Forkcast Hacker News #4: kage, lathe, paca, helix-db, luz, gentleos32, clawpatrol
34 open-source projects from Hacker News Show HN, organized into five thematic groups: AI Agent Infrastructure, Developer Experience, Systems and Languages,
Watch Episode
Episode Summary#
34 open-source projects from Hacker News Show HN, organized into five thematic groups: AI Agent Infrastructure, Developer Experience, Systems and Languages, Hardware and Embedded, and Tools and Productivity. Features kage (offline website mirroring), lathe (LLM-powered tutorial generation), paca (multi-agent AI runtime), helix-db (distributed SQLite), and many more innovative tools from the HN community.
Repositories Covered#
kage#
Drives headless Chrome, snapshots DOM, strips ALL JavaScript, rewrites asset links to local files. Output formats: folder, ZIM archive, self-contained binary, desktop app. Polite resumable crawl, respects robots.txt. MIT.
lathe#
Generates multi-part technical tutorials on demand via LLM-agnostic skills. Local web UI with table of contents, side-notes, exercises. Full provenance per tutorial. CLI in Go, 7 agent targets. Homebrew install.
paca#
Self-hosted AI-native Scrum platform. AI agents join sprint planning, pick tasks, write BDD specs, contribute to design docs. WASM plugin sandbox. In-app AI chat, activity diff and revert. Apache 2.0.
helix-db#
OLTP graph-vector database in Rust. Query via Rust, TypeScript, Go, Python DSLs as dynamic JSON. Interactive bootstrapper. Managed cloud with ACID transactions. Apache 2.0.
luz#
C++20 path tracer with zero third-party deps. Monte Carlo global illumination, BVH, adaptive sampling, NFOR denoiser, atmospheric scattering. Blender exporter. 583 commits. MIT.
gentleos32#
Hobby OS for vintage 32-bit PCs. Minimum: i386, 4MB RAM, VGA 640x480x16. Monolithic kernel, compile-time config. Runs on real hardware. Docker build support. GPL-2.0.
clawpatrol#
Security firewall between AI agents and production. Gates every action against HCL rules with CEL expressions. Plugin sandbox with Linux namespaces and Landlock. From the Deno team. MIT.
boo#
Terminal multiplexer on libghostty. Persistent sessions, faithful redraws, automation primitives: send, peek, wait, JSON output without a TTY. Written in Zig. MIT.
cutile-rs#
NVIDIA Research: tile-based GPU kernel DSL in safe Rust. 7 TB/s on B200, 2 PFlop/s GEMM. 171 tok/s Qwen3-4B on RTX 5090. CUDA 13.3+. Apache 2.0.
ezra#
Task queue backed by SQLite, speaks Redis RESP3. Single 20MB binary. 15k-30k tasks/sec on SSD, 40k-80k on NVMe. ~20MB RAM baseline. Erlang/OTP.
khadas_yolov8n_multithread#
YOLOv8n UAV detection on RK3588S at 46 FPS camera ceiling. 3-thread NPU pool. ~140MB RAM per 1080p stream. Two cameras. NPU hand-off to Qwen2.5-0.5B for LLM summaries. Runs on $90 boards.
claude-quota#
SwiftBar plugin with real-time Claude Code quota bars. Per-account 5-hour windows with green/orange/red coloring and countdowns. Reads OAuth tokens read-only from macOS Keychain. MIT.
bulk-delete-claude-chat#
Browser console script that deletes every Claude.ai conversation at once. Uses internal API endpoints. Single delete-all.js file. MIT.
agoracosmica#
Living library with 30 historical figures as AI Echoes. 360 stories, 110 councils, bilingual EN/DE. BYOK encryption (AES-256-GCM), no tracking. Self-hosted TTS/STT. Nonprofit. AGPL-3.0.
nucleus#
Container runtime: 12ms cold start vs Docker’s 500ms. PostgreSQL at 107K TPS above bare metal. No daemon. Linux primitives: cgroups v2, namespaces, seccomp, Landlock. TLA+ verified. MIT.
