Top AI GitHub Repos Worth Starring in 2026
Curated top AI GitHub repos in 2026: OpenClaw, Hermes, Pi Agent, OpenCode, DeepSeek-V4, skills, frameworks, memory, and MCP. Shortlists with stars and setup guides.
Dragos21 min read
GitHub is full of AI repos. Most are forks, thin wrappers, or things nobody committed to in six months. This page is the short list I actually use and recommend: projects with real maintenance, clear jobs, and enough community weight that you’re not gambling on a weekend demo.
Same idea as my Docker containers for home server catalog, but for AI tools instead of self-hosted apps. I cut hard. If a project stalled or the company shut down (Coqui TTS, Bark, AUTOMATIC1111 SD WebUI, Zep, Flowise freeze, Helicone maintenance mode), it’s not here.
Star counts are approximate as of late July 2026 and move fast. Treat them as a momentum signal, not a quality ranking.
How to use this page
- Pick a lane in the table below (local AI, assistants, coding, frameworks, etc.).
- Star a shortlist — three to five repos per category is plenty.
- Open one section you care about. Each has a quick-pick table, then short notes on the ones that matter most.
- Where I have a full guide on bitdoze, I link it.
Don't star 50 repos and call it research
Pick what matches what you’re building this month. A focused watchlist beats a starred junk drawer.
Pick your lane
| I want… | Go here | Typical first picks |
|---|---|---|
| Run LLMs on my hardware | Local AI | Ollama, Open WebUI, llama.cpp |
| Strong open-weight models | Open-weight models | DeepSeek-V4, Kimi K3, GLM-5.2 |
| A personal AI assistant | Assistants | OpenClaw, Hermes Agent |
| Humans + agents in one workspace | Assistants | Buzz |
| Coding agent in the terminal | Coding agents | OpenCode, Pi Agent |
| Skills / playbooks for agents | Skills | Spec Kit, superpowers, pi-skills |
| Multi-agent systems in code | Frameworks | LangChain, Mastra, Agno, CrewAI |
| Visual AI workflows | Workflows | n8n, Dify, Langflow |
| Long-term agent memory | Memory | Mem0, Cognee, Hindsight |
| Embeddings / RAG storage | Vector DBs | pgvector, Qdrant, Milvus |
| Images and video | Multimodal | ComfyUI, Remotion |
| Trace and test LLM apps | Observability | Langfuse, promptfoo |
| One API across providers | Observability | LiteLLM, OmniRoute |
| Connect tools via MCP | MCP | awesome-mcp-servers |
| Learn the stack | Learning | Generative AI for Beginners |
What changed in 2026 (short)
- OpenClaw went from a niche self-hosted assistant to ~385K stars. Local-first, many chat channels, skills that extend themselves. Creator joined OpenAI; the project moved toward a foundation model of ownership. I still run it — with a security review.
- Hermes Agent from Nous Research (~223K stars) is the other heavyweight personal agent. Built-in learning loop, skill creation from experience, OpenClaw migration path. Full write-up: Hermes setup guide.
- OpenCode (~191K stars under
anomalyco/opencode) and Pi Agent (earendil-works/pi) are the open terminal coding agents I actually install. Pi stays minimal; you add skills and extensions yourself. - Open-weight coding models (DeepSeek-V4, Kimi K3, GLM-5.2, MiniMax M3, and friends) made cheap agents practical. Full roundup: best open-source LLMs as Claude alternatives.
- Buzz by Block (July 2026) is the new workspace play: humans and agents share channels, git, and workflows on a self-hostable Nostr relay. Guide: Buzz Docker setup.
- Skills and playbooks became a real category: Spec Kit, superpowers, pi-skills, token savers, deploy skills. Agents without skills are just chat with tools.
- Agent memory is its own category now — Mem0, Cognee, Hindsight, Letta. People stopped treating “the context window” as long-term memory.
- MCP stuck as the common way to plug tools into agents. The awesome list is the index.
- Dead or stalled projects are gone from this page on purpose. Alive code only.
Local AI & model runners
Run models on your CPU, GPU, or Apple Silicon. No API key required for the happy path.
