Trending AI GitHub Repos
Trending open source AI and machine learning repositories on GitHub.
Showing 50 of 196 items
ds4
Local DeepSeek 4 engine targeting Metal, CUDA, and ROCm GPUs for fast running of the AI. If you want serious open models on consumer hardware, this is one of the most important new runtimes.
gstack
Ships Garry Tan’s full Claude Code setup, with tools that act like a small executive team. If you want an opinionated starter kit for AI first workstations, this is a strong baseline.
pstack-claude
Ports Poteto’s pstack workflow ideas to Claude and other models, with ready made agent flows. Handy if you want disciplined multi step coding agents without inventing your own process from scratch.
OpenMontage
Agentic video production stack with many tools and pipelines wired up. Ideal if you want to test "editor" style agents that plan, cut, and render full videos instead of single clips.
text-to-cad
Turns natural language or agent instructions into CAD geometry. If you are exploring AI assisted hardware or robotics design, this repo gives your agents direct hooks into CAD workflows.
marketingskills
Curated marketing skills, prompts, and workflows tailored for Claude Code and other AI agents. Great starting kit if you want agents that actually understand CRO, copy, and analytics tasks.
ponytail
Opinionated framework that makes your coding agent behave like a lazy senior engineer, preferring deletion and simplification. Use it if your agents over build and you want them to cut scope instead.
Agent-Reach
Single CLI that lets agents read and search Twitter, Reddit, YouTube, GitHub, Bilibili, and more without per site APIs. If your agents feel blind outside docs and PDFs, this is a fast way to grow their reach.
agent-skills
Production grade engineering skills and patterns for AI coding agents, distilled from real projects. If your agents flail on real repos, this is a gold mine of reusable capabilities.
impeccable
Design language and components that help AI harnesses produce cleaner, more consistent UI output. If you ship agent frontends, you can drop this in to raise design quality without hand coding every layout.
claude-mem
Adds persistent memory to many popular agent harnesses by logging, compressing, and reinjecting past context. Drop it in if you want your agents to remember prior sessions without building a custom memory layer.
google/artemis
Google’s toolkit for instrumenting and stress-testing AI-heavy systems, including agents. It wraps telemetry, replay, and fault injection into one consistent harness.
ayghri/i-have-adhd
An agent skill that restructures responses for people who struggle with dense outputs. It forces coding agents to surface the answer first and keep context minimal.
stablyai/orca
Orca wraps many coding agents into one "agent IDE" so you can run parallel agents with your own API keys. It focuses on orchestration, comparison, and safety checks.
bilawalsidhu/gods-eye-view
A browser-based globe that overlays real geospatial data so you can query, monitor, and visualize the world live. It is becoming a go-to open spatial-intelligence front end.
cloudflare/security-audit-skill
A skill package that turns coding agents into multi-phase security auditors. It structures scanning, triage, and reporting into machine-readable findings you can feed back into CI.
blader/humanizer
A text filter that strips obvious AI "tells" from generated writing. It is becoming a standard post-processing step for people shipping AI-written content.
alibaba/open-code-review
CLI tool that runs a disciplined AI code review pipeline on your diffs. It combines hard-coded checks with an LLM agent to give precise, line-level feedback.
anthropics/knowledge-work-plugins
A set of official plugins aimed at knowledge workers using Claude. They standardize tasks like research, note-taking, and summarization into reusable harnesses.
TauricResearch/TradingAgents
A suite of autonomous trading agents wired into real markets with safety rails. It is a live testbed for how agents handle money, latency, and slippage.
Tencent/WeKnora
Tencent’s framework for building knowledge-grounded agents over enterprise data. It focuses on secure connectors, retrieval, and governance for regulated workloads.
datahub-project/datahub
DataHub is a metadata and lineage platform pitched as the “context layer” for your data and AI stack. It tracks where data comes from, how models use it, and who owns what. If your AI projects sprawl across teams and tables, this helps you keep a map.
cvat-ai/cvat
CVAT is a mature labeling tool for images, video, and 3D data with AI assisted helpers. Teams use it to build high quality datasets for vision models. If you plan serious custom vision work, this is one of the few battle tested options.
memgraph/memgraph
Memgraph is an in memory graph database marketed directly for GraphRAG, AI memory, and agentic AI workloads. It speaks a Cypher style query language and targets real time graph analytics. If you hit limits with vector databases for knowledge graphs, this is a concrete alternative.
NVIDIA/Megatron-LM
Megatron-LM contains NVIDIA’s recipes and code for training huge language models across many GPUs. It focuses on tensor, pipeline, and sequence parallelism for very large setups. Use it as a reference when planning serious model training or comparing with newer training stacks.
huggingface/transformers
Transformers is the go to Python library for running and customizing modern language, vision, and audio models. It wraps dozens of model families with a common API and production ready utilities. If you prototype or ship LLMs, this is still the main entry point.
