Top GitHub AI Agent Repositories 2026: OpenClaw, Karpathy Skills, Langflow, Strix and More

GitHub’s Octoverse 2025 report contained a number that reframed the entire open-source landscape: over 4.3 million AI-related repositories now exist on the platform, representing a 178 percent year-over-year jump in LLM-focused projects. The global AI agent market reached $7.84 billion in 2025 and is projected to hit $52.62 billion by 2030, growing at a compound annual rate of 46.3 percent. The GitHub data and the market data are telling the same story from different angles: AI agent development is the fastest-growing segment of software development, and the tools developers are choosing are being validated at scale on GitHub.

In this environment, a select group of repositories has emerged as clear frontrunners, each amassing tens or hundreds of thousands of stars by offering developers the tools to build autonomous agents, deploy models locally, and streamline AI-powered workflows. This article covers the most impactful AI and agentic AI repositories of 2026, from the fastest-growing project in GitHub history to the specialized tools defining how agents are built, secured, and deployed in production.

OpenClaw: 210,000+ Stars and the Fastest-Growing Open-Source Project of 2026

OpenClaw is the breakout star of 2026 and, by the most commonly applied standard, the fastest-growing open-source project in GitHub history — at least until DeepSeek Harness arrived in August and broke its records. Created by PSPDFKit founder Peter Steinberger, OpenClaw surged from 9,000 to over 60,000 stars in just a few days after going viral in late January 2026, and has since surpassed 210,000 stars.

OpenClaw is a coding agent built around a fundamentally open architecture. Where most coding agents bundle a specific model with their tool surface and require significant configuration to use a different provider, OpenClaw was designed as a model-agnostic platform from the start. It supports multiple AI providers through a unified interface and has become the reference implementation for what an open coding agent looks like in 2026.

The repository introduced the Agent Communication Protocol (ACP), an open transport standard that defines how agents communicate with each other and with external tools. The dsh-openclaw-acp bridge, which connects DeepSeek Harness to OpenClaw’s ACP transport, illustrates how OpenClaw has evolved from a standalone tool into the coordination layer for a broader agent ecosystem. The ACP specification has been adopted by several other agent projects as a standard integration point.

OpenClaw’s 177 production-ready SOUL.md configuration files, covering 24 categories including project management, SEO, DevOps, writing, and customer support, have also contributed to its adoption. These configurations allow teams to deploy specialized OpenClaw agents for specific professional workflows with minimal setup, making the tool accessible to non-engineering functions as well as developers.

Andrej Karpathy’s CLAUDE.md Skills File: 207K Stars for One File

The most starred Claude Code behavioral skill on GitHub encodes Karpathy’s viral observations about LLM coding pitfalls into four hard rules — and it is a single CLAUDE.md file with zero runtime dependencies. When Andrej Karpathy posted his frustrations with agentic coding failures in January 2026, it went viral across the AI developer community. Developer Forrest Chang turned the observations into a CLAUDE.md behavioral principles file. The repository hit 207,000 stars, making it one of the fastest-growing AI workflow repositories ever created.

The skill targets three failure patterns Karpathy identified directly: agents making silent wrong assumptions and proceeding without checking, over-engineering that inflates a 50-line solution into 500 lines, and orthogonal changes where the agent modifies code it was never supposed to touch. The four rules in the file address these patterns with behavioral constraints rather than technical controls.

The repository at github.com/multica-ai/andrej-karpathy-skills demonstrates something important about where the GitHub AI agent ecosystem is focusing attention in 2026: not just on building more capable agents, but on making existing capable agents more reliable, predictable, and controllable. A file that encodes behavioral guardrails and reaches 207,000 stars is evidence that reliability and controllability are as valued as raw capability in the developer community.

Addy Osmani’s agent-skills Pack: 90,100 Stars for 24 Engineering Workflow Skills

Addy Osmani’s agent-skills repository — 90,100 stars and growing — has become one of the most-installed agent skill packs of 2026 because it does something that generic skills do not: it encodes the workflows, quality gates, and review discipline that senior engineers already use, then packages them so AI agents follow those workflows consistently.

The pack contains 24 production-grade engineering skills mapping to the full development lifecycle: specification, planning, building, testing, reviewing, and shipping. Eight slash commands — /spec, /plan, /build, /test, /review, /webperf, /code-simplify, and /ship — map to phases of the development cycle, with each command activating the right skills automatically. The design choice of /build auto, which sequences through specification, planning, and building in one command, reduces the overhead of running a disciplined agent-assisted development process for teams that want the rigor without the manual step-by-step invocation.