Quick picks
| # | Repo | Stars | Lang | What it does |
|---|---|---|---|---|
| 1 | Ollama | 177K | Go | Download, run, and serve LLMs with one command |
| 2 | Open WebUI | 147K | Python | ChatGPT-style UI for Ollama and OpenAI-compatible APIs |
| 3 | llama.cpp | 122K | C++ | Inference engine under most local stacks |
| 4 | vLLM | 88K | Python | High-throughput serving for production |
| 5 | AnythingLLM | 64K | JS | Desktop AI with built-in RAG over your docs |
Ollama is still how most people start. ollama run pulls a model and serves it. Desktop apps exist for macOS and Windows. Pair it with Open WebUI for a full self-hosted chat product. I covered install in Ollama with Docker and wiring it into OpenClaw in OpenClaw + Ollama.
Open WebUI is the dashboard: chat UI, RAG, multi-user options, voice bits. If Ollama is the engine, this is the cockpit.
llama.cpp is the library almost everyone builds on for local inference. You may never clone it, but it’s in your dependency tree if you care about GGUF on consumer hardware.
vLLM is what you reach for when “it works on my laptop” becomes “we need tokens per second on a GPU box.”
Open-weight AI models
This catalog is mostly tools and harnesses. For model picks (pricing, benchmarks, when to use each), use the full guides:
- Best open-source LLMs as Claude alternatives — Kimi K3, GLM-5.2, Qwen 3.6 Plus, MiniMax M3, MiMo V2.5 Pro, Mistral Medium 3.5 vs Opus 5 / Fable 5 / GPT-5.6
- Best open-source models for OpenClaw — practical OpenClaw primary + fallback
- Best cheap models for Hermes — agent pricing and DeepSeek-V4 in the cheap 1M-context stack
Quick picks (models I actually point people at)
| # | Model | Home | Role |
|---|---|---|---|
| 1 | DeepSeek-V4-Pro | HF + API | Flagship MoE, ~1M context, heavy coding / agent work |
| 2 | DeepSeek-V4-Flash | HF + API | Smaller V4 — good default behind always-on agents |
| 3 | Kimi K3 | Moonshot | Largest open-weight MoE; coding + web peaks |
| 4 | GLM-5.2 | Z.AI | Daily-driver coding workhorse |
| 5 | MiniMax M3 | MiniMax | Budget always-on agent model, 1M multimodal context |
| 6 | MiMo V2.5 Pro | Xiaomi | Long-horizon agent / tool-heavy runs |
| 7 | Qwen 3.6 Plus | Alibaba | Frontend / vibe-coding on a budget |
| 8 | Mistral Medium 3.5 | Mistral | Dense open weights, realistic self-host option |
DeepSeek-V4 is the 2026 open-weight line to know by name: Pro for heavy work, Flash when cost and latency matter. Weights live on Hugging Face; the API is live for both.
The rest of the table is the shortlist from the Claude-alternatives article. Don’t star eight model cards and call it research — pick one primary and one fallback for the agent you’re running, then read the full comparison if the bill or quality isn’t right.
AI personal assistants
Always-on agents that live on your machine or VPS, talk over chat apps, and do work for you.
Quick picks
| # | Repo | Stars | Lang | What it does |
|---|---|---|---|---|
| 1 | OpenClaw | 385K | TypeScript | Personal AI assistant, many channels, skills, local-first |
| 2 | Hermes Agent | 223K | Python | Self-improving agent with learning loop (Nous Research) |
| 3 | AutoGPT | 186K | Python | Long-running autonomous agent platform |
| 4 | Buzz | 19K | Rust | Humans + agents workspace (chat, git, workflows) |
OpenClaw is the default answer when someone says “I want my own AI assistant on a VPS.” Messaging integrations (Telegram, Slack, Discord, WhatsApp, and more), shell and browser tools, skills marketplace, and a huge community. It also needs care: broad permissions and community skills mean you should read OpenClaw security before you open it to production systems. Setup: OpenClaw guide. Alternatives: OpenClaw alternatives.
Hermes Agent is the one I put next to OpenClaw on purpose. Nous Research built a learning loop: skills from completed work, recall across sessions, user modeling. It can migrate OpenClaw config and memories. If you want an agent that improves instead of only following a static prompt pack, start here. Guides: Hermes setup, Hermes dashboards, cheap models for Hermes.
