

Stable Diffusion XL checkpoints driven through the ComfyUI assembly-line for AI artwork.

whisper.cpp with a CUDA HTTP server (ggml-small.en) is the Jetson STT, transcribing 30s of phone audio in ~1.7s.

Coqui's multilingual voice-cloning TTS named as the planned upgrade path to properly clone Carl Sagan's cadence for the voice-operator Sagan persona.

Meta's PyTorch audio-generation framework (the MusicGen/AudioGen training+inference stack) noted as the Phase-1 training dependency in dnb-brain.

CAMEL-Amy OASIS social-media agent-simulation framework that drives the Mirofish social-network simulation.

Fast Transformer inference engine that ships the CUDA aarch64 wheels powering faster-whisper on the Jetson for the curse-collar project.

Stem-splitting model referenced in faible's band flow for separating vocals/instruments from source audio.

R3F's helper kit; its transform <Html> is how real DOM text rides the lab's wood blocks, plus useGLTF/useAnimations for every character.

TypeScript SQL ORM (with drizzle-zod/drizzle-kit) used for schema and migrations in the faible/FableGroove music-label web app.

Node web framework behind faible's API/session server and spacetable's backend.

Async Python web stack serving the voice-operator dashboard (:7070) and the eBay scanner web UI.

CTranslate2-backed Whisper used as the CPU STT fallback in voice-operator and the live STT in curse-collar / eBay scanner.

Python web framework (with flask-cors) serving the Mirofish simulation backend API.

Morphologically realistic Drosophila simulator (NeuroMechFly v2) I run in the McFly project to drive an autonomous 3D fly with CPG/hybrid locomotion controllers, streamed to a browser.

The low-level tensor library underlying whisper.cpp/llama.cpp; the Jetson whisper-server runs a resident ggml-base.en.bin model on the GPU via libggml-cuda.

Animation library (with ScrollTrigger) wiring scroll progress into the lab's camera rig and reveal choreography.

CUDA Gaussian-splatting trainer I use on Modal to reconstruct 3D objects from known turntable poses (Ebay Automator) and walkable rooms from home video (memoryspace / betacam-splat).

my primary model-loading library — AutoModel/AutoProcessor for MusicGen, EnCodec, MERT, Qwen2.5 and the MisoTTS backbone across Modal apps and the dnb-brain corpus engine.

Python GPIO library for reading hardware triggers/pins on the Jetson in the curse-collar project.

Embedded vector database (via the 'vectordb' npm package) powering the long-term memory store in the Todoist ADHD coach.

The smooth-scroll engine under the 3D lab: one app-root instance paces the camera ride down the block tower, hands off to the breakout, and restarts on release.

Audio analysis library I use for BPM tagging in dnb-brain's ripping/prep phase and chunking audio for MusicGen training.

C/C++ inference engine for GGUF-quantized LLMs; the planned Jetson runtime to run Qwen2.5 (GGUF + CUDA) on-device for idea-factory and voice-operator's three-brain target.

The physics engine underlying FlyGym/NeuroMechFly that I run (MUJOCO_GL headless) for the McFly autonomous-fly simulations.

React framework behind kenmacy-portfolio (v16), Sci-non-fi, karaoke/karoshi, the wedding site and other I web apps.

Node SMTP mailer used by spacetable to send email notifications from its Express server.

Cross-platform inference runtime (onnxruntime / onnxruntime-gpu) I use to run Piper, Kokoro, Silero VAD and bge-small as ONNX models on Mac CPU and Jetson CUDA.

onnxruntime-gpu (aarch64 wheel from Ultralytics) gives the Jetson GPU-accelerated Piper/Kokoro TTS.

Point-cloud and mesh processing library used in the eBay scanner to derive object measurements from reconstructions.

Headless OpenCV handles camera capture and frame processing on the Jetson eBay scanner.

Phaser is the client game engine for marine-arena/blipwars, paired with an authoritative Node ws server.

PDF text-extraction library in the Mirofish backend for ingesting documents into the simulation.

The core deep-learning framework underpinning nearly all of my on-device/Modal model work — MusicGen, ACE-Step, EnCodec/SNAC/MERT tokenization, MisoTTS, Moshi, and the dnb-brain training scripts.

Background-removal library (U2Net model) the eBay scanner uses to cut product photos before listing.

Embedding library I use to encode text for semantic search — MiniLM for the voice-operator constellation and bge-small for idea-factory's grounded responder.

all-MiniLM-L6-v2 embedding model powering the voice-operator knowledge layer's semantic search over databanks.

