agents-libagents-lib
AgentsMCP ServersWorkflowsTutorialsComparisonsTemplatesNewsAbout
OpenWorkCodex SecurityMicrosoft AI SecurityKimi K3AiriVibeVoiceGitHub TrendingNews Roundup

AI News Roundup: July 29, 2026 — OpenWork Launches as Open-Source Cowork Alternative, OpenAI Ships Codex Security, and Microsoft Enters Cybersecurity AI

OpenWork: Open-Source Alternative to Claude Cowork Hits 17K Stars

different-ai/openwork has launched as "the open-source alternative to Claude Cowork and Codex," powered by OpenCode. At 17,400+ GitHub stars in its first days, it is rapidly gaining traction as a desktop-first workflow platform.

Key features:

  • Workflow Sharing: Create AI workflows once and reuse skills, MCPs, and connected services across tools, teammates, and machines
  • Desktop Application: Dedicated workspace for macOS, Windows, and Linux (optional — CLI works standalone)
  • MCP Integration: One OpenWork MCP integrates with Codex, Claude Code, Cursor, and compatible agents
  • OpenWork Den: Control plane for managing OpenWork across a team or organization
  • Enterprise Features: Inference at scale, team access management, skills/plugin marketplaces
  • Install via AI Agent: "Install OpenWork on my computer, set up my workspace" — agent-native setup

Why it matters: OpenWork is the first serious open-source competitor to Claude's Cowork. With 17K stars and an agent-native installation model, it signals that the multi-agent workflow orchestration market is rapidly fragmenting toward open ecosystems.

OpenAI Ships Codex Security — Security Scanning in Your CI

OpenAI has released openai/codex-security, a CLI and TypeScript SDK for security vulnerability analysis. It scans repositories, reviews changes, tracks findings over time, and integrates into CI pipelines.

Key features:

  • CLI scanner: npx codex-security scan . for quick repo-wide security scans
  • TypeScript SDK: CodexSecurity class for programmatic CI integration
  • Change review mode: Focus security checks on pull requests and diffs
  • CI-native: API key authentication via OPENAI_API_KEY for automated pipelines
  • Login flow: npx codex-security login with ChatGPT sign-in or API key support

Why it matters: OpenAI is moving beyond code generation into code security. Codex Security is the first major product tying the ChatGPT model directly to a CI-native vulnerability scanning workflow.

Microsoft Launches First Cybersecurity AI Model + Agentic Security System

Microsoft has debuted its first cybersecurity-focused AI model alongside a new agentic cybersecurity system, entering a space long dominated by security-specialist vendors. This positions Microsoft as a first-party AI security provider for enterprises.

Kimi K3 Architecture: 2.8 Trillion Parameters, NoPE Everywhere

Moonshot AI's Kimi K3 is a 2.8-trillion-parameter open-weight model — the largest of its kind. Sebastian Raschka's detailed analysis reveals a number of significant architectural innovations.

Key architectural innovations:

  • LatentMoE: Compresses large linear layers using a down-projection approach similar to multi-head latent attention (shared with Nemotron 3 Ultra)
  • Attention Residuals: Connects residuals across layers weighted by attention scores — ~4% training cost, ~2% inference cost, improved validation loss
  • NoPE Everywhere: Kimi K3 removes all RoPE positional embeddings — the first frontier model to adopt NoPE across the board
  • Native multimodal support
  • Multi-Head Latent Attention and Kimi Delta Attention hybrids

Why it matters: Kimi K3 is a watershed moment for open-weight models at scale. The NoPE-only approach is a major architectural bet that, if validated, could change how position embeddings are used across the frontier model space.

Other Notable Stories

  • Codex Security (OpenAI) — Security vulnerability scanner as CLI + TypeScript SDK, CI-native (GitHub)
  • LearnVector (Andrew Ng) — New AI company focused on "one-to-one learning experiences" (learnvector.ai)
  • Airi (moeru-ai) — Hits 45,119 stars, the leading self-hosted AI companion platform with real-time voice chat and gaming integration
  • ECC Agent Harness — Surges past 235K stars, now the most-starred AI agent project on GitHub
  • superpowers — Reaches 263K stars as the dominant agentic skills framework
  • VibeVoice (Microsoft) — 51K stars for open-source frontier voice AI
  • JCode — 13K stars for "the most RAM efficient harness"
  • FlashKDA (Moonshot) — High-performance Kimi Delta Attention kernels (883 stars, trending)
  • Transformers Transformer — Unified model for motion-conditioned robot co-design

Key Themes

  1. Security Becomes First-Class: OpenAI Codex Security and Microsoft's cybersecurity AI signal that security is shifting from a niche to a platform feature — agents that write code now ship with agents that scan it.

  2. Open-Source Cowork: OpenWork (17K stars) is the first credible open alternative to Claude Cowork. The workflow-sharing, multi-agent, multi-tool marketplace model is being democratized.

  3. Frontier Model Architecture Divergence: Kimi K3's NoPE-only design and LatentMoE approach are breaking with the established architectural defaults. Expect more models to experiment outside the standard RoPE + full-attention pattern.

  4. Voice AI Goes Open: Microsoft VibeVoice (51K stars) joining HuggingFace's speech-to-speech (7.6K stars) shows voice agent infrastructure is rapidly becoming open-source.

  5. Harness Model Dominance: ECC (235K) and superpowers (263K) continue to lead GitHub — the "wrap your existing agent" approach is winning over "replace your framework."