AutoGPT pioneered the autonomous-agent hype in 2023 and still has a huge star count. Better as a platform for long-running automation than as a “chat with me on Telegram” daily driver.
Buzz (buzz.xyz) is different enough that people mis-file it. Block released it as an open-source workspace where agents are first-class members with cryptographic identities, not bots bolted onto Slack. Channels, DMs, git hosting, workflows, desktop app, self-hostable relay. Early (pre-1.0) but already useful if your problem is “team + agents in one place.” Full path: self-host Buzz with Docker.
Security on always-on agents
OpenClaw, Hermes, and friends need shell, browser, or message access to be useful. Treat them like privileged ops tools: least privilege, review skills, don’t paste production secrets into prompts, isolate with Docker or a dedicated VPS.
AI coding agents & tools
Agents that live in the terminal or IDE, read the repo, and take multi-step actions.
Quick picks
| # | Repo | Stars | Lang | What it does |
|---|---|---|---|---|
| 1 | OpenCode | 191K | TypeScript | Open coding agent with a serious TUI, multi-provider |
| 2 | Pi Agent | 81K | TypeScript | Minimal terminal coding harness — skills + extensions |
| 3 | OpenHands | 83K | TypeScript | Autonomous software engineer in a sandbox |
| 4 | Cline | 65K | TypeScript | VS Code autonomous coding agent |
| 5 | Herdr | 23K | Rust | Terminal multiplexer built for agent herds |
OpenCode is my batteries-included pick. Multi-provider, strong TUI, plan mode, no single-vendor lock-in. Repo lives at anomalyco/opencode (old sst/opencode redirects). Guides: OpenCode setup, OpenCode Go plan.
Pi Agent is the minimal one I keep next to OpenCode. Core tools only (read, write, edit, bash); memory, MCP, sub-agents, and themes come from extensions and skills you choose. Built by Mario Zechner, now under Earendil (earendil-works/pi). Install via pi.dev or npm. Full path: Pi coding agent setup. Side-by-side: OpenCode vs Pi.
OpenHands is the “agent in a box that can drive a browser and fix its own errors” approach. Heavier, more autonomy when you want it.
Cline if you want the agent inside VS Code with terminal and browser control.
Herdr is not a coding model. It’s the runtime/multiplexer when you run several agents at once and need to see which one is blocked. Review: Herdr agent multiplexer.
Skills that plug into these agents (Spec Kit, superpowers, pi-skills, etc.) live in the skills section below.
Agent skills & playbooks
Reusable skills, methodologies, and skill packs. This is how modern agents get domain behavior without rewriting the harness.
Quick picks
| # | Repo | Stars | What it does |
|---|---|---|---|
| 1 | superpowers | 264K | Agentic skills framework + software-dev methodology |
| 2 | Spec Kit | 125K | GitHub’s toolkit for spec-driven development |
| 3 | caveman | 95K | Skill that cuts output tokens by talking like a caveman |
| 4 | claude-mem | ~89K | Persistent context across coding-agent sessions |
| 5 | scientific-agent-skills | 32K | Skills that turn an agent into a research assistant |
| 6 | pi-skills | ~2K | Skills for Pi (also usable with Claude Code / Codex) |
| 7 | pi-config | — | Curated Pi extensions: web, PDF, bash guards, Reddit |
superpowers is both a skill pack and a way of working with coding agents (structured workflows, sub-agents, brainstorming protocols). Huge star count because people were tired of freeform vibe coding.
Spec Kit is GitHub’s official SDD kit: write the PRD/spec first, then let the agent implement against it. Belongs here more than under “coding agents.”
caveman and claude-mem are practical painkillers: fewer tokens in the response, and memory that survives session restarts.
pi-skills and pi-config are what I install after Pi itself. LazyPi bundles a bigger skill pack in one command if you want a curated starter set (see the Pi setup guide).
For home-lab deploy automation, I also wrote a reusable docker-deploy skill you can drop into OpenClaw, OpenCode, or a custom agent.
AI agent frameworks
Code-first libraries when you’re building products, not only chatting with an assistant.