The adoption pattern for agent-skills reflects a broader trend in the 2026 GitHub AI ecosystem: the fastest-growing repositories are not always the ones implementing new capabilities but the ones that make existing capabilities more useful in professional contexts. An engineering skill pack that encodes senior developer workflows is more immediately useful to most professional development teams than another experimental agent framework.

Langflow, Dify, and Flowise: Visual Agent Builders Dominating the Top Lists

Three of the top five most-starred AI agent repositories on GitHub as of April 2026 are visual builders: Langflow at 146,000 stars, Dify at 136,000 stars, and Flowise at 51,000 stars. This concentration at the top of the star rankings tells a story about who is building AI agents in 2026 — and it is not only software engineers.

Langflow provides a drag-and-drop interface for building LangChain-based agent pipelines. Each component in the agent — the prompt template, the memory store, the tool set, the output parser — is a visual node that can be connected and configured without writing code. Complex multi-step agent workflows can be assembled, tested, and deployed through the interface, and the underlying Python code is generated automatically from the visual design.

Dify extends this concept with a more complete platform approach that includes deployment infrastructure, API management, and team collaboration features alongside the visual builder. It supports multiple model providers, built-in RAG pipelines, and a workflow engine for complex agent orchestration. Dify’s 136,000 stars reflect adoption across a wide range of use cases, from internal tooling at technology companies to deployed products built on the Dify backend.

Flowise focuses specifically on visual LangChain flow builders with a React-based frontend. Like Langflow and Dify, it enables non-engineers to build working agent pipelines. Its 51,000 stars reflect a more focused community of users who need the specific combination of visual building and LangChain compatibility that Flowise provides.

The dominance of visual builders at the top of the AI agent star rankings mirrors what happened in web development when no-code and low-code tools democratized site creation. Domain experts in marketing, legal, operations, and other functions who understand their use cases deeply but cannot write agent code are using these tools to build agents that solve their specific problems. The barrier to entry for building production AI has never been lower.

Strix: Open-Source AI Penetration Testing That Validates Vulnerabilities Automatically

Not every trending AI agent repository is focused on productivity or workflow automation. Strix is an open-source AI penetration testing tool that behaves like a real security researcher rather than a static scanner — and it trended consistently through July and August 2026 with approximately 7,000 new stars per week, a rate that security tooling rarely sustains.

The weekly star growth rate is notable because it suggests genuine adoption among security teams rather than attention from a viral moment that fades. Security practitioners discovered Strix, found it useful, and kept recommending it to colleagues — the pattern of sustainable adoption that distinguishes a working tool from a novelty project.

Strix distinguishes itself from conventional penetration testing scanners in a critical way: it dynamically tests applications, validates vulnerabilities with proof-of-concept exploits, and includes features like an HTTP proxy for traffic interception. Static scanners report potential vulnerabilities based on code patterns or configuration. Strix confirms whether vulnerabilities are actually exploitable, which is the information security teams need to prioritize remediation work. An AI agent that can autonomously discover and validate exploits changes the economics of penetration testing significantly — a task that previously required specialized human expertise for every validation step can now be partially automated.

Strix’s rise in the August 2026 trending data is also connected to the cybersecurity threat environment that month — the same period that produced the UK AISI’s report on rogue agent incidents, the OWASP agentic AI security updates, and the Atlassian Rovo vulnerability disclosures. Security teams actively evaluating their exposure to AI-related threats were simultaneously reaching for AI-powered tools to assess and address those exposures.

Grok Build: xAI’s Open-Source Agent CLI Released in July 2026

Grok Build is xAI’s open-source coding agent CLI and terminal UI, released in July 2026 under the Apache 2.0 license. It powers the same agent loop behind Grok’s coding stack — meaning the architecture that xAI uses internally for its own AI-assisted development is now publicly available for inspection, compilation, and local execution.

The significance of Grok Build is not primarily its feature set, which is comparable to other coding agent CLIs at its release. The significance is transparency. By releasing the complete source code under Apache 2.0, xAI gave developers full visibility into how context handling, tool execution, plugins, skills, and MCP integration are implemented in a production coding agent. That level of reference implementation detail is rare in the AI tooling space, where the most capable systems are typically closed-source.

xAI does not accept external contributions to the main repository, which limits community involvement but maintains architectural control. Developers can study the implementation, compile and run it locally, and use the architecture as a reference for their own agent building. For the subset of developers who learn best by reading and running working code, Grok Build is one of the most educational AI agent repositories of 2026.