ScreenMind#
AI screen memory with Gemma 4 E2B locally. Smart capture, vision analysis, hybrid search, conversational RAG. Voice memos, meeting transcription. AES encryption, sensitive data redaction. MCP server. MIT.
talos#
API key server, Ory ecosystem. Token derivation generates short-lived JWT and macaroon tokens. Offline verification, sidecar deployment, constant-time comparisons. Apache 2.0, v26.2.0.
mach#
Systems language with no hidden behavior. Self-hosted, no LLVM, no libc. Explicit allocators. Single binary for build, link, test, deps, cross-compile. v1.6.0. 2,668 commits. MIT.
brain-map-skill#
Converts Markdown folders into interactive HTML: Cytoscape force graph, scrubbable growth timeline, search, filter. 992-note prebuilt demo. Works as Claude Code skill. Python stdlib only. MIT.
claude_code_vs#
Bridge bringing Claude Code CLI into VS 2026. Native diff window with Accept/Reject/Reject-with-feedback. C#/C++ diagnostics sharing. Live token usage panel. WebSocket bridge. v1.0.1, MIT.
sharkclean-mcp#
MCP server for SharkClean vacuums, bypassing legacy Ayla cloud. Tools: list robots, start cleaning, clean rooms, dock, locate. PKCE browser auth. Claude Code compatible. MIT.
USTP-Secure#
Reliable unordered transport over UDP with AEAD. ChaCha20-Poly1305, AES-256-GCM. X25519 key exchange, TOFU. Tested at 33 percent loss Brazil-to-Canada. Selective retransmission, adaptive RTO. IETF draft. MIT.
zero_browser#
Browser system with zero HTML/CSS/JS. Native data-driven elements: Markdown, tables, lists, grids. Flutter/Dart, cross-platform 6 targets. v0.0.4, GPL-3.0.
state-harness#
Lyapunov-stability monitor for LLM agents. Detects token spirals, retry storms, policy drift. Classifies failure with diagnostics, zero extra LLM calls. 38.6 percent fewer search nodes on SWE-bench Verified. Rust core plus Python SDK.
azdo#
Bubble Tea TUI for Azure DevOps. PR diff viewer, work items, pipelines with color-coded stages. Metrics dashboard, 8 themes, demo mode. Secure PAT via OS keyring. Go, v0.6.2, MIT.
claireon#
MCP server inside Unreal Editor. Two tools: tool_search and python_execute. Covers Blueprints, animations, Niagara, PCG, Data Tables, MetaSounds, Level Sequences. Per-asset session locking. UE 5.5+, Beta, MIT.
rubyterm#
Terminal emulator entirely in Ruby: pure-Ruby X11 client, TrueType font renderer, damage-driven rendering. Jump-scroll for output floods. Test harness against tmux oracle. 120 commits. MIT.
ttar-tiny-tar-archivist#
Tar archiver in freestanding C, zero deps. 2.4KB binary on arm64. Syscall-only, no libc. Supports arm64, arm-v7, mips, riscv, x86, i386. GPL-3.0.
git-lrc#
AI code review hooked into git commit. 100 plus failure patterns across 10 risk categories. Issue navigator, summary deck, GitHub-style review UI. Pre-commit braking for GenAI. 452 commits, 1.4k stars.
DreamServer#
One-command private AI server. Bundles LLM inference, chat UI, voice, agents, n8n workflows, RAG, ComfyUI image gen. Auto-detects GPU for model selection. 2.1k stars, Apache 2.0.
STEPLA-1#
8-bit Harvard CPU from individual 74-series logic gates in Logisim. Hardwired control unit: every signal is a physical gate. 16 instructions, 4 registers. Bootstrap Control Unit. Target 4MHz on breadboard. 43-page spec PDF. MIT.
ctx#
Recommendation engine over 102,928-node knowledge graph with 2.9M edges. Watches build context, scores skills and agents. Louvain communities, micro-skill gate. Claude Code hooks. 803 commits, MIT.
contextspy#
Context profiler for LLM agents. HTTPS proxy intercepts requests, analyzes 8 token categories. Local SQLite, real-time dashboard. Session tracking, cost estimation. Homebrew plus PyPI. Apache 2.0.
afterburner#
JS/TS to WASM sandboxed runtime in Rust. Memory limits, timeouts, capability-gated I/O. Own package format and registry. Agent integration rewrites node commands through sandbox. MIT.
Watch#
- Video: https://www.youtube.com/watch?v=UYaHQ3sCw9U
- Cover: https://i.ytimg.com/vi/UYaHQ3sCw9U/hqdefault.jpg
Notes#
Transcript and notes will be added from Forkcast output artifacts.