Quick picks
| # | Repo | Stars | Lang | What it does |
|---|---|---|---|---|
| 1 | LangChain | 143K | Python | Default agent/app framework + ecosystem |
| 2 | MetaGPT | 70K | Python | Multi-agent “software company” simulation |
| 3 | AutoGen | 60K | Python | Microsoft multi-agent conversations |
| 4 | CrewAI | 56K | Python | Role-based agent crews and tasks |
| 5 | LlamaIndex | 51K | Python | Data/RAG-first agents and indexes |
| 6 | Agno | 42K | Python | Build and run agent platforms (teams, memory) |
| 7 | Mastra | 27K | TypeScript | TypeScript agents, workflows, memory, Studio UI |
LangChain still sits at the center of a lot of tutorials and integrations. Pair it with LangGraph when you need cycles and durable state.
CrewAI, AutoGen, and MetaGPT take different multi-agent angles: roles and tasks, multi-agent chat, and a full “virtual software company” pipeline.
Agno is the Python framework I keep recommending for real bots (Discord, teams, knowledge bases). Guides: Agno getting started, Discord bot with Agno, Agno squads, Agno MCP tools.
Mastra is the TypeScript option if your stack is already Node/Bun. Agents, Zod tools, memory, schedules, Studio UI. Guides: build an agent with Mastra, Mastra tools vs MCP, Mastra vs Eve.
Workflow orchestration
Visual builders when non-developers (or you on a Friday) need pipelines without a full custom app.
Quick picks
| # | Repo | Stars | Lang | What it does |
|---|---|---|---|---|
| 1 | n8n | 199K | TypeScript | Workflow automation + native AI nodes, self-hostable |
| 2 | Langflow | 153K | Python | Drag-and-drop agents and RAG graphs |
| 3 | Dify | 151K | TypeScript | Production AI app platform: workflows, RAG, agents |
n8n is the one I install first for “when X happens, call an LLM and do Y.” Fair-code, huge integration list, solid self-host story. Guide: n8n self-host.
Langflow is better when the product is the agent graph (prompts, retrievers, tools on a canvas). Guide: Langflow Docker.
Dify sits between them: ship AI apps with workflows, knowledge bases, and monitoring without starting from a blank Python repo.
Flowise is not on this list. Development freezes and sunset talk make it a bad default for new work.
AI memory & RAG
Give agents something longer than the current context window.
Quick picks
| # | Repo | Stars | Lang | What it does |
|---|---|---|---|---|
| 1 | Firecrawl | 158K | TypeScript | Turn websites into clean Markdown/JSON for LLMs |
| 2 | RAGFlow | 86K | Go | Full RAG engine: parse, index, cite, agentic flows |
| 3 | Mem0 | 62K | Python | Universal memory layer with a simple API |
| 4 | Cognee | 30K | Python | Knowledge-graph memory for agents |
| 5 | Letta | 24K | Python | Stateful agents with self-editing memory (ex-MemGPT) |
| 6 | Hindsight | 19K | Python | Memory that stores experiences/patterns, not only chat logs |
Mem0 is the easy “add memory” default for many apps.
Cognee builds a structured knowledge graph. I self-host it: Cognee self-host.
Hindsight aims at agents that learn from experience rather than replaying transcripts. Deploy notes: Hindsight Docker. Head-to-head: Cognee vs Hindsight.
Letta (formerly MemGPT) is still the research-rooted stateful agent platform.
Firecrawl and RAGFlow cover the document/web side: get clean data in, then answer with citations.
Vector databases
Where embeddings live for RAG and semantic search.
Quick picks
| # | Repo | Stars | Lang | What it does |
|---|---|---|---|---|
| 1 | Milvus | 45K | Go | Scale-out vector DB for large corpora |
| 2 | Qdrant | 34K | Rust | Fast vector search, clean ops story |
| 3 | Chroma | 29K | Rust | Lightweight embedding DB for prototypes and apps |
| 4 | pgvector | 22K | C | Postgres extension — often enough |
If you already run Postgres, pgvector is the boring correct choice. I use it that way. Milvus or Qdrant when scale or pure vector workloads justify another service. Related: pgvector + pgAdmin in Docker.