MiniLM sentence embeddings run locally; voice-operator computes constellation/databank embeddings (with sqlite-vec) for the 3D knowledge graph.

silero-vad detects speech/silence turns on the phone (silence threshold tuned to 2.5s for ADHD pause tolerance).

SQLite extension for on-device vector search I pair with MiniLM embeddings to back the voice-operator constellation/knowledge recall without a separate vector DB.

Web-based TidalCycles live-coding music environment; dnb-brain's radio embeds a vendored Strudel.js engine ('DJ_Dave style') to schedule the two neural brains' patterns over sampled breaks/bass in-browser.

NVIDIA's GPU inference optimizer named as the Jetson deployment target for the SAM-2 masker in the Ebay Automator 3D scanner.

Three.js renders the voice-operator constellation graph and powers the pushmatrix-archive WebXR experiments.

Bruno Simon's deep-dive three.js course — the reference for the kenmacy.com 3D-room build.

Dan Greenheck's paid drop-in sky system for three.js (TSL/WebGPU + WebGL): raymarched volumetric clouds, physically-based atmosphere, day/night cycle with moon and stars, god rays, live cloud shadows, and per-frame sky env maps. Bookmarked as the buy-vs-build benchmark for my hand-rolled lab sky.

Xenova Transformers.js runs ML models (embeddings) client-side in spacetable and todoist-adhd-coach.

Mesh-processing/bounding-box library in the eBay scanner's measurement pipeline.

High-throughput LLM inference/serving engine present in my archived SkyPilot LLM examples (llama chatbots / vLLM serving).

C++ Whisper with CUDA, run as an HTTP server (ggml-small.en) on the Jetson; voice-operator's primary STT path (30s of phone audio in ~1.7s) with a circuit-breaker fallback to faster-whisper.

The ws library runs the authoritative multiplayer game server (marine-arena/blipwars) and the editor-mcp-server transport.

Direct-to-fan music sales/streaming platform listed in faible's distribution and label strategy.

DNS, CDN and domain management referenced in my domain-status and infrastructure tracking for his various .com properties.

Convex reactive backend (convex.json, VITE_CONVEX_URL) used in the pushmatrix-archive/clawhub project.

Music distributor for getting faible/FableGroove releases onto streaming platforms; faible has a dedicated DistroKid upload guide.

AI video-generation toolset (Soul characters, Elements references) paired with Higgsfield in the gen1515 trailer production plan.

Code hosting and GitHub Actions CI — editor-mcp-server and quillman ship workflow YAMLs; my repos (voice-operator, etc.) live on GitHub (default private).

Domain registrar and DNS host; faible has a full GoDaddy DNS setup + troubleshooting guide for pointing a custom domain at Vercel.

Display-ad monetization wired into Sci-non-fi via a public AdSense client ID.

Web hosting / registrar option noted in my domain-status tracking.

Hugging Face ([redacted]) gates pyannote model access and is my general model/dataset source, also surfaced via the HF MCP and CLI skills.

Local kiwix-serve hosts an offline Wikipedia for the Professor persona's Wikipedia-RAG answers.

Kroger (King Soopers banner) product + cart API powering the voice-operator chef's grocery scanner lookups and cart-add (productId = UPC-A minus check digit, padded to 13).

Teams integration (via Azure AD app registration) scaffolded as an optional connector for spacetable.

A browser-based 3D creation platform billed as the easiest way to make anything in 3D, with a collaborative editor, asset management, and chat. Bookmarked as a low-friction way to build 3D scenes without a heavy desktop tool.

Adobe's free auto-rigger and animation library. Every animated character in the lab goes through it: rig the Meshy mesh, grab clips (breakdance, floating, sitting-talking), export FBX.

Modal (free tier) is my portable GPU dev platform — hosts the MisoTTS 8B endpoint and the MusicGen LoRA training/inference pipeline for dnb-brain.

Domain registrar used for some of my domains, called out in domain-status and squishdrop vendor/domain shortlists.

Workspace used as the shared project hub (voice-operator status page, 133-row wedding guest DB) that I and Codex both read/write, accessed via the Notion MCP.

A trust-and-verification layer and registry for Claude Code skills: browse and vet specialized skills across domains (email, web scraping, design, spreadsheets). Reference for finding and trusting agent skills.

Pinecone vector database — I provisioned a 'kennybot-memory' serverless index (llama-text-embed-v2, cosine) on AWS us-east-1 to store call/guest embeddings for the wedding bot.

