The HashFork List — curated AI resource list

The Agency is a collection of AI specialist agents you can drop into coding and automation workflows, each a domain expert with its own voice, processes and measurable deliverables, from frontend work to community management. The Agency Agents desktop app installs them into Claude Code, Cursor, Codex, Gemini CLI and OpenCode and keeps them updated. MIT licensed.

Dify is a collaborative workspace for building agentic workflows and RAG pipelines with broad AI model and tool support, combining visual building with the tooling to develop and ship on one platform. It is free under a modified Apache-2.0 license, with a commercial license required for multi-tenant services.

Transformers is the model-definition framework from Hugging Face for state-of-the-art machine learning models in text, vision and audio, giving developers the architectures and pretrained models to use and build on. Apache-2.0 licensed.

Ollama gets you up and running with open models such as Kimi, GLM, MiniMax, DeepSeek, gpt-oss, Qwen and Gemma on your own machine, with one-line installs on macOS, Windows, Linux and Docker. MIT licensed.

AutoGPT is a platform for building, deploying and running AI agents that carry out complete workflows: describe an outcome in plain English or shape every step in the visual builder, then run the agent on demand, on a schedule or from a trigger. The platform code is free under the Polyform Shield license except for competing services, with the rest of the repository under MIT.

Firecrawl is a web data layer for AI agents: it crawls and scrapes sites and returns clean, LLM-ready markdown, structured JSON or screenshots, with proxy rotation, JavaScript rendering and PDF and DOCX parsing handled. It connects to any AI agent or MCP client in one command and runs as a hosted service or self-hosted, free under the AGPL-3.0 license.

TensorFlow is an open-source machine learning framework for everyone, used to build and train machine learning models and put them into production. The repository contains the core libraries and tools of the framework and is released under the Apache-2.0 license.

DeepSeek Harness (dsh) is an open-source agent harness from DeepSeek AI in which everything is a plugin, so models, tools and behaviours are swappable components rather than fixed code. It is built on the Cordis plugin architecture and currently in developer preview with frequent compatibility-breaking changes. MIT licensed.

Hermes Agent is an open-source personal AI agent from Nous Research that works with any model provider and meets you on Telegram, Discord, Slack, WhatsApp, Signal or a full terminal interface through one gateway. It keeps agent-curated memory, creates its own skills after complex tasks, runs scheduled automations and parallel subagents, and executes on backends from local and Docker to SSH, Modal and Vercel Sandbox.

OpenClaw is an open-source personal AI assistant that runs on your devices and works inside the chats you already use, doing real tasks instead of only answering. State, memory and credentials stay on your hardware, models and agent harnesses are swappable plugins, and the project is stewarded by the nonprofit OpenClaw Foundation with no paid tier. MIT licensed.

LivePortrait is a portrait animation model from Kuaishou that brings a still photo of a face to life by driving it with a reference video or audio, with driving-video auto-cropping, privacy templates and a separate animals model, under stitching and retargeting control for precise face motion. The code is MIT licensed.

Hypit is a video-cloning agent system: drop in a viral video and its agent rebuilds the complete workflow behind it, covering footage, captions, B-roll and effects, then renders many variants from one command with optional AI generation models. Free for your own organization's purposes under a modified Apache-2.0 license.

Pixelle-Video is a fully automated short-video engine: enter a topic and its pipeline writes the script, plans images, generates video with models like WAN 2.1, adds TTS voiceover and background music, and composes the finished clip in portrait or landscape. Workflows combine ComfyUI and RunningHub with direct model APIs from providers like DashScope and Kling. Free under the Apache-2.0 license.

Open Generative AI is a free image and video generation studio that runs hundreds of AI models across 14 studios from one interface, positioned as an open alternative to AI video platforms with no subscription fees and no closed ecosystem. The MIT-licensed web app calls hosted generation APIs such as MuAPI.

OpenMontage is an agentic video production system in which AI agents run twelve end-to-end production pipelines built from over 100 tools and 700 agent skills, covering research, scriptwriting and narration, image and motion-clip generation, music, word-level captions, timeline editing and final rendering through pluggable providers. Free under the AGPL-3.0 license.

