Selection rule: only repos climbing right now. The primary metric is recent momentum — stars gained today — with lifetime stars as reference only. Data: Trendshift daily (live), cross-checked against findarepo’s 7-day growth (2026-07-13 snapshot). Large-but-flat “ranking holders” and previously featured repos whose growth has rolled over are excluded.
JustVugg/colibri — +3,100 ⭐ today (2 days running)
10.4k ⭐ total · C · Apache-2.0 · https://github.com/JustVugg/colibri
Colibri is a pure-C inference engine that runs GLM-5.2, a 744B-parameter MoE model, on a consumer machine with about 25GB of RAM. The trick is refusing to load everything into memory: only the dense portion (~9.9GB at int4) stays resident, while the 21,504 routed experts (~370GB total) are streamed from disk on demand through a per-layer LRU cache. Its components include MLA attention with a 57× compressed KV cache, native MTP speculative decoding hitting 2.2–2.8 tokens per forward at int8, AVX2 integer-dot kernels, async expert readahead, a learning cache that auto-pins hot experts, and an offline FP8→int4 converter that never needs the full 756GB checkpoint on disk at once. The whole engine is roughly 1,300 lines of pure C with no CUDA, no PyTorch and no BLAS — that constraint is the project’s entire identity. Throughput is 0.05–0.1 tok/s cold on the dev box (WSL2, 25GB RAM, ~1GB/s NVMe) and around 1 tok/s on an Apple M5 Max with 128GB RAM and a fast SSD, so the appeal is less practicality than the plain fact that it works at all — which is why it accelerated from +2,000 yesterday to +3,100 today.
Where it fits in practice: research, teaching and PoC work where you need to inspect and verify a frontier-scale MoE model’s behaviour without a GPU budget.
#LocalLLM #MoE #ExpertOffloading #Int4Quantization #SpeculativeDecoding #CLang
stablyai/orca — +1,300 ⭐ today (re-entry)
18k ⭐ total · TypeScript · https://github.com/stablyai/orca
Orca is an ADE (Agent Development Environment) for driving a fleet of coding agents at once — Claude Code, Codex CLI, OpenCode, Pi, Grok and 20+ others, all running on your own subscriptions and API keys. The core mechanic is parallel worktrees: run anywhere from 1 to 50 agents side by side, each isolated in its own git worktree, then compare the resulting diffs and merge the winner. Around that sit Design Mode, where clicking any UI element in an embedded Chromium window captures its HTML, CSS and a screenshot straight into the agent’s prompt; Ghostty-class terminals with WebGL rendering, infinite splits and restart-proof scrollback; and a review loop that lets you drop comments on any diff line and ship them back to the agent. It integrates natively with GitHub and Linear, so opening a worktree from a PR, issue or ticket carries the task context automatically and leaves the diff ready to review in-app, while iOS and Android companion apps let you monitor runs and send follow-up prompts from anywhere. Orca makes a point of never proxying or reselling API access — your keys call each provider directly from your machine. It was featured on July 12, cooled off for two days, and is back on renewed acceleration.
Where it fits in practice: throw one ticket at several agents simultaneously and pick the best diff — turning agent quality variance from a risk into a menu.
#AgentOrchestration #GitWorktree #ParallelAgents #CodingAgents #DevTooling
HKUDS/Vibe-Trading — +1,200 ⭐ today
22.9k ⭐ total · Python · MIT · 47 contributors · https://github.com/HKUDS/Vibe-Trading
From the University of Hong Kong’s data science lab (HKUDS), Vibe-Trading is an open-source research workspace that turns finance questions into runnable analysis. It works by wiring natural-language prompts to market-data loaders, strategy generation, backtest engines, reports and exports, plus a persistent research memory that carries session context across strategy code, metrics, benchmarks, validation artifacts and run cards. It also parses broker journals to diagnose your trading behaviour, extract rules and compare them against a Shadow Account, and runs multi-agent research reviews from investment, quant, crypto, macro and risk perspectives. Strategies export to TradingView (Pine Script v6), TDX and MetaTrader 5 (MQL5), and it exposes 22 MCP tools as a stdio subprocess for any MCP-compatible client — 21 of the 22 need no API key at all. Autonomous trading only runs through a broker you authorize yourself (Tiger, Alpaca, OKX, Binance, Futu, Robinhood) behind a user-committed mandate covering symbol universe, order size, exposure, leverage and daily cap, plus a filesystem kill switch, a fail-closed pre-trade gate and a full audit ledger — it holds no funds and can be halted instantly, and that safety model is a large part of why it is climbing. It hit #1 on GitHub Trending on July 12 and has kept growing since.
