Heads up: this post contains affiliate links. If you buy through them we may earn a commission at no extra cost to you — it's how we fund independent research. Full disclosure.
GitHub stars are a noisy signal — popularity and quality do not always correlate — but velocity tells you where the community’s attention is going. This roundup covers the repos with the biggest star counts and fastest growth in the agent space as of July 2026. All star counts are sourced directly from the linked repositories; none are estimated.
The broader context: AgentMarketCap’s funding tracker reports $2.66 billion raised by agentic AI companies through April 2026 — 144% more capital than the same period in 2025, but 39% fewer transactions. The market is consolidating. Large capital rounds are going to high-conviction bets. The repos below reflect where that conviction is landing.
1. OpenClaw — 380,000+ stars
GitHub: openclaw-foundation/openclaw (formerly Warelay/Moltbot) Stars: 380,000+ as of June 2026; 79,600 forks Language: TypeScript and Swift
OpenClaw is a free, open-source autonomous agent that connects large language models to your operating system via messaging apps. It supports Signal, Telegram, Discord, WhatsApp, and seven other messaging platforms. You install it once, configure it with your LLM provider credentials, and then interact with it through any connected messaging app as if you were texting an assistant that has access to your computer.
The backstory: originally released in November 2025 as Warelay, then rebranded to OpenClaw on January 30, 2026 following a trademark dispute with Anthropic. The project is now governed by the independent OpenClaw Foundation after the original creator joined OpenAI in February 2026.
The 380,000 star count makes it the most-starred autonomous agent repo by a wide margin. The messaging-app interface lowers the barrier to entry significantly compared to API-first frameworks — you do not need to build a frontend. Its plugin marketplace, ClawHub, provides one-click install for community-built extensions.
Who should watch it: Anyone building autonomous agents that need to be accessible through existing communication channels rather than custom interfaces.
2. Hermes Agent — 210,000 stars
GitHub: NousResearch/hermes-agent Stars: 210,000 as of July 1, 2026; v0.18.0 “The Judgment Release” Maintainer: Nous Research
Hermes Agent is a self-improving agent from Nous Research with a closed learning loop: it writes reusable skills from completed tasks, building a personal knowledge base that compounds over time. The 90,000+ community skills in the Hermes Skills Hub are available to any Hermes instance.
The velocity is striking: Hermes grew from 40,000 to 188,000 stars between April and June 2026, a peak rate of 24,000 stars per week. NVIDIA selected Hermes as the reference runtime for Nemotron 3 Ultra, which accelerated enterprise attention significantly.
The “self-improving” framing is specific: Hermes does not update its own model weights. It builds a library of reusable SKILL.md-format skills from tasks it completes, so it gets progressively faster and more capable at recurring workflows without retraining. This is a different and more tractable form of agent improvement than model fine-tuning.
Who should watch it: Teams building agents that will handle repetitive workflows over months — the skill accumulation pattern compounds most visibly in high-volume, recurring use cases.
3. CrewAI — 55,000 stars
GitHub: crewAIInc/crewAI Stars: 55,000; v1.15.1 (June 27, 2026) Downloads: 5.2 million monthly pip installs
CrewAI’s role-based crew abstraction — define agents by role (researcher, writer, critic) and let the framework handle coordination — remains the fastest path from an idea to a working multi-agent prototype. The 5.2 million monthly download figure reflects how widely it is used in the “get it working first” phase of agent development.
The v1.15.1 release represents a mature, stable codebase for a framework that started as a rapid-prototyping tool. CrewAI has expanded to support more complex stateful workflows than its original design, though it still trails LangGraph for production stateful deployments.
Who should watch it: Every builder starting a new multi-agent project. Run your first version in CrewAI; decide whether to port it to LangGraph based on whether you need checkpoint recovery and audit trails.
4. LangGraph — 36,500 stars
GitHub: langchain-ai/langgraph Stars: 36,500; v1.2.7 (June 30, 2026) Ecosystem downloads: 47M+ monthly across LangChain
LangGraph is the production workhorse of the agent ecosystem. Klarna, Uber, and LinkedIn are among its documented production deployments. The June 30 v1.2.7 release maintains the project’s consistent release cadence.
The project’s design — explicit state transitions, graph topology, built-in checkpointing — means production teams can reason about failure modes in ways that role-based frameworks make harder.
