Forkcast Monthly #3: system_prompts_leaks, OpenCut, code-review-graph, build-your-own-x, awesome-llm-apps, DesktopCommanderMCP, editor

Episode Summary#

Episode 3 covers 35 trending open-source projects from July 2026. Highlights include system_prompts_leaks exposing hidden AI instructions with 61K stars, the 80K-star OpenCut video editor rewriting with Rust, code-review-graph building persistent codebase maps for 38-528x token savings, the 533K-star build-your-own-x tutorial collection, RuView turning WiFi routers into presence detectors, and OmniRoute providing a free 290-provider AI gateway.

Repositories Covered#

system_prompts_leaks#

AI companies hide the instructions they give their models behind closed doors. You have no idea what rules govern the AI you use daily. 689 commits, 10.1K forks. Covers Anthropic Claude Fable 5, Opus 5, Claude Design, Claude Code. OpenAI GPT-5.6-Sol, Codex. Google Gemini 3.5 Flash, 3.1 Pro. xAI Grok, Cursor, Copilot, VS Code, Perplexity. Featured in Washington Post interactive story. CC0-1.0 license.

OpenCut#

CapCut is proprietary, pulls your projects into a walled garden, and can change its terms at any time. Creators need a free, open alternative. 80.1K stars, 8K forks, 1597 commits. Being rewritten with Rust core for desktop, mobile, and browser from one codebase. Plugin-first architecture. MCP server for AI agents. Headless mode for automation. MIT license.

code-review-graph#

AI coding tools waste tokens reading your entire codebase when they only need a small subset. On large repos, this burns money and slows reviews to a crawl. 27.8K stars, 2.6K forks, 905 commits. 38x to 528x context reduction across 6 repos. F1 0.71 across 13 commits. Works with Codex, Claude Code, Cursor, Windsurf, Zed, Continue, OpenCode, Gemini CLI. Voyage embedding provider support. Tree-sitter based. MIT license.

build-your-own-x#

Reading about how databases, operating systems, or compilers work is abstract. You learn fastest when you build the thing yourself from zero. 533.2K stars, 50.4K forks, 616 commits. 30+ technology categories: 3D Renderer, AI Model, BitTorrent Client, Blockchain, Database, Docker, Docker, Emulator, Game, Git, Neural Network, OS, Physics Engine, Programming Language, Shell, Web Server. CC0 license.

awesome-llm-apps#

Building LLM apps from scratch means reinventing the wheel for every project. You need working examples and battle-tested patterns to ship fast. 129.2K stars, 19.1K forks. Covers AI agents, local LLM, RAG, curated lists. 100+ ready-to-use examples.

DesktopCommanderMCP#

Claude Code can read your code but can’t control your terminal or edit files with surgical precision. You end up copy-pasting commands like it’s 2015. 9K stars, 1K forks, 552 commits, 56 tags, 37 releases. Terminal control, file system search, diff file editing. Works as Claude Code plugin. v0.2.44 latest. TypeScript/JavaScript. MIT license.

editor#

3D architectural editors are expensive desktop software locked behind licenses. Collaborating on spatial designs in real time requires downloading gigabytes of tools. 20.3K stars, 2.6K forks, 1245 commits. Browser-based 3D editor. Turborepo monorepo, Bun package manager. Next.js. MCP server included. Node-based scene system with Zod schemas. MIT license.

agent-skills#

AI coding agents generate code without engineering discipline. They skip tests, ignore type safety, and ship without review. You need guardrails that make agents professional. 81.1K stars, 8.7K forks, 383 commits. Works with Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Antigravity. 4 persona agents. Commands system. MIT license.

codex-plugin-cc#

You use Claude Code for your workflow but Codex has strengths in code review and background tasks. Switching tools kills your flow state. 30.6K stars, 2K forks, 29 commits, 7 tags. /codex:review for read-only review, /codex:adversarial-review for challenge review, /codex:rescue and /codex:transfer for delegation. Requires ChatGPT subscription or OpenAI API key. Apache-2.0 license.

