Tools

GitHub's official spec-driven development toolkit: the specify CLI sets up spec, plan and task workflows for Copilot, Claude Code, Codex and other agents so the spec comes before the code, plus bug-fix and idea-evaluation flows. Needs Python 3.11+ and uv.

Skill pack from YC president Garry Tan: slash commands such as /office-hours, /plan-ceo-review, /review, /qa and /ship play CEO, engineering manager, designer, QA and release engineer. Needs Git and Bun; clone the repo and run ./setup as the README describes.

YC office hours skill from gstack: startup mode tests real demand, the narrowest wedge and future fit with six probing questions; builder mode brainstorms side projects, hackathons and open source. Ends with a design doc. Maintained by Garry Tan.

Local LLM inference engine in plain C/C++ with a command-line chat and an OpenAI-compatible HTTP server; pulls and runs GGUF models straight from Hugging Face. Supports Apple Silicon, CUDA, Vulkan and other hardware, plus low-bit quantization.

Minimal, extensible terminal AI agent that can write its own prompt templates, skills, extensions and themes. Offers print, JSON and RPC modes for scripting, plus a TypeScript SDK. Needs Node.js 22.19+.

Multi-agent investment research framework where fundamental, sentiment, news and technical analysts, bull and bear researchers, a trader and risk managers debate a stock and produce a report and rating. Covers US, Hong Kong and A-shares. Needs Python and an LLM API key; not investment advice.

OpenAI's open-source general-purpose speech recognition model. The whisper command transcribes many languages, translates foreign speech into English and writes subtitle files. Needs FFmpeg; faster with a GPU.

For fuzzy requirements: the agent asks one question at a time with its own guess, clarifying users, motivation, success criteria and constraints, and restates the intent for your confirmation before planning. Needs you answering live, so not for unattended runs. From Addy Osmani's agent-skills.

Open-source local design workbench and Claude Design alternative: builds web and mobile prototypes, slides, images and motion videos from a DESIGN.md system, exporting HTML, PDF, PPTX or MP4. MCP lets Claude Code and Codex use it. Needs the macOS or Windows app; Linux from source.

High-throughput LLM inference and serving engine: vllm serve starts an OpenAI- and Anthropic-compatible server in one command. Supports 200+ model architectures from Hugging Face, quantization and distributed inference. Aimed mainly at NVIDIA, AMD and other GPUs.

Provides a sequential_thinking tool that guides the model to break complex problems into revisable, branching steps. Good for planning, analysis and tasks that need careful iteration. Official MCP reference server.

Baidu PaddlePaddle's official OCR and document parsing tool: the paddleocr command reads text from images and parses PDFs and scans into Markdown or JSON, supporting 100+ languages with models that run locally.

Official PaddleOCR MCP: text recognition and layout parsing for images and PDFs (PP-OCR, PP-StructureV3, PaddleOCR-VL), returning text or Markdown. Runs inference locally, or calls the official API or a self-hosted service.

Astral's extremely fast Python package and project manager written in Rust, replacing pip, pipx, poetry, pyenv and more. Supports lockfiles and multiple Python versions, and uvx runs Python tools without installing them.

MCP server built into the adaptive Python scraping framework Scrapling: supports plain HTTP requests, browser rendering and stealth fetching, and uses CSS selectors to pass only the target content to the AI, saving tokens and filtering out prompt injection.

Adaptive web scraping framework: scrapling extract turns a page into Markdown, text or HTML in one command, with CSS selectors and browser rendering for dynamic pages, plus an MCP server (needs the ai extra). Needs Python 3.10+ and a browser install on first use.

Analyzes a codebase with a multi-agent pipeline to build a knowledge graph of files, functions, classes and dependencies plus an interactive dashboard for architecture tours, semantic search and change impact analysis. Also handles knowledge bases; the first run uses many tokens.

Open-source, LLM-friendly web crawler that turns pages into clean Markdown for RAG and agents. Ships the crwl CLI and supports deep crawling and question-driven data extraction (extraction needs an LLM key). Free to self-host.

Turns natural-language descriptions, Mermaid or a real code repo into interactive single-file HTML diagrams: architecture, flowcharts, sequence diagrams, data flows and state machines. Validates output automatically and exports PNG or SVG.

Document parser from OpenDataLab that turns PDFs, images and Office files into Markdown or JSON, handling tables, formulas and complex layouts well, with page-by-page reading and citation anchors for agents. Parses locally by default; needs Python 3.10+.

Classic open-source OCR engine and CLI that reads text from images and outputs plain text, hOCR, searchable PDF or TSV. Supports 100+ languages; install the language data for each language you need.

All-in-one LLM fine-tuning framework: llamafactory-cli handles pretraining, SFT, PPO/DPO alignment and LoRA/QLoRA training, plus chat inference, weight merging and export, and OpenAI-style API serving for 100+ open models. Has a Web UI; training usually needs a GPU.

Multi-agent orchestration framework for Claude Code and Codex, formerly Claude Flow: swarm coordination, self-learning memory and cross-machine federated communication. The marketplace offers 30+ plugins; full features need npx ruflo init.

Multi-agent orchestration framework formerly called Claude Flow: ruflo init gives Claude Code nearly 100 specialized subagents, swarm coordination, cross-session vector memory and hooks; add --codex to configure Codex instead. Also works as a standalone MCP server or plugin marketplace.