GFPGAN is a face restoration model from Tencent ARC that develops practical algorithms for real-world face restoration, recovering old, blurry or low-quality faces in photos. The repository provides the code, pretrained models and a simple inference pipeline, free under the Apache-2.0 license.

Fooocus is an offline, free image generator built on Gradio that needs no manual tweaking: you focus on prompts and pictures, in the spirit of online tools like Midjourney. It covers text-to-image, img2img, upscaling, variations and its own inpainting and outpainting, installs in a few clicks and runs on 4 GB of NVIDIA GPU memory.

Stable Diffusion is a latent text-to-image diffusion model that creates images from text prompts and edits existing ones with img2img, released by the CompVis group with Stability AI and Runway, building on the High-Resolution Image Synthesis with Latent Diffusion Models paper. The repository ships the model code with training and sampling scripts, and the weights are published under the CreativeML Open RAIL-M license as free to use.

ComfyUI is a modular engine for AI content creation: image and video generation pipelines are built as graphs of nodes covering model loading, conditioning, sampling, upscaling and output, then run from a node editor, an API or a Python backend. It executes diffusion models locally and is extended with custom nodes. Free under the GPL-3.0 license.

Stable Diffusion web UI is a Gradio-based browser interface for Stable Diffusion that runs on your own hardware, with txt2img and img2img generation, outpainting, inpainting, color sketch and upscaling, prompt attention and matrix controls, loopback processing and a large ecosystem of extensions. It ships a one-click install and run script and is free under the AGPL-3.0 license.

Upscayl is a free and open-source desktop app that enlarges and enhances low-resolution images with AI, for Linux, macOS and Windows. Upscaling runs on your own Vulkan-compatible GPU, and you can load custom upscaling models from the project custom-models repository alongside the built-in ones.

ACE-Step is an open-source music generation foundation model that combines diffusion generation with Sana Deep Compression AutoEncoder and a lightweight linear transformer, synthesizing up to 4 minutes of music in about 20 seconds on an A100. Beyond text-to-music it supports voice cloning, lyric editing, remixing and track generation such as lyric-to-vocal and singing-to-accompaniment, with weights, training code and a Gradio UI released for building tools on top.

InvokeAI is a locally installed creative engine for Stable Diffusion models, with a web-based UI for generating and refining visual media. Its Unified Canvas covers in and out-painting with brush tools, node-based workflows build custom generation pipelines, and board and gallery management keeps images with their prompt and setting metadata. Free under a commercially-friendly license and installed with the Invoke Launcher.

Vercel Domains is Vercel's domain registrar: search for a name, check availability instantly, and buy it at registrar cost with private search, with nameservers configured automatically for your Vercel project. Domains can be managed from the dashboard, the vercel domains CLI or the Domains Registrar API for search, pricing, purchase, renewal and transfer.

Is Agentic is a free agent-readiness score for websites from Vercel, scored by Ora. Enter a URL and it audits what AI agents can discover, retrieve, understand and use, from server-rendered content and HTTP behavior to canonical URLs and Open Graph metadata, with essential, recommended and emerging checks that each carry evidence and a recommendation. Reports are readable in the browser, in the npx is-agentic CLI, through a public JSON API and an MCP server.

Ghidra MCP Server is a Model Context Protocol server that bridges Ghidra reverse engineering to AI agents, exposing 253 tools for decompilation, renaming, typing, commenting, structure creation, script execution, P-code emulation, data-flow analysis and live debugging. It runs as a GUI plugin or headless server with batch operations, naming-convention enforcement and Docker deployment, plus Ghidra Server integration for shared projects.

REA is an agent toolkit for reverse engineering anything, from app behavior down to native binaries. It connects your coding agent and terminal to tools that inspect native binaries, JavaScript and Electron apps, .NET assemblies and websites, wiring into Hopper or Ghidra so the agent can decompile a feature, explain how it works with evidence, and rebuild a version for your product. Analysis runs locally and conclusions carry their evidence and limitations.

SoL-Refiner is a video refinement model from NVIDIA that turns low-resolution output from any base video generator into 2K or 4K video with a single denoising step, instead of the multi-step refinement that adds a second sampling bottleneck. It is trained with high-resolution continual training, reinforcement-learning post-training and one-step distribution-matching distillation, with bidirectional and streaming inference.

