AI Open Source Radar — July 29, 2026

Selection rule: repositories rising right now. Recent velocity (daily trending rank) is the primary signal; cumulative stars are only a reference. Large, perennial repos that simply hold a top spot (e.g. obra/superpowers, mattpocock/skills, affaan-m/ECC, andrewyng/aisuite, huggingface/speech-to-speech) and non-AI repos are excluded.

Data-source note: OSSInsight /trending/ai is still client-rendered (empty table, unusable), and a direct fetch of github.com/trending returned a stale mid-July cache. We therefore used the OrangeBot daily snapshot (2026-07-28, refreshed hourly) as the primary velocity signal. Exact “stars today” deltas were unavailable, so each item’s momentum is shown as its 2026-07-28 daily trending rank (cumulative stars noted only where confirmed). All five entries are new to our streak log.

bradautomates/claude-video — daily trending #12 (2026-07-28, new)

github.com/bradautomates/claude-video — a skill/plugin that gives Claude the ability to “watch” any video. With a single /watch command it downloads a video, extracts frames and transcribes it, then hands the whole thing to Claude so a text model can reason about the footage.

How it works: yt-dlp (downloads from YouTube/Instagram/X/Vimeo and any yt-dlp site) + ffmpeg (frame extraction) + Claude’s multimodal Read tool. Transcription prefers free public captions and only needs a Whisper API key (Groq or OpenAI) when captions are missing. First run auto-installs yt-dlp and ffmpeg (brew on macOS; exact commands printed for Linux/Windows), so it is effectively zero-config.
Differentiator / use cases: installs as a Claude Code marketplace plugin and also runs under Codex, Cursor, Copilot and Gemini CLI. Point it at a meeting, tutorial or long YouTube link and summarize or extract from it directly.
Why it is surging: genuine “multimodal video understanding for a text agent” utility, plus a free-captions path that removes the API-cost barrier, drove fast adoption at the early-release (v0.1–v0.2) stage.

Practical angle: turn customer onboarding/training videos into transcripts and summaries from a single link, feeding a knowledge base or FAQ draft.

#VideoUnderstanding #WhisperTranscription #FrameExtraction #ytdlp #MultimodalAgents #ClaudeCodeSkill

virgiliojr94/book-to-skill — daily trending #7 (2026-07-28, new, ~8,600⭐)

github.com/virgiliojr94/book-to-skill — turns a technical book (PDF and more) directly into a ready-to-use agent skill.

What it does: takes any technical book as PDF, EPUB, DOCX, HTML, Markdown, plain text, RTF or MOBI/AZW (via Calibre) and produces a Claude Code skill that structures the frameworks, mental models, principles, techniques and anti-patterns inside.
How it works: after install, typing /your-book-slug has the agent read the relevant chapter and answer from the actual text. Chapter files are loaded on demand, so they do not consume the skill context budget until that topic is accessed — meaning even a large book carries a low token overhead. It follows the standard SKILL.md format, so GitHub Copilot CLI, Amp and Claude Code all read it identically.
Differentiator / use cases: not “book summarization” but applying an author’s frameworks while you work. Beyond books it converts structured prose like architecture decision records (ADRs) and runbooks into a reusable knowledge base, answering from source without hallucination.
Why it is surging: in the growing Claude Code skills ecosystem, the ability to easily turn a personal library or internal docs into skills clicked quickly with developers.

Practical angle: convert your team’s internal standards and playbooks into a skill so agents always answer against the latest rules.

#DocumentToSkill #AgentSkills #PDFExtraction #OnDemandContext #KnowledgeBase #ClaudeCodeSkill

microsoft/agent-governance-toolkit — daily trending #10 (2026-07-28, resurging)

github.com/microsoft/agent-governance-toolkit — a runtime security and governance toolkit for autonomous AI agents.

What it does: provides policy enforcement, zero-trust identity, execution sandboxing and reliability engineering, and claims to cover all 10 of the 2026 OWASP Agentic Top 10 risks (10/10).
How it works: it intercepts every tool call, message send and delegation in deterministic application code before the model’s intent reaches the wire, making denied actions structurally impossible rather than merely “unlikely.” It touts sub-millisecond governance latency (p99 < 0.1ms), runs on Python 3.10+ and is MIT-licensed.
OWASP Agentic Top 10 coverage: goal hijacking, tool misuse, identity abuse, memory poisoning, cascading failures, rogue agents and other risks specific to autonomous agents.
Why it is surging: amid autonomous-agent containment failures and heightened industry focus on agent governance (the same day, governance startups like Hush Security raised), this April release climbed back into the top of the daily trending list.

Practical angle: before putting agents into production, install deterministic, code-level blocking gates in front of irreversible actions (send, deploy, pay, delete).

#AgentSecurity #OWASPAgenticTop10 #PolicyEnforcement #ZeroTrustIdentity #ExecutionSandboxing #RuntimeGovernance

vudovn/ag-kit — daily trending #19 (2026-07-28, new)

github.com/vudovn/ag-kit — an “Antigravity-first” scaffolding kit packaging Markdown-based agent engineering patterns.

What it does: bundles rules, skills, specialist agents, workflows, persistent memory, MCP guidance, orchestration and a native safety hook into one open-source agent engineering kit.
How it works: the @vudovn/ag-kit npm CLI installs Antigravity Kit templates (skills, agents, workflows) into a project. Antigravity is the supported production runtime; recent releases normalized the structure to .agents folder naming and tidied up the workflow, security and integration docs.
Differentiator / use cases: focuses not on “what to prompt” but on how to structure and standardize an agent’s working environment — shipping and versioning the rules, workflows and memory structures a team repeats.
Why it is surging: demand for an “install-and-go” agent scaffold targeting the Antigravity runtime, with a safety hook and persistent memory built in, lifted early adoption.

Practical angle: drop standard agent rules and safety hooks into a new repo with one CLI command to keep an entire team’s agent usage consistent.

#AgentEngineering #Antigravity #PersistentMemory #AgentWorkflows #SafetyHook #MCPGuidance

moeru-ai/airi — daily trending #3 (2026-07-28, surging)

github.com/moeru-ai/airi — a self-hosted, you-owned “AI companion” (cyber character), aiming for a Neuro-sama-style AI VTuber.

What it does: builds an AI companion you own and self-host in cyberspace, capable of real-time voice conversation and even playing games like Minecraft and Factorio.
How it works: designed from day one on web technologies — WebGPU, WebAudio, Web Workers, WebAssembly and WebSocket — so it runs in the browser and on macOS/Windows. Game play combines LLM-based natural-language understanding (the Minecraft bot) with computer vision plus LLM (Factorio, YOLO/LLM).
Differentiator / use cases: unlike other AI-VTuber open-source projects, its web-first design makes it highly portable, and it aims to be a “character container you own” rather than a finished service — useful for streaming/virtual characters and interactive desktop companions.
Why it is surging: demand for AI companions/VTubers combined with the self-hosting, privacy-first trend reignited it into the daily top ranks.

Practical angle: run a brand-specific virtual character or live-streaming persona with real-time voice on your own infrastructure.

#AIVTuber #DigitalCompanion #RealtimeVoiceChat #WebGPU #GamePlayingAI #SelfHosted

Excluded today: pascalorg/editor (#1, a 3D architectural editor where AI is a secondary MCP feature — off-topic), affaan-m/ECC (#5, large perennial), andrewyng/aisuite (#4) and huggingface/speech-to-speech (#6) as established, pbakaus/impeccable (#16, trending since mid-July), alibaba/open-code-review (#18, no re-acceleration), plus non-AI repos.