Open Source Releases

arxiv-mcp-server — Search and read arXiv papers via MCP — MCP server that lets agents search arXiv, fetch metadata, and read full-text content. Clean integration for research-assistant workflows that actually need to cite sources instead of hallucinating them.

sst/opencode v1.17.7 — MCP workspace passing, plugin client reuse, and shell tool UX fixes — MCP servers now receive the current workspace as context. Plugin clients reuse the active server instead of assuming a default port. ACP shell tool calls show the command and working directory upfront, and plugin-provided env vars apply to PTY sessions. Quality-of-life fixes that matter when you’re knee-deep in agent loops.

lilbee 0.6.66b497 — Batteries-included terminal app for local AI: browsable model catalog, semantic/hybrid search over local files and code, vision OCR, auto-built wiki, and source-citing chat. Ships as CLI, TUI, MCP server, REST API, and Python library in a single process with no external model or database dependencies. Impressive density.

Kronos: A Foundation Model for the Language of Financial Markets — Foundation model built specifically for financial market language understanding and prediction. Targets financial NLP tasks like sentiment analysis and forecasting. Domain-specialized models keep eating generalists’ lunch in narrow verticals.

gsd-build/get-shit-done - Meta-Prompting & Spec-Driven Development System — Lightweight meta-prompting and spec-driven development system for Claude Code. Focuses on structured workflows with context engineering patterns. The name delivers what it promises.

Research Worth Reading

Efficient On-Device Diffusion LLM Inference with Mobile NPU — Tackles running diffusion-based LLMs on smartphones by leveraging mobile NPUs for the dense matrix work during repeated denoising steps. Makes parallel token denoising practical for latency-sensitive on-device use. Your phone might actually run this soon.

Can Editing 1 Neuron Fix Repetition Loops in LLMs? — Finds a reproducible failure mode in Gemma 4 instruction-tuned models where long factual enumeration prompts collapse into repetition loops. Tests whether targeted single-neuron editing breaks these loops. Mechanistic interpretability with a scalpel instead of a sledgehammer.

The Weight Norm Sets the Grokking Timescale: A Causal Delay Law — Settles a grokking literature dispute by causally intervening on weight norm during training (not just observing). Demonstrates weight norm directly controls the delayed generalization timescale. A causal law linking norm dynamics to the grokking transition. Finally, someone did the intervention experiment.

Diffusion Policy Optimization without Drifting Apart — Identifies “double-drift” in diffusion policy RL post-training: optimizing a variational surrogate causes the ELBO to diverge from true log-likelihood, destabilizing policy gradients. Proposes a corrected method for reliable policy improvement. The math checks out.

Beyond LoRA: Is Sparsity-Induced Adaptation Better? — Compares LoRA variants against sparsity-induced parameter-efficient fine-tuning methods. Examines how structural restrictions on updates affect generalization. Historical framing plus empirical analysis of when sparsity beats low-rank. LoRA isn’t the only game in town.

FedSPC: Shared Parameter Correction for Personalized Federated Learning — Addresses interference when splitting models into shared and personalized parameters in federated learning. Proposes Shared Parameter Correction (SPC) to decouple updates and improve convergence under statistical heterogeneity. FL still has sharp edges; this files one down.

AI Dev Tools

safishamsi/graphify - Code-to-Knowledge-Graph AI Skill — Turns any folder of code, SQL schemas, R scripts, shell scripts, docs, papers, images, or videos into a queryable knowledge graph using GraphRAG with Leiden clustering. Works with Claude Code, Codex, OpenCode, Cursor, Gemini CLI. Your codebase as a graph — finally queryable.

pubmed-mcp-server — MCP server for NCBI E-utilities / PubMed — Wraps NCBI E-utilities API for search, metadata/full-text retrieval, citation generation, MeSH term exploration, and related-research discovery. Useful for biomedical research agent pipelines. PubMed via MCP — someone had to do it.

wshobson/agents - Multi-Harness Agentic Plugin Marketplace — Plugin marketplace supporting Claude Code, Codex CLI, Cursor, OpenCode, GitHub Copilot, and Gemini CLI. Includes MCP integration, prompt engineering patterns, and workflow orchestration. One marketplace to rule them all — or at least try to.

Today’s Synthesis

MCP is quietly becoming the connective tissue for agent tooling. The arxiv-mcp-server and pubmed-mcp-server prove domain-specific knowledge bases can expose clean MCP interfaces — no custom integrations per agent framework. Meanwhile sst/opencode v1.17.7 ships workspace passing so MCP servers actually know what repo you’re in, and wshobson/agents demonstrates cross-harness plugin compatibility across Claude Code, Codex, Cursor, OpenCode, and Gemini CLI. The actionable takeaway: if you’re building a research assistant or code-aware agent, stop writing one-off API wrappers. Expose your tools as MCP servers (or consume existing ones) and they’ll work in any compatible client. Pair arxiv + pubmed + graphify for your codebase, and you’ve got a multi-source research agent that runs in whatever editor you prefer tomorrow.