AI repository comparison: Nativ, QM and codebase memory

This overview compares ten open source repositories around model inference, agent workflows and code review. The speaker focuses on what each tool does and whether its headline claims have supporting tests.

For local LLM use, Nativ is described as a Swift app for macOS 14 or later with an MLX VLM server, cached Hugging Face model discovery and OpenAI- and Anthropic-compatible endpoints. turbo-fieldfare takes a narrower approach: it streams Gemma 4 26B experts from SSD. The speaker reports roughly 2GB of RAM use, 14.3GB on disk and 5 to 6 tokens per second on an 8GB M2 MacBook Air. Its stated limits are text only, Apple silicon and Metal 4.

RTK compresses shell output for a coding assistant. The speaker cites a JetBrains test in which average task cost rose 7.6%, distinguishing shorter output from lower bills. codebase-memory-mcp instead indexes repository structure into a graph; the cited study found fewer tokens and tool calls but lower response quality.

The workflow comparison covers qm's self-hosted shared workspaces and OpenWorker's permission modes. OpenWorker is described as an open beta with unsigned Windows builds. The speaker also reports repository uploads by Grok Build CLI 0.2.93. Codex Security reviews code through an editable threat model, while T3MP3ST coordinates authorized security testing and labels much of its functionality experimental. no-ai-slop supplies editing rules for model-written prose.