Criterion: ranked by recent star velocity (how much a repo gained yesterday-to-today) over the past 24 hours by absolute time (UTC), not by cumulative size. Data sources: Trendshift daily (freshness verified by decoding the top Live-mentions X post IDs — 2026-07-21 18:58 UTC snapshot, ~2 hours before run time) and findarepo 1-day deltas. Cumulative stars are reference only.
bojieli/ai-agent-book — +5.7k ⭐ today (~14k total) · 2 days running
https://github.com/bojieli/ai-agent-book
This is the open-source home of the free Chinese-language book Understanding AI Agents In-Depth: Design Principles and Engineering Practice by Bojie Li — it bundles the full text, a compiled PDF, and chapter-by-chapter code. Across 10 chapters it walks through the core building blocks of production agents, and each chapter ships runnable implementation examples (e.g. AdaptThink for self-adaptive reasoning modes, embodied-agent training on the AWorld framework). Beyond the Chinese original, community-contributed English, Tamil, and Vietnamese PDF translations are included, which broadened its reach fast. It jumped from roughly 2k to ~14k stars in a matter of days and held the #1 spot on Trendshift’s daily board for a second day; its 1-day gain (+5.7k) is larger than yesterday’s, confirming it is still accelerating rather than plateauing. The topics covered include:
- Context engineering — KV-cache-friendly design, prompt/dynamic prompts, and context compression
- User memory and knowledge bases — RAG pipelines and retrieval beyond flat text
- Tools — tool taxonomy, the MCP protocol, perception/execution/collaboration tool types, and event-driven async agents
- Coding agents — code as “a tool that can build new tools,” the meta-capability of a general agent
Practical use: drop it straight into team onboarding or an internal study group as an agent-engineering primer, and start a PoC from the per-chapter code.
Tags: #AIAgent #ContextEngineering #RAG #MCP #CodingAgent #AgentArchitecture
lidge-jun/opencodex — New surge, #3 on Trendshift daily (~1k total)
https://github.com/lidge-jun/opencodex
opencodex is a local “provider proxy” that lets you run OpenAI Codex and Claude Code on any LLM backend. It translates Codex’s Responses API into whatever wire format your provider speaks in real time, passing streaming, tool calls, reasoning tokens, and images through in both directions. When you omit the provider/ prefix it routes to your default provider, or auto-matches on the model name pattern (claude-* → Anthropic, gpt-* → OpenAI). You install it from npm (@bitkyc08/opencodex), inject it into Codex with ocx init, launch the proxy with ocx start, and it also ships a web dashboard (ocx gui). The pitch — wire up any model without waiting for official support — pushed it into the daily top ranks right after a fresh preview release. Its five adapters cover:
- Anthropic Messages
- Google Gemini
- Azure
- OpenAI Responses passthrough
- Every OpenAI-compatible Chat Completions endpoint — 40+ built-in providers (Claude, Gemini, Grok, GLM, DeepSeek, Kimi, Qwen, Ollama, and more)
Practical use: point Codex/Claude Code at models your team already subscribes to and run coding agents without vendor lock-in.
Tags: #CodexCLI #ProviderProxy #ClaudeCode #LLMGateway #OllamaLocal #ResponsesAPI
ayghri/i-have-adhd — +2.7k ⭐ today (~6.7k total) · 2 days running
https://github.com/ayghri/i-have-adhd
i-have-adhd is a “skill” file that stops a coding agent (Claude Code in particular) from burying the answer in verbose explanation. It disciplines the agent’s output style to lead with the key conclusion and stay concise — “ADHD-friendly.” Because it’s a rule/skill layered onto the agent rather than a separate installable tool, adoption is lightweight and the effect is felt immediately in workflows where long-winded answers slow you down. It spread via an X (Twitter) viral moment on July 18–20 and jumped from ~2.4k to ~6.7k stars in a single day (+2.7k), a larger daily gain than yesterday’s +1.7k, keeping it near the top of Trendshift’s daily board for a second day. (Its daily rank eased from #3 to #12 as other repos surged, but its absolute daily velocity increased.)
Practical use: layer it onto your team’s standard agent prompt to enforce an “answer-first” format for reviews and responses.
Tags: #AgentSkill #ClaudeCode #PromptEngineering #DeveloperUX #OutputFormatting
Excluded: koala73/worldmonitor (Trendshift daily #2 but a large 62.8k-star project growing only ~+200/day — a “rank holder,” not a fresh spike); tirth8205/code-review-graph (yesterday’s pick, daily rank slipped #2→#5 with no re-acceleration); MoonshotAI/kimi-cli (decayed out of the top ranks).
Daily AI Brief 
Leave a comment