Forkcast Hacker News #10: HNewhere, palmier-pro, CheapSecurity, xy, verified-3d-mesh-intersection

Episode Summary#

Episode 10 covers 30 open-source projects from Hacker News, led by palmier-pro (12.6K stars) for native macOS AI video editing and skills (4.9K stars) for browser automation. Also featured are xy for GPU-accelerated plotting, ctrlb-decompose for log noise reduction, emem for multi-agent memory with citations, and 25 more projects spanning developer tools, security, and coding agent infrastructure.

Repositories Covered#

HNewhere#

Every time you click an HN link, it opens a new tab on the external site. You have to manually switch back to HN to read the discussion. This wastes time and breaks your reading flow. Lightweight userscript. Works with Tampermonkey/Greasemonkey. Zero configuration needed. 95 stars. Simple JavaScript injection.

palmier-pro#

Video editing tools lack native AI collaboration. You have to export, send to a separate AI tool, then re-import. The workflow is fragmented and slow. Native Swift macOS app. 12.6K stars. Open-source CapCut alternative. Built specifically for AI-assisted video editing workflows. Real-time preview.

CheapSecurity#

Commercial CCTV systems are expensive, cloud-dependent, and send your footage to third-party servers. Raspberry Pi owners want local security monitoring without subscriptions. Runs on Raspberry Pi and other SBCs. Self-hosted, no cloud required. Python-based. 173 stars. Motion detection and recording. Low resource usage.

xy#

Existing Python plotting libraries are either slow for large datasets, lack GPU acceleration, or don’t compose well for interactive dashboards. You need sub-second rendering for millions of data points. GPU-accelerated via WebGPU. 389 stars. Composable API. Interactive panning and zooming. Built by the Reflex team. Python-native with TypeScript rendering backend.

verified-3d-mesh-intersection#

AI-generated code for 3D mesh operations is unverified and unreliable. Bugs in CSG boolean operations cause geometry errors in CAD, games, and 3D printing that are nearly impossible to debug. 93 lines of formal Lean4 specification. Mathematically proven correct. 74 stars. CSG boolean operations on 3D meshes. Trust the spec, not the implementation.

yap#

macOS dictation is cloud-dependent, privacy-invasive, and requires internet. Existing local alternatives need large model downloads and complex setup. 256 stars. On-device voice dictation for macOS. No model download required. Uses Apple’s built-in speech frameworks. Works offline. Swift native. Real-time transcription.

letsseal#

Document signing certificates are expensive, vendor-locked, and require trust in centralized CAs. Small businesses and open-source projects can’t afford $200+ per year for code signing. 227 stars. Free, self-hosted document signing. Based on Let’s Encrypt model. Automated certificate issuance. JavaScript/Node.js. Zero-cost code signing.

ctrlb-decompose#

LLMs burn tokens processing verbose logs full of timestamps, request IDs, and stack traces. You pay per token and context windows fill with irrelevant noise before the actual error. 219 stars. Rust-based log decomposer. Strips timestamps, UUIDs, and boilerplate. Reduces token count by 60-80%. CLI and library modes. Works with any log format.

cynative#

Understanding live infrastructure requires logging into multiple dashboards, running ad-hoc commands, and piecing together status from disparate sources. There’s no single CLI that explains what’s actually running. 158 stars. Go CLI for infrastructure observability. Read-only, safe for production. Explains running services, ports, and dependencies. Single binary deployment.

skills#

AI agents need browser automation but existing tools require complex setup, API keys, and cloud services. You want agents that can browse, click, and interact with web pages locally. 4.9K stars. Browser automation skill library for AI agents. Python-based. Local-first. No API keys needed. Works with Claude Code and other agents. Click, scroll, type, extract.

zil-lean#

Authorization logic in AI systems is ad-hoc and error-prone. Google Zanzibar proved scalable authorization, but implementing it requires complex infrastructure. There’s no lightweight DSL for it. 15 stars. Lean4 formalization of Zanzibar authorization model. Datalog DSL. Mathematical proofs of authorization correctness. Compiles to efficient query plans.

minute#

Meeting transcription tools send audio to the cloud. Privacy-sensitive teams need local transcription that works offline without sending conversations to third-party servers. 7 stars. Offline meeting transcription for macOS. Uses Whisper for speech-to-text. llama.cpp for summarization. No cloud dependency. TypeScript/Electron.

base-gpui#

Base UI headless components are web-only. Zed’s GPUI framework needs native headless UI primitives for building desktop applications without Electron overhead. 41 stars. Rust port of Base UI to GPUI. Headless components for native desktop apps. GPU-accelerated rendering. No web technologies needed.

magpie#

AI agents handle code and tasks but can’t manage finances. Freelancers and small teams need automated bookkeeping that integrates with their existing AI workflows. 4 stars. Go CLI for AI-powered bookkeeping. Parses receipts, categorizes expenses, generates reports. Integrates with coding agents via MCP.