The-Swarm-Corporation/AutoHedge
AutoHedge wires agent swarms into trading workflows, from data gathering to risk checks and execution. It is a sandbox for building autonomous hedge fund style systems. If you play with AI in markets, read this carefully before reinventing your own fragile setup.
humanlayer/skills
Another skills collection, focused on agents that work alongside humans in existing tools. It packages patterns for communication, escalation, and context gathering. Use it if you want agents that feel less like black boxes and more like assistants that explain themselves.
OpenWhispr/openwhispr
OpenWhispr is a cross platform dictation app that can use local or cloud speech models. It targets privacy first workflows with good UX. If you want always on transcription without streaming everything to a big provider, this is a strong starting point.
aipoch/open-science
Open Science is a local first research workbench that mixes agents, notebooks, and data connectors under one roof. It tracks provenance so experiments can be reproduced. If you run a small lab or quant fund, this is an opinionated template for your AI stack.
magnitudedev/magnitude
Magnitude is an open source server for running local models behind a single API, wired into popular agents like Pi and Codex. It chooses the best model for your hardware automatically. If you want to own your stack but keep setup simple, this is worth testing.
cathrynlavery/diagram-design
A library of clean HTML plus SVG diagram templates tuned for Claude Code, Codex, and Pi. It gives agents a visual language that humans can actually read. If your AI writes architecture docs, wiring it to these templates will immediately make outputs less chaotic.
mattpocock/skills
A curated set of small, composable prompts that make coding agents behave like disciplined senior engineers. Skills cover grilling requirements, triaging bugs, and managing context. Copy these into your own repos if you want agents that feel less like toys and more like coworkers.
affaan-m/ECC
ECC is a giant harness for AI coding agents, with skills, memory, security rules, and editor integrations in one place. It turns local and cloud agents into something closer to a programmable IDE. If you are serious about agents writing production code, this is quickly becoming the default experimentation surface.
ruvnet/ruflo
Ruflo is an "agent meta harness" for running swarms of agents with memory, routing, and RAG baked in. It targets multi agent workflows, not single chat bots. Use it when you want several specialized agents cooperating over serious tasks.
NousResearch/hermes-agent
Hermes Agent aims to be a general purpose agent framework that can grow with user needs. It focuses on orchestrating tools, memory, and models rather than locking you into one stack. Use it if you want a serious, open playground for long running agents.
openai/skills
OpenAI's Codex skills catalog exposes reusable behaviors for coding agents, like refactors, tests, and migrations. It treats skills as building blocks you can wire into your own workflows. Even if you do not use Codex, the structure is a useful pattern for skill libraries.
anomalyco/opencode
OpenCode is a full coding agent stack that runs locally or in the cloud. It wraps models, tools, and editors behind a unified configuration. If you want a starting point for an in house coding agent instead of wiring plugins by hand, start here.
abhigyanpatwari/GitNexus
GitNexus runs in the browser and turns a repository into a knowledge graph with an embedded RAG style agent. It is aimed at exploring large unfamiliar codebases locally.
livekit/agents
LiveKit Agents is a framework for realtime voice and video agents that can listen, talk, and see. It abstracts away the hard parts of low latency media streaming.
tashfeenahmed/freellmapi
freellmapi aggregates dozens of free model providers behind one OpenAI compatible endpoint. It is a quick way to prototype without paying for every call.
handsomestWei/patent-disclosure-skill
This skill helps agents draft Chinese patent disclosures, mine key inventive points, and track policy changes. It packages a niche but high value workflow into an off the shelf tool.
pollen-robotics/microduck_rl
microduck_rl hosts reinforcement learning environments for the Microduck robot platform. It gives roboticists a ready sandbox for testing control policies before touching hardware.
punkpeye/awesome-mcp-servers
A curated list of Model Context Protocol servers that plug tools into agents like Claude and others. It is becoming the default index for extending agent stacks.
unclecode/crawl4ai
Crawl4AI is a web crawler built for language models, handling JavaScript heavy pages and structured extraction. It feeds cleaner data into retrieval and agent pipelines with minimal glue code.
p-e-w/heretic
Heretic strips safety filters from language models by learning how the guardrails behave, then routing prompts around them. It shows how fragile many current safety layers remain.
tt-a1i/archify
Archify is an agent skill for generating architecture and workflow diagrams as self contained HTML. It lets AI tools output diagrams directly instead of plain text descriptions.
K-Dense-AI/scientific-agent-skills
This library turns generic AI agents into science specialists by wiring them into 100+ tools and databases across biology, chemistry, and medicine. It gives researchers ready made skills instead of forcing them to hand craft workflows.
THU-MAIC/OpenMAIC
OpenMAIC simulates a multi agent interactive classroom where different AI agents play students and teachers. It is a testbed for studying how teaching agents learn, coordinate, and scale.
mvanhorn/last30days-skill
This agent skill scans Reddit, X, YouTube, Hacker News, Polymarket, and the wider web, then returns a grounded summary. It is a plug in for agents that need up to date context fast.