Dev Digest — May 8, 2026

🔥 HOT RELEASES

OpenAI Codex Chrome Extension Codex now works directly inside Chrome on macOS and Windows. It can test web apps, use DevTools, navigate across parallel tabs in the background, and handle repetitive browser tasks — all without taking over your browser. Per-site access controls included. Not just coding anymore — this is agentic browser work. 🔗 X: https://x.com/OpenAI/status/2052480800004956323 📖 Changelog: https://developers.openai.com/codex/changelog#codex-2026-05-07

DeepSeek-TUI — Terminal Coding Agent The week's breakout project. A Rust-based terminal coding agent for DeepSeek V4 that exploded to 20K+ stars. Three modes: Plan (read-only), Agent (approval gates), YOLO (full auto). Supports MCP, sub-agents, session resume, workspace rollback, 1M-token context. Your code stays local. MIT licensed. 🔗 X: https://x.com/midudev/status/2052012365562954011 💻 GitHub: https://github.com/Hmbown/DeepSeek-TUI

TokenSpeed — Open-Source LLM Inference Engine LightSeek Foundation released TokenSpeed, targeting TensorRT-LLM performance for agentic workloads. ~9% lower latency at batch 1, ~11% higher throughput at 100 TPS. Compiler-backed SPMD modeling, C++ FSM scheduler, and one of the fastest MLA kernels on NVIDIA Blackwell. MIT license. 🔗 X: https://x.com/Marktechpost/status/2052512628300460512 💻 GitHub: https://github.com/lightseekorg/tokenspeed

XcodeBuildMCP 2.5.0 Biggest update ever for the Xcode ↔ MCP bridge. New response formats (text/JSON/NDJSON/raw), streaming progress for builds, structured content support for agents, JSON schemas for response verification, and optimized token usage. 🔗 X: https://x.com/camsoft2000/status/2052512893015613446 📖 Site: https://www.xcodebuildmcp.com/

Expo CI/CD Workflows API Expo shipped a REST endpoint to trigger CI/CD workflows with a single HTTP request. No CLI or dashboard needed — POST to /workflows/dispatch with a bearer token and workflow YAML. Any platform can now kick off Expo builds. 🔗 X: https://x.com/expo/status/2052530653263167603

🧪 INTERESTING REPOS

ds4 + pi-ds4 (Armin Ronacher / antirez) antirez's ds4 is a minimal local model runner, and mitsuhiko (Flask creator) built a pi extension for it. One command pi install on a 128GB Mac and you get DeepSeek running locally with tool streaming. Mitsuhiko wrote a blog post arguing we need to focus local model efforts more. 🔗 X: https://x.com/mitsuhiko/status/2052508753472143614 💻 GitHub: https://github.com/mitsuhiko/pi-ds4 📖 Blog: https://lucumr.pocoo.org/2026/5/8/local-models/

video-use — AI Video Editor Browser-use team released an open-source video editing tool for Claude Code. Drop raw footage + resources in a folder and it handles cuts, filler word removal, subtitles, color grading, animations, and final render. No timeline, no manual editing. 🔗 X: https://x.com/0xluffy_eth/status/2052676640229282172 💻 GitHub: https://github.com/browser-use/video-use

InsForge — Backend for Coding Agents Full-stack Postgres backend (Auth + Storage + Compute + Hosting + AI Gateway) purpose-built for coding agents. Also provides context engineering for Claude Code — users report 3x fewer tokens used. 🔗 X: https://x.com/DailyDoseOfDS_/status/2052668450733228345 💻 GitHub: https://github.com/InsForge/InsForge

agent-skills (Addy Osmani) Production-grade engineering skill pack for AI coding agents. Shell scripts, best practices, tool use patterns. Works with Claude Code, Cursor, Codex. 3K+ stars in a day. 💻 GitHub: https://github.com/addyosmani/agent-skills

ruflo — Claude Agent Orchestration Multi-agent swarm platform with Claude Code native integration. Self-learning workflows + RAG. 2K+ stars in a day. 💻 GitHub: https://github.com/ruvnet/ruflo

local-deep-research Local deep research agent supporting all local/cloud LLMs, 10+ search engines (arXiv, PubMed, private docs), ~95% SimpleQA accuracy. Everything encrypted locally. 💻 GitHub: https://github.com/LearningCircuit/local-deep-research

