Hottest Repos
Hmbown/DeepSeek-TUI — 18.4K stars (+5,787 today). The run that won't stop. From 1.1K five days ago to 18.4K — a 17x explosion. The Rust terminal coding agent for DeepSeek V4 just posted its biggest single day yet, gaining nearly 6K. Three modes (Plan/Agent/YOLO), 1M-token context, LSP integration. This is no longer a niche — it's the leading non-Claude coding agent. https://github.com/Hmbown/DeepSeek-TUI
addyosmani/agent-skills — 32.7K stars (+3,058 today). Biggest daily gain yet. The engineering-discipline skills repo jumped from 30.2K to 32.7K overnight — nearly 6x its daily average. 20 skills across the full dev lifecycle, anti-rationalization tables, explicit failure modes. The "behavioral files" category has its infrastructure hero. https://github.com/addyosmani/agent-skills
mattpocock/skills — 64.6K stars (+20.7K this week). Week's biggest weekly gainer. Matt Pocock's .claude directory gained more stars in a week than most repos earn in a lifetime. The TypeScript wizard's approach to agent guardrails clearly resonates. Skills-as-code is now a mass-market category. https://github.com/mattpocock/skills
anthropics/financial-services — 11.2K stars (+1,367 today). Day 2, accelerating. Daily gain more than doubled from yesterday's 540 to 1.3K. 10 named agents (Pitch Agent, KYC Screener, GL Reconciler), 11 MCP connectors (FactSet, S&P, Moody's), 40+ slash commands. Anthropic showing exactly how enterprise vertical AI adoption works — pre-built agents, not generic chat. https://github.com/anthropics/financial-services
ruvnet/ruflo — 46K stars (+10.9K this week). Eighth consecutive week of acceleration. The Claude agent swarm orchestrator crossed 46K, up from 43.4K two days ago. The daily gains have shifted from spiky to consistent — 1.5-2.5K daily like clockwork. Multi-agent orchestration is a permanent category. https://github.com/ruvnet/ruflo
AI & ML
z-lab/dflash — 3.4K stars (+654 today). New — speculative decoding gets practical. Block diffusion model that drafts tokens in parallel for faster LLM inference. Supports 17+ models (Gemma 4, Qwen 3.6, Kimi K2.5, DeepSeek V4 coming). Works with vLLM, SGLang, Transformers, and MLX. Training recipe coming soon. The kind of infrastructure that makes local inference actually usable. https://github.com/z-lab/dflash
VectifyAI/PageIndex — 29.4K stars (+953 today). Vectorless RAG crosses 29K. Builds a hierarchical tree index from documents, then uses LLM reasoning (not vector similarity) to retrieve. 98.7% on FinanceBench — state-of-the-art for document RAG. The "similarity ≠ relevance" pitch is landing. https://github.com/VectifyAI/PageIndex
TauricResearch/TradingAgents — 71K stars (+15.5K this week). Crossed 71K, still accelerating. Weekly gains (15.5K) are the highest ever. Eighth consecutive week of massive growth. The multi-agent trading framework has become finance-AI's gravitational center. https://github.com/TauricResearch/TradingAgents
LearningCircuit/local-deep-research — 6.1K stars (+564 today). 95% SimpleQA with Qwen3.6-27B on a 3090. Supports all local and cloud LLMs, 10+ search engines (arXiv, PubMed, private docs). Everything local and encrypted. The "deep research without API bills" play. https://github.com/LearningCircuit/local-deep-research
AIDC-AI/Pixelle-Video — 13.2K stars (+5K this week). Ninth week trending. The automated short-video engine has gone from novelty to institution. Text-to-finished-video with consistent 500-1K daily gains. https://github.com/AIDC-AI/Pixelle-Video
Developer Tools
InsForge — 8.8K stars (+459 today). New — the backend platform for coding agents. Postgres-based with auth, storage, compute, hosting, and AI gateway. Acts as a semantic layer between agents and backend primitives — agents can fetch docs, configure services, and inspect state. Think "Supabase, but designed for agents to operate." Apache-2.0. https://github.com/InsForge/InsForge
docusealco/docuseal — 15.5K stars (+899 today, +2.7K this week). Open-source DocuSign, quietly exploding. Ruby/Rails document signing platform — create, fill, sign. Self-hosted, AGPL. Nearly 3K stars gained this week with no marketing push visible. The self-hosted business-tool wave continues. https://github.com/docusealco/docuseal
1jehuang/jcode — 4.8K stars (+3.3K this week). New — the performance-obsessed agent harness. Rust-based coding agent focused on multi-session workflows, minimal RAM (fraction of alternatives), sub-second boot. Supports Claude, OpenAI, any LLM. Emphasizes "raise the skill ceiling" over "lower the barrier." https://github.com/1jehuang/jcode
rtk-ai/rtk — 43.8K stars (+847 today). The token diet pill keeps growing. Rust CLI proxy that filters command outputs before they hit your LLM — 60-90% token reduction. Now at 43.8K from 41.4K three days ago. A utility that saves real money. https://github.com/rtk-ai/rtk
Hidden Gems
Augani/openreel-video — 1.5K stars (+208 today). Browser-based video editor, zero installs. Open source CapCut alternative. 130K+ lines, WebGPU accelerated, multi-track timeline, keyframe animations, color grading, LUT support, ProRes export — all running client-side in Chrome. No uploads, no watermarks. Genuinely impressive feature depth for an OSS browser tool. https://github.com/Augani/openreel-video
always-further/nono — 2.2K stars. Agent sandboxing from the Sigstore creator. Capability-based sandboxes with fine-grained policies, built specifically for AI agents. Credential injection, agent multiplexing, snapshots, supply chain security. Profiles for Claude Code, Codex, OpenCode. The security layer agents need but don't have yet. https://github.com/always-further/nono
modelcontextprotocol/ext-apps — 2.2K stars. MCP Apps protocol — UIs inside AI chatbots. Official spec and SDK for embedding interactive app UIs served by MCP servers directly into AI chat interfaces. The next layer of MCP beyond tools and resources. https://github.com/modelcontextprotocol/ext-apps
Notable Trends
Agent skills repos are now a mega-category. mattpocock/skills (64K), forrestchang/andrej-karpathy-skills (113K), addyosmani/agent-skills (32K), ComposioHQ/awesome-codex-skills (7.4K). Four repos totaling 217K+ stars, all essentially collections of markdown behavioral files. The demand for "how to make agents not do stupid things" dwarfs almost every other category on GitHub right now.
Finance AI has become a full ecosystem, not a trend. TradingAgents (71K), Anthropic financial-services (11K), dexter (24K), PageIndex citing FinanceBench as its benchmark. When the benchmark, the agents, the data connectors, and the infrastructure all exist in the same trending list, it's an ecosystem. Enterprise buyers are clearly paying attention.