Browser 3D/WebGL engine; I run an MCP server (editor-mcp-server) to drive the PlayCanvas Editor with an LLM, plus a playcanvas-gallery.

Free CC0 textures, HDRIs, and 3D models, no attribution owed. The /lab blocks are surfaced with its woods (oak veneer, black painted planks, dark wood, rough), alternated per box.

Print-on-demand merch fulfillment considered for faible label merchandise and wedding-site product ideas.

Crowdfunded vinyl press-on-demand service evaluated for faible's physical-record / label strategy.

Music streaming/sharing platform included in faible's distribution channel mix.

Artist-facing Spotify platform targeted by faible's release/automation plan for distribution and playlist pitching.

Supabase (Postgres + auth + edge functions, @supabase/supabase-js) is the backend for kenmacy.com, the wedding site, Sci-non-fi and ai-assembly-line.

Alternative music distributor compared against DistroKid in faible's label/distribution planning.

No-code site builder; the karoshi.co site was built in Webflow and later exported to static HTML for Vercel hosting.

Zep long-term memory service ([redacted]) wired into the Mirofish Exploration backend for conversational memory.

Hosted temporal knowledge-graph memory service (zep-cloud SDK) giving the Mirofish agents persistent memory across simulation rounds.

NVIDIA's GPU compute platform that whisper.cpp, Piper, and onnxruntime-gpu run on across the Jetson and Modal GPUs.

Container runtime — the Kroger grocery scanner runs as a Docker container on the Jetson, and Mirofish ships a docker-image build workflow.

CUDA underpins all GPU work on the Jetson Orin Nano (whisper.cpp, TTS, Ollama) and is the most-referenced stack keyword across the repo.

Ollama serves the local Qwen models on the Jetson (keep_alive=-1, 100% GPU); OLLAMA_HOST is also configured in spacetable.

Local model server on the Jetson; voice-operator's independent (non-cloud) brain talks to Ollama /api/chat running qwen2.5:7b (and qwen2.5:3b for story personas) with keep_alive=-1.

The underlying SQL database behind Supabase and faible's Drizzle-backed app, accessed via postgresql:// connection strings.

File-based embedded database used all over my local tools — voice-operator databanks/tasks, familycouncil, spacetable (better-sqlite3), and the dnb-brain corpus harvester.

Linux init/service manager running the three voice-operator units (sip/web/firewall) on the Jetson with Restart=always.

WireGuard-based mesh VPN that puts all my machines on one private network - the zero-config way to reach the Jetson, the Mac, and home boxes from anywhere without opening ports.

Common Extensibility Platform — SuperMoBrainz is a CEP panel using CSInterface.js to bridge a JS UI to After Effects.

DTMF tones (RFC 2833 events) drive the extension dial-map and the private [redacted]/[redacted] backend-override codes in voice-operator.

Adobe's host-scripting language (host.jsx) and ScriptUI toolkit that SuperMoBrainz uses to automate After Effects from the CEP panel.

Autodesk's exchange format; what Mixamo speaks. Everything animated arrives as FBX and leaves as GLB via assimp or Blender.

8kHz narrowband telephony codec — voice-operator passes µ-law audio straight through between the HT801 and the OpenAI Realtime brain.

The quantized model-weight container format used by llama.cpp; my Jetson port plan stores Qwen2.5 as GGUF to swap the runtime without changing weights.

ggerganov's data-over-sound library (WASM/JS, Reed-Solomon ECC) that faible plans to use to bake a secret URL/slug into an audio 'screech'.

The JPEG of 3D and the delivery format for everything in the lab: baked-vertex-color characters, animation-only clips, props. GLB binaries are whitelisted per-file into git so Vercel can build without Blender.

Low-Rank Adaptation fine-tuning I use on Modal — a Sagan-voice LoRA on qwen2.5:7b (plugs into Ollama) and a MusicGen-melody jungle/DnB LoRA via musicgen-dreamboothing.

The open protocol/SDK (@modelcontextprotocol/sdk) underpinning every connector above; my custom Stitch proxy and the PlayCanvas server are both built on it.

Lightweight pub/sub messaging protocol referenced for IoT/device control in my voice-operator and weather device planning.

Low-latency audio codec compressing the QuiLLMan/Moshi voice stream across the websocket for near-real-time response.

RAG pattern used for the Professor's Wikipedia answers (kiwix) and the planned Pinecone-backed kennybot memory.