PageIndex (VectifyAI) Vector-free, reasoning-based RAG document indexing. 950+ stars in one day. 💻 GitHub: https://github.com/VectifyAI/PageIndex

Polytoria 2.0 — Open Source Game Engine Game creation platform rebuilt from the ground up after a year of development. Now fully open source. 🔗 X: https://x.com/PolytoriaGame/status/2052664541595496525 💻 GitHub: https://github.com/Polytoria/polytoria-game

🎥 WORTH WATCHING

Multi-Agent Architecture That Actually Ships (Factory AI) Luke from Factory AI breaks down a system that ran 16 days straight using orchestrators, workers, and validators. Validation contracts before implementation, structured handoffs, adversarial verification. 🔗 https://www.youtube.com/watch?v=ow1we5PzK-o

Qwen3.6-27B in Action — Agentic Workflows Alibaba's Tongyi Lab demos their new dense model tackling complex agentic coding workflows. 🔗 https://www.youtube.com/watch?v=7ejQSGaiyQE

Claude Code Team: Automate Engineering Workflows in 30 Minutes Straight from the people who built Claude Code — a free walkthrough on automating your engineering workflow. 🔗 X: https://x.com/cgtwts/status/2052717242186690974

💡 TECHNIQUES & IDEAS

Claude Code's 5-Folder System A breakdown of the Agent Development Kit structure gaining traction: CLAUDE.md (repo rules), skills/ (reusable workflows), hooks/ (guardrails via shell scripts), subagents/ (isolated context), plugins/ (team distribution). Turns Claude from a chat tool into scalable engineering infrastructure. 🔗 X: https://x.com/berryxia/status/2052719498021773349

CrewAI Ships Checkpointing Agent workflows now get Google Docs-style autosave. Resume from any point, fork from saved states, edit past outputs and watch changes ripple. Visual TUI to inspect everything. 100% open source. 🔗 X: https://x.com/akshay_pachaar/status/2052691407421538805 💻 GitHub: https://github.com/crewAIInc/crewAI

OpenAI Realtime Prompting Guide New guide for GPT-Realtime-2 voice apps: tuning reasoning effort, preamble design, tool behavior, handling unclear audio, entity capture, and state management in longer sessions. 🔗 https://developers.openai.com/api/docs/guides/realtime-models-prompting

TRAE Editor for Unity + XR ByteDance's TRAE AI is collaborating with PICO on AI-native XR development — bringing AI coding directly into Unity for faster VR/AR prototyping. 🔗 X: https://x.com/Trae_ai/status/2052672708958441544

🔮 EMERGING TRENDS

Local Models Getting Real — mitsuhiko's passionate blog post argues we need to focus local model efforts more seriously. With antirez's ds4, you can pi install and have a full local coding agent on a Mac. DeepSeek-TUI's explosion (0 → 20K stars in days) shows massive demand for local-first AI coding.

Agent Infrastructure Maturing — InsForge, ruflo, agent-skills, CrewAI checkpointing — the ecosystem is moving past "can agents code?" to "how do we make agent workflows reliable, resumable, and production-grade?" Backend-as-a-service specifically for agents is a new category.

Browser as Agent Workspace — Codex's Chrome extension, Browser Use's growing ecosystem, and Higgsfield's MCP-powered ad automation all point to the browser becoming the primary surface for agent work beyond code.

Context Engineering > Prompt Engineering — The Claude Code 5-folder system, InsForge's 3x token reduction, and the general discourse around CLAUDE.md and skills suggest the community is shifting from "write better prompts" to "architect better agent environments."

Full digest: https://digest.99.cooking