strix#

Security audits are expensive, slow, and require specialized pentesters. Small teams can’t afford to hire security consultants for every release. 46K stars, 4.8K forks, 647 commits, 23 tags. Autonomous AI hackers. SARIF output. Integrates with CI/CD. Built on LiteLLM, Caido, Nuclei, Playwright. Linux ARM64 standalone. Apache-2.0 license.

speech-to-speech#

Voice AI is locked behind cloud APIs with latency, privacy concerns, and per-minute billing. Running voice agents locally requires stitching together multiple tools. 9.3K stars, 1.3K forks. From Hugging Face. Local voice agents using open-source speech models. End-to-end pipeline. Audio processing category.

worldmonitor#

Staying on top of global events means jumping between dozens of news sites, social media, and monitoring tools. There’s no unified situational awareness interface. 77.1K stars, 11.8K forks. New 2026. AI-powered news aggregation, geopolitical monitoring, infrastructure tracking. Unified situational awareness interface. Local LLM support. Monitoring category.

meetily#

Meeting transcription tools like Otter.ai send your confidential conversations to the cloud. You need meeting notes without giving away your trade secrets. 27.6K stars, 3K forks. 4x faster Parakeet/Whisper live transcription. Speaker diarization. Ollama summarization. 100% local processing. Rust core. macOS and Windows. Self-hosted. #1 open-source meeting note taker.

buzz#

Team chat tools treat AI agents as bolt-on bots. They can’t participate in conversations naturally, share context, or coordinate like human team members. 18.8K stars, 1.9K forks. New 2026. By Block (Square). Rust-based. Apache 2.0. Agent skills built-in. Modern real-time messaging with AI-native design. Self-hosted category.

bitchat#

When the internet goes down in protests, disasters, or censorship regimes, every messaging app becomes useless. You’re completely cut off. 33.6K stars, 5.4K forks. Bluetooth mesh for offline, Nostr for internet. 30 locales. Noise Protocol encryption. 7-hop multi-relay. IRC-style commands. Emergency wipe triple-tap. Apple Silicon and Intel support. Public domain license.

destructive_command_guard#

AI coding agents can accidentally run rm -rf, force-push to main, or delete your production database. One wrong command and your entire repo is gone. 5.5K stars, 223 forks. New 2026. Blocks dangerous git and shell commands. Safety layer for AI agents. Prevents data loss from autonomous command execution.

GeoLibre#

GIS software like ArcGIS is expensive and heavy. Developers and researchers need a lightweight, open platform for geospatial visualization that works on desktop and web. 4.6K stars, 478 forks. New 2026. Cloud-native GIS. Desktop and web environments. Responsive layout for mobile. Data visualization. Lightweight architecture.

Catch2#

C++ testing frameworks are either clunky C-style or require heavy build system integration. Writing tests should feel natural, not like fighting the tool. 21.4K stars, 3.5K forks. C++14, C++17 and later. Unit-tests, TDD and BDD. Industry standard for C++ testing. Static analysis category.

lingbot-map#

3D scene reconstruction from video requires expensive hardware and hours of processing. Real-time streaming reconstruction was impossible for most teams. 16K stars, 1.8K forks. New 2026. Feed-forward 3D foundation model. Streaming data input. Computer vision and 3D generation. Data streaming category.

ai-job-search#

Job hunting means writing dozens of tailored CVs and cover letters by hand. Each application takes an hour of customization that burns you out before you even get an interview. 28.8K stars, 10.2K forks. New 2026. Built on Claude Code. Evaluates jobs, tailors CVs, writes cover letters, prepares you for interviews. Fork-and-fill workflow. Resume building and workflow automation.

ossie#

Building immersive AR/VR applications requires proprietary SDKs and platform-specific toolchains. Open standards for immersive computing are fragmented. 1.7K stars, 207 forks. Apache Foundation project. Open source immersive environment tooling.

ai-agent-book#

AI agent architecture knowledge is scattered across blog posts, papers, and Twitter threads. There’s no single, structured resource covering design principles with working code. 28K stars, 2.9K forks. Book by Li Bojie. Covers AI agent design principles and engineering practice. Full text with compiled PDF. Per-chapter companion code. AI agent and programming examples.