A collection of skills for AI coding agents from Vercel Labs, packaged in the Agent Skills format. They audit a deployed Vercel project for cost and performance, enforce React, Next.js and React Native best practices, review UI code against accessibility and UX rules and docs against the Vercel writing handbook, and deploy apps to Vercel as claimable deployments from a conversation.

Archify is an agent skill that turns anything you want to understand, plan or share, from an idea or a plan to an entire codebase, into an interactive HTML visual you can explore, customize and share. Describe the subject to your coding agent and it produces the diagram, from travel itineraries and learning maps to source-backed architecture maps of a repository. It works with Claude Code, Codex, Cursor and OpenCode.

A curated collection of Codex plugin examples from OpenAI: each plugin pairs a manifest with skills, MCP server configuration, agents, commands and hooks, and the repo ships the marketplace manifests Codex installs from. Bundles cover Figma, Notion, Expo, Netlify, Remotion and Google Slides, plus iOS, macOS and web app building workflows.

HumanLayer is a multiplayer coding agent workspace and cloud, a control plane where a team's agent sessions, plan artifacts and code diffs come together so the team can collaborate and ship. It runs harnesses like Claude Code and Codex from your workstation to background agents on remote workers, with bring-your-own compute, environment and harness. The GitHub repo is the public issues home for the product, rebuilt from the older open code.

Ponytail is a skill and plugin for coding agents that forces the laziest solution that actually works: before writing code it walks a ladder from skipping the task entirely (YAGNI), through reusing what the codebase, standard library, native platform features and installed dependencies already offer, down to the minimum code that works. It stays active every response at lite, full or ultra intensity and installs into Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot and other harnesses.

A coding-agent skill that reshapes every reply for a reader with ADHD so the answer is never buried: the next action leads, multi-step work is numbered, state is restated each turn, tangents are suppressed and time estimates are specific. Invoke it with /i-have-adhd and it stays on until you say stop adhd mode. No ADHD diagnosis needed.

ECC is an MIT-licensed engineering system and toolbox for coding agents: it gives them a coordinated plan, test, implement, review, verify, remember, improve loop instead of raw prompting. It ships 68 specialized agents, 293 skills, hooks, rules, unified memory and AgentShield security scanning. It works best with Claude Code and also covers Codex and Kimi Code, with capability-limited adapters for Cursor, OpenCode, Gemini, Zed and GitHub Copilot.

Kolibri-1 is Aleph Alpha's open-weight mixture-of-experts reasoning model for German and English, released under Apache 2.0 with 78B total and 3.46B active parameters. It supports explicit reasoning, tool calling and context up to 1,048,576 tokens, ships FP8 weights with a 78 GB footprint, and is served through Aleph Alpha's vLLM plugin.

Impeccable is a design skill and design vocabulary for coding agents, giving them the design expertise to build interfaces from scratch, improve existing ones, and stay true to your design system. Its commands create, critique, refine and harden UI work and flag common AI-generated tells. It installs into Codex CLI, Claude Code, GitHub Copilot, Cursor, Gemini CLI, OpenCode and other agent harnesses.

A ready-made recipe for running Qwen3.8-27B on RTX 3090s through a one-command SGLang container, with INT4 AWQ, speculative decoding and multimodal input. Packaged builds for one, two or four GPUs give you a local chat app for text, images and video — 75 tokens/s on a single card, 290 on four.

A DeepSeek Harness plugin for archived sessions: one cross-workspace list to restore any of them or delete it for good, plus a red Delete permanently item in the session context menu and bulk delete with select all. It covers what dsh itself never does — reclaiming session files from disk.

Drop-in memory layer for AI agents and apps that gives them context which persists across conversations. Use the open-source Python or JavaScript SDK or the managed platform to give agents long-term memory in production.

A curated collection of 1000+ agent skills from official dev teams and the community, hand-picked rather than mass-generated. Works in Claude Code, Codex, Gemini CLI, Cursor, GitHub Copilot, OpenCode, Windsurf and more.

Prints a token-efficient Go CLI, a Claude Code skill and an MCP server for any API or website, so agents get muscle-memory commands instead of hunting through docs. It reads official API docs and existing community CLIs, keeps a local SQLite mirror, and layers compound insight commands the raw API cannot answer.

