The Inference Report

The GitHub trending set shows developers investing heavily in agent infrastructure, not the agents themselves, but the connectors, skills, and interfaces that let AI systems interact with real tools. Chrome DevTools MCP, Unity MCP, and page-agent all solve the same underlying problem: giving language models actionable control over systems they couldn't previously touch. That's a genuine shift in what developers are building for. Contrast this with the token-optimization jokes (caveman Claude, token-counting skills) that are viral because they're funny, not because they solve production problems. The distinction matters. Real traction is going to tools that expand what agents can do, browser automation, game engine integration, terminal multiplexing, while the viral repos are mostly commentary on how the underlying models work.

The secondary pattern is self-hosted infrastructure for privacy-sensitive workloads. Immich dominates photo management, Meetily brings meeting transcription fully local with Rust and Ollama, and MatrixOne positions itself as a vector-aware database for agent memory. These aren't novel conceptually, but they're gaining traction because the cloud-first default is breaking down for teams handling sensitive data or operating under latency constraints. Skills repositories (mattpocock's engineering skills, the 337-skill Claude collection, .NET skills for agents) are also trending, though they're closer to documentation than code, useful for standardizing how agents interact with domain-specific systems, but they don't solve problems so much as catalog solutions. The real engineering work is happening in the MCP servers and agent multiplexers like herdr, which actually make it possible to chain multiple tools and models together in a terminal without reinventing orchestration each time.

Jack Ridley