The awesome-infra-for-ai list received 19 new repositories. Ontos-AI/knowhere (2,457 stars) — An ingestion pipeline that turns messy unstructured documents into persistent, navigable memory for AI agents — parsing, hierarchy extraction, multimodal structuring and graph construction in one pass. waybarrios/vllm-mlx (1,512 stars) — A vLLM-style inference server for Apple Silicon built on MLX: continuous batching, paged and prefix KV cache, and both OpenAI and Anthropic APIs from a single Metal-backed process. kenryu42/cc-safety-net (1,491 stars) — A PreToolUse hook that blocks destructive commands and secret access before AI coding agents run them, parsing command semantics so shell wrappers and flag reordering cannot bypass it. tensorlakeai/tensorlake (988 stars) — A sandbox-native compute platform for AI agents: stateful Firecracker MicroVMs with snapshots, cloning and live migration, plus a serverless function runtime for long-running orchestration. sgl-project/sglang-omni (831 stars) — A multi-stage serving runtime from the SGLang project for omni, speech and text-to-speech models, exposing OpenAI-compatible audio and chat endpoints with streaming output. Jia-Ethan/claude-keysmith (637 stars) — A preview-first deployment tool that installs, verifies and revokes custom instruction files for Claude Code across project, local and user scopes without hand-editing CLAUDE.md. CodeAbra/iai-personal-memory-engine (623 stars) — A fully local memory engine for AI coding assistants, exposed over MCP: verbatim recall with encrypted local storage, benchmarked retrieval quality and injected memory packs that cut search tokens. sandbaseai/sandbase-harness (611 stars) — A local-first runtime for AI agents providing persistent sessions, sandboxed tool execution, credential vaults, memory, audit trails and a built-in console, all running on your own infrastructure. caura-ai/caura (430 stars) — Shared, governed memory for fleets of AI agents: agents write plain text, Caura turns it into searchable multi-tenant memory with scoping, trust tiers and cross-agent outcome propagation. vivekchand/clawmetry (395 stars) — Real-time local observability dashboard for AI agent runtimes: auto-detects installed coding agents and meters their sessions, tools, models, providers and token usage in one view. ukanwat/aaabench (363 stars) — An open-ended benchmark harness that hands a coding agent a live Unreal Engine 5 editor over MCP and measures whether it can build a whole open-world game unaided.