Today's take
The strongest signal is a shift from configuring agent capabilities to supporting sustained execution. cayu-dev/cayu enters the Top 50 with a durable runtime, while yetone/magpie climbs from #48 to #10 by centralizing model configuration across agents: the surrounding execution stack is attracting attention alongside skills.
Usage points in a similar direction, with personal and coding agents leading the app rankings and batch-labeled models topping model usage. The dates differ: repository rankings are for October 9, while model and app usage covers October 7. The latter captures only traffic routed through OpenRouter with public attribution; tokens measure usage, not people.
Open source
Today's Top 50 includes 1 new entrant and 2 repositories flagged as rising. Three themes matter more than the count: agent engineering, separation of tool interfaces from model access, and reusable creative workflows. Stars indicate attention rather than demonstrated adoption.
- Agent engineering: #2 mattpocock/skills adds 2,095 stars in 1 day, packaging engineering practices as reusable skills. #26 cayu-dev/cayu adds 542 stars to reach 692 total and supplies a domain-specific, long-horizon agent framework with a built-in durable runtime. They address complementary layers: how an agent works and how its work keeps running. Cayu is the new entrant to track most closely.
- Model access becomes its own layer: #10 yetone/magpie adds 893 stars, reaches 6,540 total and jumps from #48 to #10. Its menu-bar model configuration suggests growing interest in separating agent interfaces from underlying model choices. #5 Niko1221/Strata approaches access friction through local inference and compatible APIs.
- Creative workflows become reusable: #3 storytold/artcraft adds 1,964 stars to reach 5,749 total, targeting artists, designers and filmmakers. #31 yihui-dev/awesome-opus5-5-videos adds 474 stars and climbs from #42 to #31 by collecting videos and their prompts. Attention extends beyond individual outputs to reusable ways of producing them.
Models
Batch-labeled variants dominate the leading positions in the October 7 snapshot, but growth differs considerably across models. These figures do not identify which apps drove a model's growth.
- DeepSeek V4.1 Flash (batch): #1 with 7.35T tokens, up 50% versus the previous day. DeepSeek V4 Flash 0731 records 693B and -15%, while DeepSeek V4 Flash 0423 records 442B and -6%. That is divergence across variants, not sufficient evidence of migration between them.
- Other leading models: GLM 5.3 Flash (batch) processes 1.67T tokens, up 3%; Claude Opus 5.5 (batch) reaches 789B, up 43%, showing substantially faster growth.
- Movers outside the top ten: Among the Top 50 models outside the top ten, Grok 4.7 leads the supplied movers at 105B tokens and +53%. GPT-6.1 Sol combines a larger 294B total with +37% growth.
Apps
The app rankings support sustained execution and development workflows as important sources of token usage. Three of the top five apps are classified as CLI agents and one as an IDE extension. Without app-to-model allocation data, their model choices cannot be inferred.
- Usage leaders: Hermes Agent ranks first at 1.49T tokens, followed by Claude Code at 1.31T. Personal and coding agents both occupy leading positions.
- Development workflows: Kilo Code records 592B tokens, pi 499B and Cline 475B, showing substantial coding-agent usage across several interfaces.
- Rising apps: Against the average of the preceding three equal-length periods, codex-fleet conductor ranks first on the rising list with 260B current tokens; Claude Code and pi follow. Growth magnitudes were not supplied, so percentages cannot be stated.
What to watch
Watch whether attention to reusable skills and durable runtimes persists alongside growth in agent app usage: the evidence supports a shared direction, but does not establish a direct causal link.
