Alibaba open-sourced the AI code reviewer it ran internally for two years. Open Code Review wraps an LLM in deterministic engineering, reporting higher precision at one-ninth the tokens of a general-purpose agent on the same model. https://github.com/alibaba/open-code… #CodeReview #AiAgent #Alibaba
PixelRAG renders web pages and PDFs to screenshot tiles and searches the pixels with a LoRA-tuned vision embedder, so tables, charts, and layout survive the parse step: https://github.com/StarTrail-org/PixelRAG #RAG #Multimodal #Vision
Kun is a desktop agent workspace that starts coding from a written requirement, threading design, plans, code and review into one GUI. It defaults to three low-cost Chinese model providers so the extra rounds stay affordable. https://github.com/KunAgent/Kun #AiAgent #Coding #Electron
HTML Anything ships no model and no API key. It detects a coding-agent CLI you already logged into and drives it to turn notes or Markdown into ship-ready single-file HTML for WeChat, X, and Zhihu. https://github.com/nexu-io/html-anything #AIAgents #Html #LocalFirst
academic-research-skills-codex repackages Claude Code's Academic Research Skills so OpenAI Codex runs the same five human-in-the-loop workflows, pinned to a specific upstream commit. https://github.com/Imbad0202/academic-research-skills-codex #Codex #ClaudeCode #AcademicResearch
text-to-cad is a library of eleven agent skills that turn a coding agent into a CAD engineer, carrying a plain-language request from STEP model through DXF, slicing, and a cautious 3D-printer start. https://github.com/earthtojake/text-to-cad #Cad #AIAgents #Robotics
A curated, daily-updated library of 12,699 GPT Image 2 prompts in 16 languages where each entry is a structured JSON blueprint credited to an author and source. Browse by use case, style, or subject: https://github.com/YouMind-OpenLab/awesome-gpt-image-2 #GptImage2 #Prompts #AwesomeList
garden-skills is a curated, MIT-licensed set of Agent Skills for Claude Code, Cursor, and Codex that hard-codes taste and process: a five-dial Design Read, 25 named style recipes, 23 video themes, and checkpoint-gated workflows. https://github.com/ConardLi/garde… #AgentSkills #ClaudeCode #WebDesign
SkillOpt treats a single skill document as trainable state, editing it through validation-gated updates until a frozen model improves, then ships a compact best_skill.md that adds zero inference-time calls. https://github.com/microsoft/SkillOpt #AgentSkills #SelfEvolvingAgents #LLM
FreeLLMAPI aggregates the free tiers of 18 LLM providers into a single OpenAI-compatible /v1 endpoint, then leans on per-key rate tracking and automatic failover to keep every request under quota. https://github.com/tashfeenahmed/freellmapi #LLM #OpenaiApi #Proxy
Terax bundles a terminal, editor, git graph, web preview, and an agentic AI panel into a roughly 8 MB Tauri and Rust binary that runs on your own API keys or local models, with no account and no telemetry. https://github.com/crynta/terax-ai #Terminal #RustLang #Tauri
OpenSquilla is a microkernel AI agent whose on-device SquillaRouter sends every turn to the cheapest capable model across 20-plus providers, and it learns to route from its own traffic. https://github.com/opensquilla/opensquilla #AIAgents #LLM #Routing
Browser Harness keeps its core to about a thousand lines and connects an LLM straight to Chrome over CDP. When a helper is missing, the agent writes it mid-task and keeps it for next time: https://github.com/browser-use/browser-harness #BrowserAutomation #LLM #AiAgent
Open Design is a local-first desktop app that hands design work to the coding agent already on your laptop, then exports prototypes, decks, images, and video as real HTML, PDF, PPTX, and MP4 files. https://github.com/nexu-io/open-design #DesignTools #AIAgents #Byok
DeepSeek-Reasonix is a Go coding agent for the terminal that protects DeepSeek's prompt prefix cache, from stable startup summaries to two separate model sessions, so long runs stay cheap. https://github.com/esengine/DeepSeek-Reasonix #Deepseek #CodingAgent #golang