Multimodal & generation
Quick picks
| # | Repo | Stars | Lang | What it does |
|---|---|---|---|---|
| 1 | ComfyUI | 123K | Python | Node-based image/video/audio generation |
| 2 | Deep-Live-Cam | 95K | Python | Real-time face swap / live video transforms |
| 3 | Remotion | 55K | TypeScript | Make videos programmatically with React |
| 4 | HyperFrames | 39K | TypeScript | Write HTML, render video (agent-friendly) |
| 5 | OpenVoice | 37K | Python | Instant voice cloning with style controls |
| 6 | F5-TTS | 15K | Python | Strong zero-shot voice cloning |
ComfyUI is the default image (and increasingly video) workbench after AUTOMATIC1111 stalled. Node graphs, huge ecosystem, still moving.
Remotion is the React path for video: compositions as components, render with Node/Lambda, full control over timeline and data. Good fit when agents or scripts should compose motion graphics and explainers instead of only calling a black-box video API. Company-friendly license for commercial use (check their licensing for production).
HyperFrames is closer to “write HTML, get a video” for agent pipelines. Use Remotion when you want a real React codebase; HyperFrames when the agent should ship HTML frames.
Voice: F5-TTS and OpenVoice cover most local cloning needs. Coqui and Bark are out because the upstream stories ended.
Observability & gateways
Quick picks
| # | Repo | Stars | Lang | What it does |
|---|---|---|---|---|
| 1 | LiteLLM | 55K | Python | OpenAI-compatible proxy over 100+ providers |
| 2 | OmniRoute | 35K | TypeScript | Free MIT AI gateway — many providers, free tiers, one endpoint |
| 3 | Langfuse | 32K | TypeScript | Tracing, evals, prompt management for LLM apps |
| 4 | promptfoo | 24K | TypeScript | CLI evals and red-teaming, CI-friendly |
LiteLLM is the glue when you don’t want every service hardcoding a different SDK. Cost tracking, fallback, load balancing. Guide: LiteLLM Docker.
OmniRoute is the free-tier-friendly gateway: one endpoint, a huge provider list (including free ones), token compression, and smart fallback. Useful when you want coding agents (Claude Code, Codex, Cursor, OpenCode, Pi) to keep working without rewriting configs every time a free endpoint dies. MIT, actively maintained.
Langfuse is what I reach for to see traces and cost. Guide: Langfuse Docker.
promptfoo is for “does this prompt still pass the suite?” before you ship.
MCP & tool connectivity
MCP (Model Context Protocol) is how agents talk to tools without every vendor inventing a private plugin format.
Quick picks
| # | Repo | Stars | Lang | What it does |
|---|---|---|---|---|
| 1 | awesome-mcp-servers | 92K | — | Directory of MCP servers for apps and APIs |
| 2 | MCP Python SDK | 24K | Python | Official SDK for servers and clients |
If you’re new to the protocol, start with MCP for beginners. For concrete wiring into agents, see the Agno and Mastra MCP pieces linked above.
Learning & reference
Quick picks
| # | Repo | Stars | What it does |
|---|---|---|---|
| 1 | prompts.chat | 151K | Large open prompt library |
| 2 | System prompts of AI tools | 130K | Collected system prompts from popular tools |
| 3 | Generative AI for Beginners | 108K | Microsoft course, lessons + code |
| 4 | LLMs from Scratch | 88K | Build a ChatGPT-like model from first principles |
Use the courses to learn. Use the prompt repos to steal structure, not to paste blindly into production agents.
Default picks by use case
| Use case | Default pick |
|---|---|
| Local chat | Ollama + Open WebUI |
| Open-weight model | DeepSeek-V4-Flash / MiniMax M3 (cheap) or GLM-5.2 / Kimi K3 (quality) — full list |
| Personal assistant | OpenClaw or Hermes Agent |
| Team + agents workspace | Buzz |
| Terminal coding agent | OpenCode or Pi Agent |
| Minimal coding harness | Pi Agent + pi-skills / pi-config |
| Spec-driven coding | Spec Kit or superpowers |
| VS Code agent | Cline |
| Python multi-agent product | Agno or CrewAI |
| TypeScript agent product | Mastra |
| Visual automation | n8n |
| Visual AI apps | Dify or Langflow |
| Simple agent memory | Mem0 |
| Graph memory | Cognee |
| Learning memory | Hindsight |
| Vector search (simple) | pgvector |
| Vector search (scale) | Milvus or Qdrant |
| Image generation | ComfyUI |
| Programmatic video (React) | Remotion |
| Agent-friendly HTML video | HyperFrames |
| LLM proxy | LiteLLM or OmniRoute |
| Observability | Langfuse |
| MCP discovery | awesome-mcp-servers |
Getting started (one weekend)
- Local chat —
ollama run+ Open WebUI in Docker (guide). - Assistant — OpenClaw or Hermes on a small VPS (OpenClaw, Hermes).
