CastClaw（观星阁） is an open-source framework for agentic time series forecasting. It enables LLM agents to autonomously perform data analysis, model selection, feature engineering, and iterative forecasting refinement through structured tool-augmented workflows. CastClaw（观星阁） bridges classical time series methods with agentic reasoning, supporting diverse forecasting scenarios ranging from short-term point prediction to long-horizon multi-step forecasting, with built-in evaluation and interpretability mechanisms.

CastFactory（炼星坊） is an open-source framework for LLM-driven time series forecasting model training, designed to make it extremely easy for users to build and adapt TSF models with modern large-model pipelines. It provides a unified and practical workflow for continued pre-training (CPT), supervised fine-tuning (SFT), and reinforcement learning (RL), helping researchers and practitioners develop, optimize, and evaluate LLM-based forecasting systems with much lower engineering overhead.

FutureCast（天星台） is a unified evaluation suite for time series models and agentic forecasting systems. It organizes benchmark management, metric computation, error attribution, and report generation into a consistent evaluation workflow, helping researchers compare model capabilities and system-level gains across datasets, scenarios, and task settings including forecasting, classification, and anomaly detection.

NeoResearch（智多星） is an autonomous research agent system for time series forecasting model development. It connects research hypothesis generation, candidate recipe search, controlled experiments, evaluation diagnosis, and research memory into a reproducible loop, supporting systematic iteration from problem definition to validated forecasting model candidates.

Academic Search is an open-source academic literature research skill for Claude Code. It unifies paper discovery across arXiv, Semantic Scholar, Google Scholar, CNKI, and other scholarly sources, while supporting query expansion, citation tracing, BibTeX export, PDF-first retrieval, and multi-source deduplication. With recency-aware ranking and browser-assisted fallback for hard-to-access platforms, it helps researchers surface recent, high-value papers and their code resources more efficiently.

PaperScout is an autonomous academic paper search agent trained with Process-Aware Sequence-Level Policy Optimization (PSPO). It formulates literature search as a multi-turn sequential decision-making problem, allowing agents to dynamically decide when to Search for new directions and when to Expand along citation and related-paper paths. PaperScout turns scholarly retrieval into an agentic RL environment for learning efficient tool-use policies and producing stronger recall under realistic research queries.

TabClaw is an open-source agentic framework that empowers LLMs to reason over complex, real-world tabular data. It decomposes table-centric tasks into structured sub-goals, equips agents with code execution, schema-aware lookup, and formula tools, and coordinates them through multi-step decision workflows. TabClaw is designed to tackle challenges beyond flat QA — including multi-hop joins, conditional aggregation, and cross-table inference — making it suitable for enterprise data analysis and scientific table understanding.

Claw-R1 is an open-source framework for training reasoning-intensive, tool-using LLM agents via reinforcement learning. It extends standard RL environments with agentic action spaces — including tool invocation, multi-turn interaction, and environment feedback — and supports GRPO-based policy optimization with customizable reward signals. Designed for reproducibility and extensibility, Claw-R1 enables researchers to study how RL shapes emergent reasoning behaviors, tool-use strategies, and decision robustness in large language models.

Agent-R1 is a large-model agent training framework for end-to-end reinforcement learning fine-tuning. It supports GRPO-based policy optimization with reward signals derived from environment feedback, and equips agents with multi-tool orchestration, persistent long-term memory, and reflective self-correction across interaction turns. Grounded in the DeepSeek-R1 paradigm, Agent-R1 enables researchers to train LLM agents that not only reason under uncertainty but also learn to use tools strategically and recover from failures autonomously.