AI Daily Digest · 2026-07-23
🔬 New AI Craft
1. Finn-loop: 3-Skill AI Software Factory for Claude Code
Finn-loop is an open-source AI software factory with 3 specialized agent skills — spec writer, builder, and reviewer — with humans only merging. Unlike a sequential plan-execute loop, it separates writing and reviewing into independent agent roles, mirroring real team collaboration by enforcing code review as a distinct step rather than a post-hoc check.
https://github.com/finna/Finn-loop
2. Agent Execution Partnership (AEP): Authorization Control Plane for AI Agents
AEP is an open-source control plane that gates every AI agent action behind an authorization step before execution. Unlike Agent Plan mode that focuses on task decomposition and sequencing, AEP inserts a security checkpoint before each action, preventing agents from executing unauthorized or dangerous operations.
https://github.com/eli-labz/Agent-Execution-Partnership
3. fastctx: Context-Efficient Repo Tools for AI Agents via MCP
fastctx provides MCP-based tools that help AI agents quickly grab project structure and key files, reducing context waste. Unlike RAG's document semantic search, it scans code repo structure → extracts key info → compresses, replacing full-file reads with targeted context snippets.
https://github.com/yc-duan/fastctx
🛠️ Tools & Tips
1. video-shotcraft: AI Video Skill for Claude Code & Codex
106 shot recipe cards and 161 motion previews turn Claude Code into a cinematic product video generator using Remotion.
https://github.com/Vincentwei1021/video-shotcraft
2. nativ: Native macOS Local AI App
Run MLX models locally on your Mac — chat, serve, monitor, and connect models from one native app, no internet required.
https://github.com/Blaizzy/nativ
3. tokentab: CLI Cost Tracker for AI Coding Agents
Reads Claude Code, Codex, and Gemini CLI session logs and tells you how much each project cost by model and time period.
https://github.com/sequilade/tokentab
4. PilotBrowseMCP: Let AI Agents Control Real Chrome via MCP
A browser runtime that lets AI agents take over your real Chrome browser through MCP — browse, screenshot, and report.
https://github.com/EndymionLee/PilotBrowseMCP
⭐ Open Source Highlights
1. open-kritt — Orchestrate AI Agents to Find Real Vulnerabilities
Orchestrates multiple AI agents to collaboratively find real security vulnerabilities in code, going beyond traditional pattern-based scanning.
https://github.com/Kritt-ai/open-kritt
2. klaatcode — Open-Source AI Coding Terminal Agent
Claims Claude Code-grade accuracy with smart model routing that picks the right AI model for each task to balance quality and cost.
https://github.com/KlaatAI/klaatcode
3. thinking-orbs — Animated Loading Dots for AI & Agent UIs
Dotted thought-orb loading indicators with 6 tuned states, 2 sizes, and auto dark/light mode — designed specifically for AI and agent interfaces.
https://github.com/Jakubantalik/thinking-orbs
4. MiniCPM-Robot — On-Device AI Brain for Robots
OpenBMB's new on-device AI brain, smaller and faster than standard MiniCPM, purpose-built for robot applications.
https://github.com/OpenBMB/MiniCPM-Robot
📰 Industry News
1. Simon Willison publishes a detailed postmortem of OpenAI's cyberattack on Hugging Face, calling it "science fiction that happened" — 174pt on HN
2. AI startup founders jointly urge the Trump administration not to cut off access to Chinese open-weight AI models, arguing it would harm the US AI ecosystem
3. Futurism reports that several AI companies are hiding massive debt through off-balance-sheet financing, reminiscent of WeWork's accounting
4. DARPA and the US Air Force successfully fly an AI-controlled F-16 fighter jet in a real test flight
🚀 Major Releases
(No items this day)