cloudrift#

AWS bills grow silently with idle resources, oversized instances, and forgotten services. Teams waste thousands per month on infrastructure they don’t know they’re running. 9 stars. TypeScript CLI. Analyzes AWS CloudWatch metrics to find idle resources. Recommends right-sizing. Reports potential savings. Open-source alternative to commercial tools.

cloudflarebase#

Firebase is proprietary, expensive at scale, and locks you into Google’s ecosystem. Developers want a self-hosted alternative with the same developer experience on Cloudflare’s edge network. 23 stars. Firebase alternative built on Cloudflare Workers. Real-time database, auth, storage, and functions. Edge-first architecture. Svelte frontend.

writekin#

Generic AI writing assistants produce generic output. You want an AI that writes in YOUR voice, but fine-tuning requires expensive GPU servers and ML expertise. 12 stars. macOS app. Fine-tunes small LLMs on your writing samples. Runs locally on Apple Silicon. No cloud, no GPU server needed. Swift native.

colmena#

Reproducing firmware CVEs is tedious and non-repeatable. Security researchers waste hours manually setting up environments that can’t be shared or verified by others. 3 stars. Python tool for reproducible firmware CVE environments. Generates verifiable receipts. Enables sharing and peer review of security findings.

minions-army-harness#

Claude Code works well for individual tasks but lacks structured multi-agent orchestration. You need a lightweight harness that coordinates multiple Claude sessions. 4 stars. Python agent harness. Coordinates multiple Claude Code sessions. Task decomposition and result aggregation. Lightweight alternative to complex frameworks.

DocSlicer#

Document parsers are either slow, lose structure, or require GPU. RAG pipelines need fast, accurate document parsing that preserves headings, tables, and relationships. 8 stars. 31 pages per second on CPU. Structure-aware parsing for PDF, DOCX, and PPTX. Preserves headings, tables, and layout. Python. No GPU required.

VideoTomb#

YouTube videos are publicly accessible once uploaded. Creators who want to share private content with specific audiences have no built-in encryption option. 4 stars. Client-side video encryption. Works with YouTube and other platforms. JavaScript. End-to-end encrypted video sharing.

faster-enhancer.c#

Real-time audio/video denoising on embedded devices is either too slow, too power-hungry, or produces artifacts. Edge devices need stable, low-latency noise reduction. 7 stars. Pure C library. Real-time denoising for embedded devices. Stable output quality. Low CPU usage. Suitable for IoT and edge computing.

emem#

Multi-agent systems forget context between sessions and can’t cite their sources. You need persistent memory that tracks where information came from and remains queryable over time. 52 stars. Rust-based. Long-horizon memory with citation tracking. Multi-agent compatible. MCP server. Local-first. Persistent across sessions.

all-ears#

Meeting notes tools are cloud-dependent and combine all sources into one messy document. You want local, composable meeting notes split by source with proper attribution. 6 stars. Swift daemon for local meeting transcription. Browser extension routes meeting audio. Split by source. GPL-3.0. 172 commits.

cognikernel#

AI coding agents forget project context between sessions. You re-explain architecture decisions, coding conventions, and design patterns every time you start a new session. 7 stars. Event-sourced, typed, fail-open memory. Python package. Works with Claude Code and Codex. 157 commits. Apache-2.0.

learning-mode#

AI coding agents write code for you, but you never actually learn anything. If you could paste what Claude wrote and have the task done, it wrote too much. 5 stars. Claude Code skill. Guides learning instead of implementing. Forces the agent to teach rather than do. MIT license. 19 commits.

Orb-Backend#

AI assistants wait for you to ask questions. You want proactive AI that monitors your services, checks on your projects, and reaches out when something needs attention. 7 stars. Python backend. Proactive AI assistant. MCP server with 74 tools. Messages you first. Self-hosted. MIT license.

fedterm#

Claude Code terminals require switching contexts and managing separate windows. You want instant access to Claude from anywhere on macOS without leaving your current workflow. 4 stars. macOS native terminal. Spotlight-style popup. Claude-native integration. Swift/SwiftUI. 8 color schemes. MIT license.

sessiongrep#

AI coding agents forget everything between sessions. You want your next session to recall the last one, but existing memory tools are cloud-dependent or require complex setup. 24 stars. Rust-based. SQLite + FTS5 for fast search. Works with Claude Code, Codex, and Cursor. CLI, TUI, and MCP server interfaces. Apache-2.0.

injex#

Python dependency injection libraries are either heavyweight, untyped, or don’t validate the dependency graph at startup. You catch wiring errors at runtime instead of import time. 19 stars. Zero dependencies. Typed. Validates dependency graph before startup. Fast resolves. Works with FastAPI, Sanic, Quart, FastStream. 134 commits. MIT license.

Watch#

• Video: https://www.youtube.com/watch?v=sRplXBm3LLM
• Cover: https://i.ytimg.com/vi/sRplXBm3LLM/hqdefault.jpg

Notes#

Transcript and notes will be added from Forkcast output artifacts.