#cloudflare
17 posts
Prism is a product line at 1.0: AGPL playground and metered control plane, iOS and Android clients, MIT MCP for agents, ~93 models, and honest gaps. Conrad Rockenhaus maps the doors, hosts, and repos.
Postern is at v1.4.1. v1.4.0 bound read and send to a person, not a shared mailbox god-token; v1.4.1 fixes the webmail attachment-send race. Conrad Rockenhaus covers identity-bound registry credentials, MCP multi-person setup, the projection census that keeps IMAP honest, and what still stays estate-wide on purpose.
play.skyphusion.org is a public Prism deploy now: first-party accounts, no Cloudflare Access wall, and mandatory per-user BYOK. Conrad Rockenhaus explains AUTH_MODE=public, fail-closed billing, what self-host still looks like, and how the model catalog grew after the cutover.
Skyphusion Labs published @skyphusion/security-audit: adversarial LLM reviews of GitHub PRs and full repos through Cloudflare Workers AI and AI Gateway. Conrad Rockenhaus covers pr vs repo mode, the public/private data boundary, redact hygiene that must not eat real code, and why a gate must be able to go red.
The Vivijure control panel is no longer Cloudflare-only. Conrad Rockenhaus walks the constellation after the host split: vivijure-cf on Workers, vivijure-local on Docker (home PC or any cloud VM), vivijure-core unifying both, vivijure-mcp for agents, plus Slate, RunPod, own-GPU doors, and the finish engines. Same studio API either way.
Postern tagged v1.0.0 on July 9, 2026: the first production-ready release of the self-hostable Cloudflare mailbox for humans and agents. Notes on what the tag means, the five-week path from send-only cf-email-relay to a full store with webmail and IMAP, the acceptance smoke that gates the release, and the honest cost of making real mail clients happy.
mud-bots is my open source suite of AI players for The Hollow Grid, the Cloudflare Workers MUD from Skyphusion Labs. Open-source models on Workers AI log in like human players, read the structured event channel, and face the game's real moral choices: free the caged or take the loot, defend the refugee or join the strong. Notes from Conrad Rockenhaus on the thesis, the two bots that chose well from opposite directions, and why the bots double as live QA.
A full tour of the Vivijure constellation by Conrad Rockenhaus of Skyphusion Labs: the vivijure Studio control plane, CPU media containers on your own iron, the Slate Discord screenwriter, the vivijure-backend RunPod GPU engine, the vivijure-local-12gb and vivijure-local-16gb consumer GPU doors, cloud i2v modules, and the vivijure-musetalk, vivijure-upscale, and vivijure-audio-upscale finish engines. All AGPL-3.0, all on GitHub, and after a two-week hardening sprint the studio is almost ready for full public release.
Common Thread pairs a twelve-section methodology paper with a Cloudflare Workers reference implementation: archive public social artifacts, extract deterministic behavioral features, and emit calibrated cluster-level attribution claims with full chain of custody. For journalists, OSINT practitioners, and researchers who need reproducible investigations without platform-internal data.
Postern replaces our retired send-only cf-email-relay with a full mailbox: inbound via Email Routing, outbound via Email Sending, D1 and R2 storage, FTS and optional semantic search, webmail and read-only IMAP, a Go SMTP relay for legacy callers, and one structured API that agents and humans both use.
SidVicious_exe is a self-hosted Discord roadie forked from Slate with the film stack removed: Claude via AI Gateway, web search, a Vectorize knowledge base, and image generation across Workers AI and gateway models, with a deliberately irreverent personality.
Slate is the Discord front door to Vivijure: a collaborative screenwriter assistant that keeps a structured storyboard brief in channel, generates portraits and thumbnails, searches the web and a knowledge base, and submits finished bundles to the Vivijure studio API when the crew is ready.
The fourth Vivijure showcase, Vivijure Speaks: a talking character lip-synced to its own dialogue and upscaled, rendered start to finish on a self-hosted GPU with nobody steering it. Two shots, about two and a half seconds, and voiced. The honest part is that it came out silent once, and then a from-scratch re-fire found two more orchestration bugs before any user could. Notes on the three control-plane and backend fixes that gave it a voice.
The first film rendered end to end on Vivijure Studio: NEON HALFLIFE, a silent 1080p ten-shot cyberpunk render on a self-hosted GPU. The point is not the picture; it is that the first unattended run came out clean (zero clips dropped) and then healed its own finish-phase stall across a session restart with nobody watching. Notes on why the studio ships silent by default, and why a system recovering itself is the milestone worth showing.
Introducing The Hollow Grid, a multiplayer MUD that runs entirely on Cloudflare Workers and Durable Objects: one Durable Object holds a whole world, WebSocket hibernation lets it sleep at $0 when empty, and a single alarm drives all of time. It is also a small federation, where separate world deployments share one backend Grid (one faction war, one character that travels between worlds) over a service-binding RPC contract. Notes on the design rules that make it testable, the structured event channel that drives the client, a bot, and the tests alike, the cross-Worker bug that taught us how service bindings really behave, and the human-and-AI collaboration that built it.
A small two-part utility for sending transactional email through Cloudflare Email Sending: a Worker with a service-binding RPC and a token-gated HTTP endpoint, and a tiny Go SMTP relay that bridges the services that can only speak SMTP. Notes on the one-validation-path design, why the SMTP bridge is the genuinely useful part, the container wrinkle that made the relay multi-listen, and wiring Jenkins and Uptime Kuma to send through it.
Architecture and lessons from Prism, a multimodal AI playground as a single Cloudflare Worker: 35 chat models across five providers, voice chat, RAG, projects, Discord ingestion, web search, SSE streaming, Workflows for durable video and music jobs, and a public demo mode where each user brings their own AI Gateway credentials.