- Coding — OpenCode or Pi against a real repo (OpenCode, Pi).
- Skills — add Spec Kit or pi-skills instead of pasting giant system prompts.
- Automation — one n8n workflow that calls an LLM (n8n).
- Memory — Mem0 or Cognee on something you already run (Cognee).
- Observability — Langfuse if you’re past demos (Langfuse).
Don’t star the whole page on day one. Pick one row from the table, ship something, then expand.
Guides on this site (related)
| Topic | Article |
|---|---|
| OpenClaw install | OpenClaw setup |
| OpenClaw safety | OpenClaw security |
| OpenClaw alternatives | Alternatives roundup |
| Open-weight models | Claude open-source alternatives |
| Models for OpenClaw | OpenClaw model picks |
| Hermes Agent | Hermes setup |
| Pi Agent | Pi setup |
| OpenCode vs Pi | Comparison |
| Buzz workspace | Buzz Docker setup |
| Deploy skill | Docker deploy skill |
| Mastra | Build with Mastra |
| Agno | Agno start |
| OpenCode | OpenCode setup |
| Memory comparison | Cognee vs Hindsight |
| Home server context | Docker containers list |
FAQ
Why isn't project X on this list?
Usually one of: abandoned or code-frozen, pure hype with no maintenance, duplicate of something better on the list, or too narrow for a “top dogs” catalog. Star count alone doesn’t get you in. OpenClaw and Hermes are here because people run them daily, not only because they trend.
Should I use OpenClaw or Hermes Agent?
OpenClaw if you want the biggest ecosystem, channels, and skills community. Hermes if you care more about the learning loop and Nous Research’s approach. Plenty of people run both. Start with one, migrate later if needed — Hermes has an OpenClaw migration path.
Where does Buzz fit?
Buzz is not a drop-in OpenClaw clone. It’s a collaboration workspace (chat + git + workflows) where agents have first-class identities. Use it when the problem is multi-human, multi-agent collaboration. Self-host with the Buzz Docker guide.
OpenCode or Pi Agent?
OpenCode if you want a full TUI, plan mode, and lots built in. Pi if you want a small core and you like assembling skills/extensions yourself. I use both. Details: OpenCode vs Pi.
Python or TypeScript frameworks?
Python still has more agent framework surface area (LangChain, CrewAI, Agno). TypeScript is fine and often better if your product is already Node/Bun — use Mastra. Don’t rewrite your whole stack just to match a framework language.
How often will this page update?
When major projects rise, die, or change ownership. The 2026 story already killed or sidelined several previous “must install” tools. If a repo on this page goes dark, it comes off.
Bottom line
Most of GitHub’s AI section is noise. The useful core is small:
- Run models: Ollama, Open WebUI, llama.cpp, vLLM
- Open weights: DeepSeek-V4, Kimi K3, GLM-5.2, MiniMax M3 (full guide)
- Personal assistants: OpenClaw, Hermes Agent
- Team + agents workspace: Buzz
- Coding agents: OpenCode, Pi Agent, Cline, OpenHands
- Skills: Spec Kit, superpowers, pi-skills, claude-mem
- Build products: LangChain, Agno, Mastra, CrewAI
- Workflows: n8n, Dify, Langflow
- Multimodal: ComfyUI, Remotion, HyperFrames
- Memory / RAG: Mem0, Cognee, Hindsight, RAGFlow, pgvector or Qdrant/Milvus
- Ops: LiteLLM, OmniRoute, Langfuse, MCP directory
Star what matches your next project. Ignore the rest until you have a real reason to care.