Where it fits in practice: stand up a research pipeline that takes a strategy idea through backtest, validation and reporting, while keeping any live execution fenced in by a mandate.
#TradingAgent #Backtesting #MCP #Fintech #PineScript #MultiAgentReview
Dicklesworthstone/destructive_command_guard — +962 ⭐ today (new)
2.2k ⭐ total · Rust · https://github.com/Dicklesworthstone/destructive_command_guard
dcg is a guardrail that stops AI coding agents from running catastrophic commands like git reset --hard, rm -rf ./src or DROP TABLE users and destroying hours of uncommitted work in seconds. It works by inspecting each shell command an agent tries to run in sub-millisecond time: if it matches a destructive pattern, dcg returns a JSON denial to the agent and prints a human-readable warning plus safer alternatives to stderr; otherwise it passes silently. Technically it leans on SIMD-accelerated regex filtering to hold that sub-millisecond latency, scans heredocs and inline scripts (things like python -c "os.remove(...)") for hidden destructive patterns, and enforces a hard 200ms processing ceiling so no command can ever hang — anything over the limit is allowed with a logged warning. It ships 50+ modular security packs spanning git, filesystem, PostgreSQL, Docker, Kubernetes, AWS, GCP and Azure.
Supported harnesses cover essentially the whole field:
- Claude Code, Codex CLI 0.125.0+, Gemini CLI
- GitHub Copilot CLI, VS Code Copilot Chat, Cursor IDE
- Hermes Agent, Grok (xAI), Antigravity CLI (agy)
- OpenCode, Pi, Aider, Continue
Word spreading on X about hooking it into the pre-validation hooks most harnesses already expose is what produced a +962 day against just 2.2k lifetime stars — the steepest relative slope on this list.
Where it fits in practice: the cheapest possible safety net when you want to give agents shell access but selectively block the irreversible mistakes — one hook line to install.
#AgentGuardrails #CommandSafety #ShellSecurity #RustLang #SIMDRegex #PreToolUseHook
Nutlope/hallmark — +853 ⭐ today (new)
3.4k ⭐ total · CSS · MIT · https://github.com/Nutlope/hallmark
Hallmark is an anti-AI-slop design skill that makes the UIs Claude Code, Cursor and Codex generate look made rather than generated. It works by enforcing a tight rule set at the prompt stage so the model cannot fall back on the defaults every LLM was trained into — the familiar hero → 3-feature → CTA → footer rhythm. Those rules are distilled from Anthropic’s frontend-design skill, the Claude cookbook on frontend aesthetics, and the 2026 “tactile rebellion” movement. Its sharpest differentiator is demanding structural variety rather than merely visual variety: two pages built from different briefs should feel like different sites, not colour swaps of one template. It also runs a verification loop, scoring its own output before shipping it.
Every design is scored 1–5 across six axes, and anything below 3 triggers a revision pass:
- Philosophy
- Hierarchy
- Execution
- Specificity
- Restraint
- Variety
It first surfaced on the trending list back on May 21, went quiet, and has now re-ignited with +853 in a day against 3.4k lifetime stars.
Where it fits in practice: fixes the “every agent-built landing page looks the same” problem with a single skill file — useful for raising the floor on design quality without a designer in the loop.
#FrontendDesign #AntiAISlop #AgentSkills #UIGeneration #DesignSystems
Excluded today (for reference)
- Graphify-Labs/graphify (85k total, +1.4k today): growth is top-tier but it’s a large, long-running trending holder — permanently excluded.
- OpenCut-app/OpenCut (+2.4k today) and 1c7/chinese-independent-developer (+1.9k today): high growth, but not AI projects.
- Shubhamsaboo/awesome-llm-apps (+1.1k today) and mattpocock/skills (168k total): large, mature curation lists.
- MadsLorentzen/ai-job-search and iOfficeAI/OfficeCLI: featured yesterday (07-14), both dropped out of today’s daily growth leaders — decay.
Daily AI Brief 
Leave a comment