도구

Open-source LLM evaluation platform: one opencompass command picks the model and datasets and runs the evaluation across hundreds of benchmarks. Works with local Hugging Face models and API models, with vLLM and LMDeploy for faster inference.

Converts and two-way syncs Jupyter notebooks with plain Python scripts or Markdown, so agents can edit research and teaching notebooks like ordinary text and diffs stay readable. Needs a local Python environment; install with pip.

Cleans a LaTeX project before arXiv submission: removes comments, auxiliary files and unused .tex files and images, and can compress images, so internal notes are not published with the source. From Google Research; installable with Homebrew on macOS.

Python library and CLI for extracting main web content: fetches pages, crawls via sitemaps and RSS feeds, strips headers, footers and other noise, and outputs TXT, Markdown, JSON, XML and more with metadata. Also processes downloaded HTML offline.

Anthropic's official OS-level sandbox CLI: restricts file and network access for agents, MCP servers and any command without containers, using sandbox-exec on macOS and bubblewrap on Linux. Currently a research preview.

Google's official skills for building ADK agents: dev workflow, scaffolding, the ADK Python API, evaluation with LLM judges and adaptive rubrics, deployment to Agent Runtime, Cloud Run or GKE, publishing and observability. Works with agents-cli; local dev needs only an AI Studio key.

Semi-automated research assistant for CS and AI researchers that links ideation, Zotero literature management, experiment analysis and reporting, paper writing and reviewer responses, and keeps project knowledge in Obsidian. Branches for Codex and other agents exist.

Semi-automated research workflow for computer science and AI researchers: ideation, Zotero literature management, experiment analysis and reports, paper writing and polishing, and rebuttals, with optional Obsidian knowledge base. 40+ skills; also installable as a Claude Code plugin.

Detects shot changes in video and splits long footage into scenes, exporting cut lists and per-scene thumbnails, useful for shot analysis, organizing footage and finding edit points. Runs locally; needs Python 3.10+, plus FFmpeg or mkvmerge for splitting.

LM Studio's official CLI: list, load and unload local models, start and stop the local API server, and check status and logs; ls, ps and other commands support --json output for scripts and agents. Requires the LM Studio desktop app.

CLI and TypeScript runtime for discovering and calling MCP servers from the terminal and scripts: list tool signatures and call tools directly, or generate a standalone CLI or typed client from a server. Auto-imports existing MCP configs from Claude Code, Codex and others.

Teaches agents to code against the latest Gemini API and Python/TypeScript SDKs: multi-turn chat, function calling, structured output, and image and speech generation, while avoiding deprecated older models. Maintained by Google's Gemini team.

Gemini API development skills: gemini-api-dev (core API and SDKs), gemini-live-api-dev (real-time bidirectional audio/video streaming) and gemini-omni-flash-api (Omni video generation and editing), filling in APIs released after model training cutoffs. Maintained by Google's Gemini team.

Hugging Face's official hf CLI: search, download and upload models and datasets, manage repos and storage buckets, and run jobs on Hugging Face's cloud. Output adapts automatically for agents, and hf skills add installs companion skills. Uploads and similar actions need a login.

NVIDIA's official plugin that finds and installs the right skills from NVIDIA's catalog for your task, spanning CUDA acceleration, NIM inference, NeMo training, RAG blueprints, Omniverse simulation and robotics. Listed in both the Claude and Codex official marketplaces.

NVIDIA's official, verified catalog of hundreds of skills covering CUDA-X libraries, NeMo training, NIM and Dynamo inference, RAG blueprints, Jetson, Omniverse and robotics simulation. The install adds only nvidia-skill-finder, which recommends other skills and installs them once you confirm.

AWS's generally available agent toolkit and plugin marketplace: plugins such as aws-core, aws-agents and aws-data-analytics bundle the AWS MCP Server with evaluated skills for IaC, serverless, containers and billing. Calling AWS APIs requires configured credentials.

About 40 research skills from Google DeepMind: search arXiv, PubMed and OpenAlex, query UniProt, PDB, ClinVar, ChEMBL and more, run AlphaGenome variant analysis and render structures in PyMOL. Requires uv; AlphaGenome needs an API key, and keys raise limits for OpenAlex, ClinVar and others.

MCP server from the opengeos GeoAI project: segment buildings, roads and vegetation in satellite imagery from text prompts, detect vehicles and ships, spot changes between images, estimate canopy height and fetch NAIP, Sentinel-2 and Landsat imagery. Install from the repo (Python 3.10+).

Teaches agents to build, debug and review stateful AI agents with the Cloudflare Agents SDK: state sync, scheduled tasks, RPC, MCP servers, email and streaming chat, checking the latest official docs first. Maintained by Cloudflare.

Search and download arXiv papers and read the authors' LaTeX source section by section, export BibTeX and follow topic updates, with papers cached locally. No API key needed.

Microsoft's official skill collection: 170+ skills focused on Azure SDKs (Python, .NET, TypeScript, Java, Rust) and Microsoft Foundry development, plus plugins, custom agents and MCP configs. Microsoft recommends picking skills per project rather than installing them all.

869 medical skills compiled by FreedomIntelligence from a dozen open-source libraries: SOAP notes, discharge summaries, drug interactions, clinical guidelines, PubMed/FDA lookups, genomics and device regulation. Made for OpenClaw, usable elsewhere; some need API keys. Doctors must review output.

Unified search across 20+ academic sources including arXiv, PubMed, bioRxiv, Semantic Scholar and OpenAlex, returning deduplicated, normalized results and downloading full text via open-access channels first. Most sources need no API key.