Open Source Releases

Gread — MCP server giving AI agents access to all public GitHub repos — An MCP server that lets your coding agent browse any public GitHub repo like it’s local. Auto-detects docs repos, works with Claude Code, Cursor, Copilot, etc. Finally, “read the source” without leaving your editor. 🤖

Claude Code v2.1.169 – Safe mode flag and /cd command — --safe-mode starts Claude Code with zero customizations — handy for debugging config conflicts. New /cd command switches working directories mid-session without nuking your prompt cache. Small but meaningful QoL. 🛠️

tenax-tn 0.7.1.dev20260609 — JAX tensor network library with symmetry-aware block-sparse tensors. Built for quantum many-body physics sims. If you’re doing scientific ML with tensor networks, this is your jam; otherwise, hard pass. 📄

Research Worth Reading

PathoSage: Towards Multi-Source Evidence Adjudication in Pathology via Experience-Aware Agentic Workflow — Agentic workflow for computational pathology that adjudicates evidence across multiple sources. Tackles the hallucination problem in patch-level morphological reasoning. Medical AI folks: this is how you do RAG right for high-stakes domains. 📄

OmniMem: Perturbation-aware Memory Compression for Streaming Audio-Visual LLMs — Compresses KV caches for long-form audio-visual understanding by being smart about what to perturb. Memory-efficient streaming for video LLMs. If you’re building long-context multimodal systems, the compression strategy here is worth stealing. 📄

Syll: Open-Source Personal Automation with Cross-Surface Execution — Self-hosted personal agent that operates across APIs, shells, web, and desktop GUIs. Built for teachability and auditability — you can actually see and correct what it does. Refreshing to see “open source” and “auditable” in the same sentence for agent frameworks. 🛠️

A case study of evaluating AI agents on a neuroscience data-to-discovery pipeline — Real-world eval of agents on a neuroscience pipeline that takes domain experts months. Spoiler: agents struggle with correctness and robustness in scientific code. Good reality check if you’re selling “AI replaces scientists” snake oil. 📄

Why Limit the Residual Stream to Layers and Not Tokens? Persistent Memory for Continuous Latent Reasoning — Extends Chain of Continuous Thought (CoCoNuT) with persistent memory across tokens, not just layers. Latent reasoning without decoding at every step. Theoretical but the persistent memory angle could matter for long-horizon reasoning architectures. 📄

Automatic Extraction of Structured Information from Brain MRI Reports Using an Open-Weight Large Language Model — Tests open-weight LLMs on Dutch neuroradiology reports (947 samples). Structured extraction from messy clinical text. Practical medical NLP — shows what’s actually achievable with open models on non-English clinical data. 📄

AI Dev Tools

Egonex-AI/Understand-Anything — Interactive Code Knowledge Graph Explorer — Drops any codebase into an interactive knowledge graph. Queryable, searchable, works as a skill for Claude Code, Codex, Cursor, Copilot, Gemini CLI. Finally, “how does this connect to that?” without grepping for hours. 🛠️

Today’s Synthesis

The combination of Gread , Understand-Anything , and Claude Code’s new /cd command gives you a genuinely practical “read any codebase” workflow today. Point Gread at a public repo — no clone needed — and your agent can browse it like local files. Feed that context into Understand-Anything to auto-generate a queryable knowledge graph of call chains, dependencies, and architectural patterns. Then use /cd to hop between your actual project and the referenced code without losing prompt cache, letting you trace “how does this library actually implement that interface?” in a single session. This isn’t theoretical: install the MCP server, add the skill to Claude Code, and you’ve replaced hours of git clone → grep → less → git clone another repo with a conversation. The --safe-mode flag is the cherry on top — spin up a clean agent instance when you need unbiased answers about unfamiliar code, not your customized shortcuts. 🛠️🤖