The AI agent landscape is finally getting the census it deserves. Favz, a new indexing tool, has scanned 166,014 public GitHub repositories to determine exactly which tools developers are configuring for their AI agents. Unlike star-count metrics that favor popularity, Favz reads the actual config files checked into repos—counting MCP servers, skills, and plugins to reveal what is set up, not just what is talked about. The collector began operations on September 22, 2026, with a full initial pass over GitHub completed on October 7, 2026.

The Web Frontend Stack Dominates

The data exposes a clear winner in the Model Context Protocol (MCP) ecosystem: web development tools. Playwright (@playwright/mcp) leads with 4,594 repos currently configured, followed closely by Context7 (4,335) and Supabase (4,351). A deeper analysis of a random sample of 2,039 repos shows that four of the top ten MCP servers—Playwright, shadcn, Chrome DevTools, and Next.js DevTools—are exclusively for building web frontends. The most common pairing is Playwright with Context7, found in 33 of the sampled repos, suggesting that browser control combined with current documentation access is becoming the de facto standard stack for public AI agent setups.

Hosted Servers Are Replacing Local Packages

A significant architectural shift is underway in how developers launch these tools. In 2025 Q3, only 13% of newly added MCP servers were hosted URLs, with 63% being local npm packages. By 2026 Q3, hosted URLs have surged to 36% of new adoptions, while npm packages have dropped to 40%. This trend indicates a move away from local dependency hell toward managed, remote endpoints. The data also highlights the fading relevance of early reference servers; sequential-thinking, once a staple, saw new adoptions plummet from 17 in 2026 Q1 to just 2 in 2026 Q3, signaling that the experimental phase of agent tooling is ending.

The Long Tail of Custom Configuration

While the headline numbers highlight popular tools, the reality of agent configuration is fragmented. Favz identified 24,207 distinct tool names across the dataset, but 76% of them (18,496 names) appear in exactly one repository. This massive long tail suggests that most developers are not adopting standardized products but are instead writing custom hooks, skills, and agents for their specific workflows. Furthermore, the study found that skills often travel in packs; the top skill pairs are members of the same install bundle, meaning a single skill’s adoption count is often just a proxy for its parent pack’s popularity.

Key Takeaways

  • Playwright and Context7 are the most adopted MCP servers, with over 4,300 repos each currently configured.
  • Hosted MCP servers now account for 36% of new adoptions, nearly tripling their share from a year ago.
  • 76% of all tool names appear in only one repository, highlighting a heavy reliance on custom, non-standardized configurations.
  • Early reference servers like sequential-thinking are losing ground, with new adoptions dropping to near-zero in Q3 2026.

The Bottom Line

The era of 'demo-ware' agents is over; developers are settling on a pragmatic, web-centric stack with a clear preference for hosted infrastructure over local complexity.