Open source alternative to Semrush and Ahrefs: keyword research, rank tracking, competitor insights, backlink analysis and site audits in one self-hostable web app. Bring your own DataForSEO API key and pay per use, and let coding agents query your SEO data directly through the bundled MCP server and agent skills.

OpenViking is a self-evolving context database for AI agents that unifies agent memory, knowledge RAG and skills in one store. Self-host it with Docker or a one-click Railway deploy, and keep what agents know, retrieve and can do in a single place.

A Prompt-as-Code library for GPT-Image 2 and 2.5: 500+ reverse-engineered image cases and 20+ industrial prompt templates, each published with its full prompt and generation record, plus reusable agent skills. A same-prompt comparison gallery for 2.5 lives on the companion site.

Phonon is the speech recognition engine from Fermion Research, running the Phonon models on Apple silicon through MLX, on x86-64 and Arm CPUs, and on NVIDIA GPUs. Transcribe files from the command line or live from the microphone, or serve an OpenAI-compatible transcription endpoint — Phonon-2, a 164 MB download, is the current model.

Superpowers is a complete software development methodology for your coding agents, built on top of a set of composable skills and some initial instructions that make sure your agent uses them.

A more powerful vision toolkit—give text-only models in DeepSeek Harness eyes: image Q&A, long-screenshot OCR, UI restoration, and GUI visual tasks in one toolkit and Skill.

Every agent has a budget. Every task has a receipt. AI spend should look like payroll, not like a surprise. Paperclip gives each agent a budget with a hard cap, meters every task, and rolls costs up in real time — overruns are impossible, not just unlikely.

Open source home automation that puts local control and privacy first. Powered by a worldwide community of tinkerers and DIY enthusiasts.

Turn any technical book, document folder, or collection of sources into a unified agent skill — ready to study, reference, and use while you work in GitHub Copilot CLI, Amp, Claude Code, Hermes Agent, or OpenClaw.

Recipes for serving LLMs locally on RTX 3090s (and other NVIDIA cards): working compose configs, patches and measured numbers, across vLLM, SGLang, llama.cpp, ik_llama and exllamav3. Pick a model for your card count, launch it with one command, and get an OpenAI-compatible API.

Qwen3.8-27B is a big model. This build makes it fast on consumer hardware: Up to 117 t/s decode on two RTX 3090s at full 262K context ~1.4x to 1.5x faster than the same model with native MTP4 speculation ~2.5x faster than plain autoregressive decoding 98% top-1 agreement with the original BF16 model (mean KLD 0.0007) 28 GiB on disk, 2.02 GiB drafter, full 262K context on 2×24 GB

AnythingLLM is the all-in-one AI application that lets you build a private, fully-featured ChatGPT—without compromises. Connect your favorite local or cloud LLM, ingest your documents, and start chatting in minutes. Out of the box you get built-in agents, multi-user support, vector databases, and document pipelines — no extra configuration required. AnythingLLM supports multiple users as well where you can control the access and experience per user without compromising the security or privacy of the instance or your intellectual property.

LocalAI is the open-source AI engine. Run any model - LLMs, vision, voice, image, video - on any hardware. No GPU required. A small core, not a bundle. Each backend wraps a best-in-class engine (llama.cpp, vLLM, whisper.cpp, stable-diffusion, MLX...) in its own image, pulled only when a model needs it. You install nothing you don't use.

Four principles in one file that directly address these issues: Principle Addresses Think Before Coding Wrong assumptions, hidden confusion, missing tradeoffs Simplicity First Overcomplication, bloated abstractions Surgical Changes Orthogonal edits, touching code you shouldn't Goal-Driven Execution Leverage through tests-first, verifiable success criteria

Skills For Real Engineers My agent skills that I use every day to do real engineering - not vibe coding. Developing real applications is hard. Approaches like GSD, BMAD, and Spec-Kit try to help by owning the process. But while doing so, they take away your control and make bugs in the process hard to resolve. These skills are designed to be small, easy to adapt, and composable. They work with any model. They're based on decades of engineering experience. Hack around with them. Make them your own. Enjoy.