Vibe-Trading: Natural Language to Backtests and Live Trades

Vibe-Trading, built by the University of Hong Kong’s Data Science Lab, converts natural language prompts into backtests, alpha benchmarks, and optional live trades through supported brokers. It includes 452 pre-built alpha factors, point-in-time data handling to prevent look-ahead bias, and rigorous validation techniques that separate it from typical AI trading automation tools.

The project trended throughout July 2026 and illustrates a category of AI agent repository that is distinct from the developer tooling majority: domain-specific agent applications built for professional non-developer users. A quantitative researcher who can describe a trading strategy in natural language but who does not want to implement a backtesting framework from scratch represents the target user for Vibe-Trading.

The inclusion of point-in-time data handling — a feature that ensures historical simulations do not accidentally use data that would not have been available at the simulation date — reflects genuine domain expertise in quantitative finance baked into the project’s design. This kind of domain-specific sophistication is a distinguishing characteristic of the best specialized AI agent repositories: they encode professional knowledge into the agent’s workflow, not just general AI capability.

The ByteDance Agent Harness: No.1 GitHub Trending in February 2026

ByteDance’s agent harness repository reached the top position on GitHub Trending in February 2026 and accumulated more than 25,000 stars. Described in the awesome-ai-agents-2026 comprehensive list as offering durable coordination with crash recovery, human approval gates, kill switch functionality, and open protocol support through AXP, the ByteDance harness represents the enterprise-oriented end of the agent orchestration spectrum.

The emphasis on crash recovery and human approval gates in the ByteDance harness architecture reflects the production deployment requirements that distinguish enterprise agent infrastructure from research or developer tools. An agent that can recover from crashes mid-task and pause for human approval before taking high-impact actions is an agent that can be deployed in contexts where reliability and oversight matter — not just in development environments where failures are tolerable.

Key Trends Across the 2026 GitHub AI Ecosystem

Looking across the repositories that defined GitHub’s AI trending charts in 2026, several consistent patterns emerge that are worth understanding as strategic signals rather than individual data points.

Visual builders are democratizing agent development. The presence of Langflow, Dify, and Flowise at the top of the star rankings reflects genuine adoption by non-engineering users who are building and deploying agent applications without writing code. This is not a niche trend — it mirrors the trajectory of web development when no-code tools went mainstream.

Behavioral reliability is as valued as raw capability. The 207,000-star Karpathy CLAUDE.md file and Addy Osmani’s 90,100-star engineering skills pack both address agent behavior rather than agent capability. Developers who already have access to powerful models are voting with their stars for tools that make those models more predictable, controllable, and aligned with professional workflows.

Plugin architectures are winning over monoliths. DeepSeek Harness’s 200,000 stars in 15 days and the 11,944-repository plugin ecosystem that emerged within two weeks validated the market’s readiness for composable agent infrastructure. The projects that allow developers to customize components independently are growing faster than those that require forking the entire codebase to change one behavior.

Security tooling is catching up. Strix’s sustained 7,000 stars per week, arriving precisely during the period when the AI safety crisis of August 2026 was making headlines, suggests that security teams are actively building AI agent capability alongside their defensive security programs. AI-powered security tools and security tools for AI systems are both growing simultaneously.

The ecosystem is going global. The ByteDance harness, Vibe-Trading from the University of Hong Kong, DeepSeek Harness from China’s leading AI lab, and the Bharat Agentic AI Hackathon all reflect a geographical distribution of AI agent development that was less visible in prior years when the GitHub AI trending list was dominated by US-headquartered projects.

Conclusion: The Building Blocks of a New AI Infrastructure Stack

The repositories profiled in this article are more than trending GitHub projects. They are the building blocks of a new AI infrastructure stack that is being assembled in public, at unprecedented speed, by a developer community that has grown from a few thousand enthusiasts to millions of active contributors in under three years.

For developers deciding what to learn and build with, the repositories in this list represent the most battle-tested, community-validated starting points available. For enterprise teams evaluating which open-source tools to incorporate into their AI agent deployments, the star counts and fork rates provide a signal — imperfect but meaningful — about where the community has concentrated its attention and found genuine value.

With 4.3 million AI repositories on GitHub and the market growing at 46 percent per year, the pace of innovation shows no signs of slowing. The projects that matter most are not always the most technically sophisticated — they are the ones that solve real problems for real developers at the moment those developers are ready to build. The 2026 trending lists are a record of which projects passed that test.

Follow our site for ongoing coverage of GitHub AI repositories, open-source agent frameworks, and the evolving agentic AI ecosystem.