An open-source Codex and Claude skill that turns raster logos into animated SVGs, but only after a QA-verified static vector passes as the final-frame contract. Smoothness beats raw IoU. https://github.com/nolangz/pixel2motion #SvgAnimation #LogoMotion #ClaudeSkill
CodexPro runs a local MCP server that lets ChatGPT Developer Mode read, edit, and verify a single repo through a defined tool set, with token-gated tunnels and guardrails on writes and bash. https://github.com/rebel0789/codexpro #MCP #Chatgpt #CodingAgent
Apple's coreai-models bundles Core AI export recipes, a Swift runtime package, and coding-agent skills for Claude Code, Codex, and Gemini. Curated gallery, no PRs, macOS/iOS 27+. https://github.com/apple/coreai-models #Apple #OnDeviceAi #Coreai
MCP servers expand your AI agent's capabilities—and its attack surface. Here's what to lock down before something goes wrong. https://www.agentpalisade.com/resources/mcp-security-checklist #AI #infosec #LLM
An AI-agent skill that turns finished Markdown into paste-ready WeChat HTML, using six themed component libraries and two Python validators to enforce the platform's filtering rules deterministically. https://github.com/isjiamu/gzh-design-skill #AgentSkill #Wechat #Markdown
AI agents act on the world — they read files, run commands, and call APIs. Here is a practical framework for limiting damage when their reasoning is subverted. https://www.agentpalisade.com/resources/ai-agent-security-checklist #AI #infosec #LLM
Forsy-AI/agent-apprenticeship treats completed agent work as portable 'experience compilations' you can install into later runs. Broad agent support, a 39k-record seed dataset, and value-per-task claims worth checking. https://github.com/Forsy-AI/agent-apprenti… #AIAgents #PostTraining #AgentTraces
RAG systems introduce new attack surfaces beyond the prompt layer—ingestion pipelines, vector stores, and retrieval logic each require independent security controls to avoid silent data exposure. https://www.agentpalisade.com/resources/rag-security-checklist #AI #infosec #LLM
A local tool that makes any LLM 'watch' a video by picking scene-change keyframes, deduping near-identical shots, and transcribing audio into a folder a model can read. https://github.com/HUANGCHIHHUNGLeo/claude-real-video #Python #LLM #Video
Why guard instructions and input classifiers can't fully solve prompt injection, and a 16-point checklist covering prompt construction, indirect injection, output handling, action gating, and red-team testi… https://www.agentpalisade.com/resources/prompt-injection-review-checklist #AI #infosec #LLM
LLM applications inherit traditional web vulnerabilities while adding new attack surfaces from non-deterministic components — a practical checklist covering trust boundaries, output validation, data exposu… https://www.agentpalisade.com/resources/llm-application-security-checklist #AI #infosec #LLM
AI coding assistants produce code that compiles and looks correct—but research shows roughly 40–45% of AI-generated samples contain security vulnerabilities, and developers tend to trust the output more… https://www.agentpalisade.com/resources/ai-generated-code-security-checklist #AI #infosec #LLM
A practical breakdown of which SMB workflows benefit most from AI automation—organized by business function—plus a simple filter for spotting good candidates before you build anything. https://www.agentpalisade.com/resources/ai-automation-use-case-library #AI #productivity #automation
Dao Code is an MIT-licensed TypeScript terminal coding agent for DeepSeek V4 that keeps a byte-stable prefix and runs memory and self-review on cache-reusing forks, claiming ~95.8% cache hits on real OSS bug fixes. https://github.com/tigicion/dao-code #Deepseek #CodingAgent #TypeScript
n8n treats prototype-to-production as the real problem for AI agents: model-swappable, self-hostable, with human approvals and audit trails as first-class parts. A builder's look at what that bet buys you. https://github.com/n8n-io/n8n #AI #automation #Workflow