The 433.92MHz ISM band with ASK/OOK modulation carrying the Patpet collar's remote commands, replayed via CC1101.

Raw-Python SIP/RTP server (phone_operator.py) is the heart of voice-operator, carrying the analog phone audio to/from the AI.

Session Initiation + Real-time Transport protocols implemented from scratch in voice-operator's phone_operator.py to run the touchtone phone call as raw VoIP.

Session Initiation Protocol + Real-time Transport Protocol; voice-operator implements a raw-Python SIP/RTP server (phone_operator.py) bridging an analog phone via a Grandstream HT801 ATA to the AI brains, including RFC 2833 DTMF for the [redacted]/[redacted] backend codes.

The automatic-speech-recognition layer of the phone operator, served by whisper.cpp+CUDA with faster-whisper as fallback.

The speech-synthesis layer of the phone operator, implemented through Kokoro/Piper/Chatterbox with a markdown-stripping sanitizer.

The turn-taking endpointing concept (TURN_SILENCE_MS tuned to 2.5s for ADHD pauses) realized via silero-vad in the phone operator.

Real-time browser media stack referenced for the voice-operator's future bidirectional echo-cancellation (AEC) path.

Bidirectional realtime transport — the multiplayer authoritative servers (marine-arena, emojimo) and Moshi voice streaming all run over ws.

WS2812 addressable-LED library/strip the ESP32 weather-for-mimi device uses for ambient weather color output.

Sony analog broadcast video format — my dad's 1980s Betacam home-movie tapes are the source footage for the betacam-splat Gaussian-splatting project.

Sub-GHz (433.92MHz) ASK/OOK transceiver chip at the heart of the shock-collar/curse-collar trigger — TX/RX to replay the collar's command frames.

USB-serial 7.9" programmable turntable (CT commands via pyserial) that rotates objects for the eBay scanner's multi-view 3D capture.

WiFi/BLE MCU (esp32dev via PlatformIO) running the weather-for-mimi device with OTA updates, NeoPixel and OLED output.

51x23x8mm WiFi MCU that drives the voice-operator's 1602 LCD caption display over USB serial and is slated for the 3D-printed phone case.

Analog Telephone Adapter ([redacted]) that bridges my physical touchtone phone to the Jetson's raw SIP/RTP server.

Grandstream HT801 analog telephone adapter bridges the touchtone phone to the Jetson's SIP server over a direct link.

16x2 character LCD (HD44780) the Feather Huzzah cycles caption pages on, showing live voice-operator call captions.

Likely-Heltec-V3 Meshtastic LoRa node planned into the voice-operator 3D-printed case for off-grid mesh messaging.

1080p USB webcam on the Jetson capturing turntable frames for the eBay scanner reconstruction.

Meshtastic (likely a Heltec V3) is destined for the 3D-printed phone case as a LoRa mesh radio alongside the OLED and Feather Huzzah.

8GB ARM64 edge-AI board (knebot @ [redacted]) that hosts the entire voice-operator stack, eBay scanner, and curse-collar — the recurring deploy target.

The actual 433.92MHz dog-collar (FCC ID 2AHHYP-COLLAR650) I i reverse-engineering to trigger from the CC1101/voice-operator.

USB SDR dongle used (with rtl_433) to capture and decode the 433.92MHz collar transmissions during reverse-engineering.

I2C OLED display module driven by the ESP32 in weather-for-mimi (and the OLED class used in voice-operator displays).

HID barcode scanner read from /dev/input on the Jetson, feeding UPCs into the Kroger grocery-list pipeline (Code39 letters+digits).

AI app builder that generates, edits, and ships real SwiftUI mobile apps from a plain-language description, no coding required. Filed here as an AI dev tool rather than a web-design reference.

Arduino framework (with over-the-air ArduinoOTA updates and WiFiManager captive-portal config) for the weather-for-mimi firmware.

Open Asset Import Library; my FBX-to-GLB bridge in the character pipeline between Mixamo exports and the web lab.

The 3D workhorse: source of truth for the lab's mountains and props, headless Python bakes (vertex-color restores for Mixamo rigs, animation-only GLB exports), and FBX-to-GLB wrangling.

Cross-platform Gaussian-splatting trainer named alongside gsplat as a candidate training backend for the memoryspace video-to-3D pipeline on Modal.

Structure-from-motion tool for recovering camera poses; used (or skipped, when turntable angles are exact) ahead of Gaussian-splat training in my 3D-reconstruction pipelines.

