AI Open Source Radar — July 21, 2026

Basis: top of Trendshift’s daily list (snapshot for July 20, 2026 UTC). Cache freshness was verified by decoding the timestamps of the top live-mention posts (20:52–20:56 UTC) — not a stale cache. Recent star growth within the trending window is the primary signal; cumulative stars are shown for reference. Yesterday’s picks all cooled and are excluded today: MoonshotAI/kimi-cli (fell from #2 to #23 daily) and lopopolo/harness-engineering and ibelick/ui-skills (both dropped out of the top ranks).

bojieli/ai-agent-book — +4,300 ⭐ (recent · #1 daily) · ~6,000 ⭐ total

URL: github.com/bojieli/ai-agent-book

This is the open-source home of the book Understanding AI Agents: Design Principles and Engineering Practice (English title “AI Agents in Depth”) by systems-and-AI researcher Bojie Li, publishing the full text, a compiled PDF, and per-chapter sample code. The book/ directory holds the complete manuscript and PDF, while chapter directories such as chapter1/ contain topical example code (for instance, context management), and translations are provided in English, Tamil and Vietnamese alongside the Chinese original. What sets it apart is that it doesn’t stop at theory: each chapter’s concepts are paired with runnable code, so readers can learn and reproduce everything from agent design principles to real engineering in a single repository. It reached #1 on GitHub Trending (all languages) on July 19, surging from roughly 2k to 6k stars in a few days, and it still holds #1 on Trendshift’s daily list today.

Practical / customer use: drop the per-chapter code straight into an internal agent study group or onboarding curriculum.

Tags: #AIAgentDesign #AgentEngineering #ContextEngineering #OpenTextbook #LLMOrchestration

tirth8205/code-review-graph — +1,700 ⭐ (recent · #2 daily · GitHub Trending #1 on Jul 19) · ~21,000 ⭐ total

URL: github.com/tirth8205/code-review-graph

code-review-graph is a local-first code-intelligence tool (MCP + CLI, Python, MIT) that indexes a codebase as a persistent knowledge graph so AI coding tools read only the files that matter. It parses the repository into an AST with Tree-sitter, storing functions, classes and imports as nodes and calls, inheritance and test coverage as edges, then computes the minimal set of files the assistant needs at review time. Key characteristics:

Re-indexes a 2,900-file monorepo in under two seconds and watches for file changes to rebuild the graph automatically.
Delivers 6.8x average context compression, up to 49x on large monorepos (benchmarked).
Ships 28 MCP tools (get_impact_radius, get_review_context, detect_changes, get_architecture_overview, and more).
Automatically attaches context_savings metadata to tool responses so agents can surface the savings to humans with no extra prompting.

The surge followed a major release (six community PRs and six new MCP tools) that carried it to #1 on GitHub Trending on July 19; it sits at #2 on Trendshift’s daily list today.

Practical / customer use: use it as a context gate to cut token cost and hallucinations from Claude Code or Codex on large repos.

Tags: #CodeIntelligence #KnowledgeGraph #TreeSitter #ContextCompression #MCPServer #LocalFirst

ayghri/i-have-adhd — +1,700 ⭐ (recent · #3 daily) · ~2,400 ⭐ total

URL: github.com/ayghri/i-have-adhd

i-have-adhd is a Claude Code skill (Shell, a single SKILL.md) that forces the model’s output format so Claude Code or Codex stops burying the answer under a wall of text. Activate it with /i-have-adhd and it applies ten rules to the current session, including:

Lead the first line with the action or answer.
Number multi-step tasks.
Report progress each turn.
State concrete times and quantities.
List at most five items at once.
Suppress tangents and filler.
Ban cliché openers like “Great question.”
End with a single, clear next step.

Because it runs from one SKILL.md file with no hooks or background process, the rules are easy to customize, and it essentially ports the principles of an ADHD toolbook into an LLM output spec. Developers shared it repeatedly on X (Twitter) from July 18–20 as the skill that “kills AI’s filler,” driving it from roughly 700 to 2,400 stars in a few days.

Practical / customer use: attach it to answer-first workflows such as code review or ticket responses to make AI output more readable and actionable.

Tags: #ClaudeSkill #PromptEngineering #AgentOutput #DeveloperUX #SKILLmd