Notes

@ctatedev introduces , a built-in terminal component for Native SDK with a real PTY, libghostty-vt rendering, truecolor, selection, scrollback, and session recording. Useful infrastructure for agent orchestrators and software-factory tooling that need an embedded shell instead of bolting one on later.

@dr_cintas points to claude-video, a plugin that lets Claude or other coding agents inspect videos frame by frame using yt-dlp, scene-aware ffmpeg extraction, and fallback transcription when captions are missing. It turns screen recordings, demos, tutorials, and launch videos into something agents can actually debug, summarize, and analyze instead of treating them as opaque media.

@localhost_5173 points to Kero, Egoist's native terminal workspace for macOS. It is built with Swift and libghostty, with split panes, Git integration, project grouping, and a file tree, aiming for a more Mac-native alternative to the usual Electron-heavy terminal workspace stack.

@ihteshamali points to BBC Sound Effects, a remarkably deep public archive with more than 30,000 recordings you can search by keyword, category, duration, and geography. It is the kind of resource that saves hours when you need ambient audio, historical textures, or oddly specific real-world sounds without digging through junk libraries.

@haydenbleasel shipped Files SDK 2.2, a unified storage layer for object and blob backends, with first-class NestJS support, a lightweight Cloudflare R2 client, and React Native plus Expo support. Useful if you want one small API across S3-, R2-, and blob-style storage without committing your app to a single provider's client surface.

@marcelroed introduces Gigatoken, a tokenizer built for language-model workloads that reportedly runs ~100-1000x faster than tiktoken and Hugging Face's already-multithreaded Rust implementations. The interesting part is that it can slot into existing Hugging Face and tiktoken flows, or use its own faster API when tokenization throughput becomes a real bottleneck.

Cloudflare's changelog points to Email Sending event subscriptions, which let you subscribe a sending domain to Queues and react programmatically to bounces, complaints, and other outbound mail lifecycle events with full message context. A small but useful building block if you're treating transactional email as part of your application pipeline rather than a black box.

@DhravyaShah introduces Burrow, a full dev machine that runs in one browser tab by porting Bun's transpiler to WebAssembly and layering in git, a just-bash terminal, package management, port forwarding, previewing, and an in-browser coding agent. A good glimpse of how much local-feeling developer tooling can now collapse into a single client-side surface.

@usekernel introduces browser telemetry, a debugging layer for browser agents that records a real-time structured timeline of network responses, console logs, captcha events, and other browser activity. The pitch is better observability with fewer blind tool calls, and they claim it can cut token usage by as much as 67%.

@argofowl built NameThat, a visual dictionary for UI elements meant to close the naming gap between what designers can point at and what coding agents need to hear. You can describe an element in plain language, get the proper term and API symbol back, then jump into cheat sheets like AppKit vs SwiftUI or the three-way translation table for more precise prompts.

@fayazara built PDF Studio, a browser-native PDF tool powered by qpdf compiled to WebAssembly. It handles locking, unlocking, merging, and splitting locally in the browser with nothing uploaded, and it also runs on Cloudflare Workers. A nice example of serious document tooling moving fully to the client and edge.

@WillowVoiceAI released Willow Frontier Mini, a free cloud dictation model for desktop and phone with zero-data-retention. The pitch is simple: faster, more accurate speech-to-text than local dictation or general API defaults, without metered usage. Interesting as a sign that voice tooling is moving from premium assistant feature toward commodity infrastructure.

@lilianweng published Harness Engineering for Self-Improvement, a thoughtful map of how recursive self-improvement may depend less on models rewriting themselves and more on the harnesses around them: eval loops, auto-research scaffolding, and deployment systems that help produce the next better model.

@orhundev points to dnsglobe, a new Rust + Ratatui terminal app for checking DNS propagation across 34 public resolvers at once. Instead of bouncing between browser tools, you can watch records spread worldwide from the command line and compare resolver answers side by side.

Pasha Sviderski highlights Docker Compose init containers: native pre_start hooks that run one-off ephemeral containers before a service starts. That finally gives Compose a first-class way to handle setup jobs like database migrations, secret fetching, or schema prep without awkward wrapper scripts or separate bootstrap services.

Bilgin Ibryam points to pg_durable, Microsoft's new PostgreSQL extension for durable workflows inside the database. Instead of bolting on an external orchestrator, it keeps state, queues, retries, and recovery in Postgres itself, which makes it an interesting fit for teams that already trust Postgres for HA, backups, and operational simplicity.

Viking points to Agentic Inbox, Cloudflare's self-hosted email client with an AI agent running entirely on Workers. It uses Email Routing and Email Sending for transport, keeps each mailbox in its own Durable Object with SQLite, stores attachments in R2, and plugs in Workers AI plus the Agents SDK for inbox search and reply drafting.

@neogoose_btw points to fff, a file search toolkit for humans and AI agents that keeps an in-memory index and is now showing absurdly fast indexed grep numbers: 2 ms queries on Chromium’s 500k-file repo on the same machine where hot-cache ripgrep took about 2 seconds. If the benchmarks hold up in real workflows, this is one of the more interesting code-search tools of the year.

@CFchangelog points to Wrangler auth profiles, which lets you keep separate Cloudflare logins per project and switch accounts automatically by working directory. Small feature, big quality-of-life win if you juggle client orgs, side projects, and production accounts: less wrangler login churn, less account confusion, and a cleaner multi-account workflow.

Older notes