RuView#

Smart home presence detection requires cameras that invade privacy or expensive radar sensors. WiFi signals already fill every room, but nobody uses them for sensing. 87.8K stars, 12K forks. Featured on Trending 8 times in 30 days. WiFi-based spatial intelligence. Vital sign monitoring. Presence detection. No video required. Rust implementation.

skills#

AI coding agents lack engineering rigor. They skip error handling, ignore edge cases, and ship code that works in demos but breaks in production. 196.8K stars, 17.4K forks. New 2026. Featured 7 times on Trending in 30 days. Production-grade engineering skills. TypeScript, testing, and software engineering. AI skills category.

hallmark#

AI-generated UIs look generic and soulless. Every button is the same blue, every card has the same shadow, every page follows the same boring template. 20.1K stars, 1K forks. New 2026. Featured 7 times on Trending in 30 days. Anti-AI-slop design skill. Works with Claude Code, Cursor, and Codex. AI skills category.

jenkins#

CI/CD pipelines need a battle-tested, extensible automation server that integrates with everything. Most alternatives either lack plugin depth or enterprise maturity. 26.3K stars, 9.7K forks. Featured 6 times on Trending in 30 days. The industry-standard automation server. Workflow automation category.

airi#

AI companions are cloud-locked and personality-less. You want a self-hosted virtual companion you actually own, with real-time voice and game-playing abilities. 46K stars, 4.7K forks. Self-hosted. Real-time voice chat. Plays Minecraft and Factorio. Web, macOS, Windows. Digital human and chatbot category. 14636 repo ID.

Vibe-Trading#

Quantitative trading requires deep financial knowledge, data engineering, and strategy backtesting. Individual traders can’t compete with institutional quant desks. 28.9K stars, 4.9K forks. New 2026. Featured 5 times on Trending in 30 days. Multi-agent LLM trading. MCP protocol integration. Fintech and workflow automation.

open-code-review#

Code review tools either catch obvious style issues or use LLMs that hallucinate. You need both deterministic bug detection and AI-powered analysis working together. 16.8K stars, 1.2K forks. New 2026. Battle-tested at Alibaba. Hybrid: deterministic pipelines + LLM Agent. Precise line-level comments. Built-in ruleset for NPE, thread-safety, XSS, SQL injection. OpenAI and Anthropic compatible. Apache 2.0.

posthog#

Product analytics tools are fragmented. You need one tool for analytics, session replay, feature flags, experiments, and error tracking, not five separate SaaS subscriptions. 37.4K stars, 3.2K forks. Featured 4 times on Trending in 30 days. AI observability, analytics, session replay, flags, experiments, error tracking, logs. MCP integration. Self-hosted.

jcode#

AI coding agent harnesses consume hundreds of megabytes of RAM just to run. On constrained machines or when running multiple agents, memory becomes the bottleneck. 14.3K stars, 1.6K forks. New 2026. Featured 4 times on Trending in 30 days. Most RAM efficient harness. AI agent and AI coding assistant categories.

caveman#

AI coding agents burn through your API budget with verbose responses. Every token costs money, and most of what the agent says is filler you don’t need. 94.7K stars, 5.7K forks. New 2026. Featured 4 times on Trending in 30 days. Cuts 65% of tokens. Caveman mode. AI skills category.

ego-lite#

AI agents need browser automation but existing tools require separate browsers. Your logins never transfer, and agents fight you for the same tabs. 6.7K stars, 330 forks. New 2026. Featured 4 times on Trending in 30 days. Browser built for sharing with AI agents. Dedicated Spaces per agent. Parallel workspaces. Supports Codex and Claude Code. AI agent category.

amnezia-client#

VPNs get detected and blocked by deep packet inspection in restrictive countries. Standard protocols are trivially fingerprinted, leaving users without internet access. 14.2K stars, 1.1K forks. Anti-censorship focused. C++ based. GPL-3.0. Cross-platform desktop and mobile. Marketplace for protocol configs.

bun#

Node.js is mature but slow to start. Webpack configs are nightmares. npm is sluggish. You need separate tools for runtime, bundling, testing, and package management. 95.1K stars, 5.2K forks. All-in-one: runtime, bundler, test runner, package manager. Zig core for speed. Drop-in Node.js replacement. Bundler category.

Watch#

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

Notes#

Transcript and notes will be added from Forkcast output artifacts.