Node-based generative-image workflow engine; faible ships a core-node ComfyUI 'tape builder' workflow (mask + recolor album-art onto a cassette template), and gen1515's 2026-04 remix fed Midjourney stills through ComfyUI into Seedance.

FFmpeg handles audio/video transcoding across voice-operator (recording, DJ stream) and dnb-brain (corpus harvesting).

Radiance-field 3D technique at the core of betacam-splat (turning 1980s Betacam home movies into walkable scenes) and the memoryspace project.

Modern DOOM source port running McDOOM, my McDonald's-flavored mod: DECORATE actors replacing the stock monsters, Freedoom as the base WAD, PixelLab sprites.

ylacombe's LoRA DreamBooth wrapper for Meta MusicGen that I run on Modal to fine-tune musicgen-melody on his jungle/DnB seed tracks in dnb-brain.

Local Markdown note vault that emojimo can read/append to via an /api/obsidian endpoint pointed at OBSIDIAN_VAULT_PATH.

Embedded build/dependency manager (esp32dev + esp32dev_ota envs) compiling and flashing the weather-for-mimi ESP32 firmware.

Headless Chromium harness behind every 'verified' claim on this site: SwiftShader WebGL screenshots, scene probes, and adversarial checks of the 3D lab before anything deploys.

ISM-band decoder used with the RTL-SDR to sniff/identify the 433MHz dog-collar protocol frames.

yt-dlp pulls audio for the DJ persona's 'Spotify-style' search and for harvesting the dnb-brain training corpus.

Portfolio of Ali Imam, an India-based design engineer who builds products, sites, and brands. Kept as an inspiration reference for design-plus-code craft and interaction detail.

Browser tool of ready-made motion templates for showing off designs: drop in an image or video, tweak, and export MP4/WebM, no motion-design experience needed. Handy for product shots and portfolio reels.

A faster, GDPR-friendly drop-in alternative to Google Fonts (privacy-first web font CDN). Reference for web typography without the tracking.

React copy-in animated components installed via the shadcn CLI (magnetic docks, gradient effects, particle typography). Reference for ready-made interactions. (Moved from componentry.fun.)

shadcn/ui component library built on React, Tailwind, and Framer Motion: 78+ animated components and effects plus marketing blocks. A go-to reference for polished, motion-forward web UI.

Emil Kowalski, design engineer at Linear and creator of Sonner (toast) and Vaul (drawer). His site and blog on animation, motion, and UI craft are a gold-standard reference for design skill and taste.

A collection of free SVG generators and color tools: gradients, patterns, textures, blobs, noise, and backgrounds. A go-to toolkit for vector web-design assets.

Web app that generates unique SVG design assets, backgrounds, blobs, waves, stacked waves, and gradients, exported as SVG or PNG for your design tools. A staple for quick section backgrounds.

The animated companion to shadcn/ui: 150+ free, open-source animated React components and effects (Tailwind + Motion) for building polished landing pages fast. A staple web-design reference.

ibelick's open-source UI kit of animated React components (Framer Motion + Tailwind, shadcn copy-paste style): text effects, magnetic elements, spotlights, tilt, morphing text. Reference for polished motion primitives.

Animated React UI library, backgrounds, text effects, and interactive bits you copy in. Reference for eye-catching React micro-interactions and hero moments.

Neobrutalist React component library (Radix/Base UI + Tailwind v4, shadcn): thick borders, hard shadows, loud color, 50+ components and 158+ blocks. Reference for bold, non-generic UI.

Tool for beautiful moving gradients that exports to Framer, Figma, and React. Reference for lush animated gradient backgrounds.

Generates SVG section dividers (waves, curves, slants) you tune and drop between page sections. Handy layout tool for shaping the seams of a web page.

shadcn/ui + Next.js + Tailwind + Motion.dev component library, 105+ ready-to-use animated components you pull in via CLI. Reference for slick landing-page sections and interactions.

React + Tailwind v4 + Motion/GSAP component library, shadcn-compatible, with polished animated blocks (hero, pricing, FAQ) and an AI-agent-friendly install API. Reference for motion-forward UI.

A gallery of handmade, ready-to-customize website templates (Notio, Axis, and more) meant to be tweaked and shipped fast. Reference for clean modern layout structure.

Tailwind CSS color-scale generator: build a palette and preview it live across real components. Reference for building custom Tailwind color systems.

A collection of copy-in web UI components and effects, bookmarked as a web-design reference (site blocks automated fetching, so details are eyeballed rather than scraped).