Today's most notable movement on GitHub reflects a clear pattern: agentic AI is reaching production-grade, domain-specific workflows. Rather than standalone chat interfaces, the trending projects this week are frameworks designed to slot into real business processes — financial modeling, long-horizon software development, time-series forecasting, and code intelligence. The tooling layer is maturing fast.
Anthropic Releases Open-Source Financial Services Agent Toolkit
Anthropic's financial-services repository crossed 31,900 stars today, making it one of the most-starred Anthropic repositories on GitHub. It provides eleven production-ready agent plugins for financial workflows — covering pitch deck generation, earnings call review, KYC document parsing, GL reconciliation, and more — deployable either as Claude Cowork plugins or through the Managed Agents API. What makes it stand out is its dual-path architecture: the same source files serve both as no-code plugins and as headless, API-driven agents. Firm-specific customization happens through YAML and Markdown rather than code changes, lowering the bar for adoption at organizations with limited AI engineering capacity.
github.com/anthropics/financial-services
GLM-5 Targets Long-Horizon Agentic Tasks with a 1M-Token Context Window
Zhipu AI's GLM-5.2 — the latest in the GLM-5 family — is built explicitly for software development tasks that unfold over many steps: multi-file repository generation, extended debugging sessions, and iterative improvement loops that standard context windows cut short. The flagship model scales to 744 billion parameters with 40 billion active at inference, and ships with a stable one-million-token context window. Scores of 62.1 on SWE-Bench Pro and 81.0 on Terminal-Bench 2.1 suggest the claims about real-world coding tasks have some grounding beyond leaderboard positioning. Deployment is supported via SGLang, vLLM, and Transformers.
TimesFM: Google's Time-Series Foundation Model Climbs Back Into Trending
Google Research's TimesFM crossed 23,700 GitHub stars as it returned to the daily trending list. Unlike general-purpose LLMs fine-tuned for forecasting, TimesFM is pretrained specifically on time-series data and designed for zero-shot forecasting across domains — demand planning, energy load, financial indicators, and sensor readings. The appeal is practical: getting accurate forecasts without labeled training data for your specific domain is the norm in production, and a strong zero-shot baseline changes the calculus on when custom training is actually worth the investment.
github.com/google-research/timesfm
DeusData/codebase-memory-mcp: A Knowledge Graph for Your Repository
Codebase Memory MCP indexes a repository into a knowledge graph and exposes it as a Model Context Protocol server, returning query results in under a millisecond. Rather than relying on embedding-based similarity search, it builds structural relationships between symbols, files, and call chains — which matters for questions like "what calls this function" or "what does this module depend on" that dense retrieval handles poorly. The MCP interface means any Claude environment, IDE extension, or agent with MCP support can plug it in without custom integration work.
github.com/DeusData/codebase-memory-mcp
Lightricks/LTX-2: Open-Weight Audio-Video Generation with Fine-Tuning
LTX-2 from Lightricks is an open-weight model for generating synchronized audio and video from a prompt, shipping with an official Python package that covers both inference and LoRA fine-tuning. With 7,500 GitHub stars, it lands in a space long dominated by closed APIs. The combination of open weights and a fine-tuning path makes it more practical for production use cases where default outputs need to match a specific visual style or brand identity — something closed video generation APIs have made difficult.