Who should watch it: Anyone running agents in production environments where debugging, audit, and recovery from partial failures matter.
5. OpenAI Agents Python SDK — 27,700 stars
GitHub: openai/openai-agents-python Stars: 27,700 Key 2026 additions: sandboxed execution (April 2026), native voice/realtime agent support
The OpenAI Agents SDK’s provider-agnostic architecture — 100+ LLM support through a unified interface — distinguishes it from SDKs tied to a single model provider. The April 2026 sandboxed execution feature allows agents to run code safely without OS access risks. Native voice and realtime support makes it the only open-source framework with production-ready voice agent capabilities.
Who should watch it: Teams that need voice agents or want to switch LLM providers without rewriting agent logic.
6. Mastra — 25,800 stars
GitHub: mastra-ai/mastra Stars: 25,800; @mastra/core@1.48.0 (July 1, 2026) Weekly npm downloads: 300,000+ Funding: $22M Series A, Spark Capital, April 2026
Mastra is the TypeScript-native answer to the Python-dominant agent framework ecosystem. The 300,000 weekly npm download figure indicates significant adoption in Node.js production environments. The Observational Memory system, which the Mastra team documents as achieving 4–10x token savings over naive context management, is the technically distinctive feature.
Who should watch it: TypeScript and Node.js teams. If you are already in the JavaScript ecosystem, Mastra’s ergonomics are significantly better than using Python frameworks through any bridging approach.
7. AG2 / AutoGen — 4,700 stars (new repo; formerly 54,600 as microsoft/autogen)
GitHub: ag2ai/ag2 Stars: 4,700 on new repo (formerly microsoft/autogen with 54,600) Context: Independent governance since late 2024
AG2 is the continuation of Microsoft’s AutoGen project under independent governance. The star count on the new repo reflects the relatively recent rebrand rather than the maturity of the codebase. The multi-agent conversational patterns first documented in AutoGen underpin a significant portion of the research community’s current work on agent interaction.
Who should watch it: Researchers and teams building conversational multi-agent systems. The independent governance move improves the project’s long-term sustainability relative to a single-vendor-owned fork.
8. Awesome AI Agents 2026 — community index
GitHub: caramaschiHG/awesome-ai-agents-2026 Description: 300+ agent resources across 20+ categories, updated monthly
Not a framework, but worth bookmarking: the Awesome AI Agents 2026 curated list is the most comprehensive single-page index of agent resources currently maintained. Updated monthly, it covers frameworks, papers, deployment patterns, and community resources across 20+ categories.
| Repo | Stars | Category | Key fact | |
|---|---|---|---|---|
| OpenClaw | 380,000+ | Autonomous agent | OS access via messaging apps; ClawHub plugin marketplace | |
| Hermes Agent | 210,000 | Self-improving agent | 24k stars/week peak Apr-Jun 2026; NVIDIA reference runtime | |
| CrewAI | 55,000 | Multi-agent framework | 5.2M monthly pip downloads; fastest to prototype | |
| LangGraph | 36,500 | Workflow orchestration | Production at Klarna, Uber, LinkedIn; checkpointing | |
| OpenAI Agents SDK | 27,700 | Multi-LLM framework | Voice + realtime agents; 100+ LLM support | |
| Mastra | 25,800 | TypeScript framework | 300k weekly npm downloads; $22M Series A | |
| Pydantic AI | 16,000 | Typed Python agents | Structured output validation; model-agnostic | |
| AG2 / AutoGen | 4,700 (new) | Conversational multi-agent | Continuation of AutoGen; independent governance |
Common questions
What is the difference between Hermes Agent and OpenClaw?
Hermes Agent (Nous Research) is a self-improving agent with a skill accumulation system — it builds a personal library of reusable skills from tasks it completes. OpenClaw is an OS-connected autonomous agent you control via messaging apps. Hermes is oriented toward power users who want an agent that gets smarter over time; OpenClaw is oriented toward users who want an autonomous agent accessible from wherever they already communicate. Both are pre-built agents you configure and extend, not frameworks you use to build custom agents.
Is the funding tracker data meaningful for choosing a framework?
Funding is a signal but not a determinant. Mastra's $22M Series A is meaningful primarily because it ensures the project will have engineering resources for at least 18-24 months. It does not tell you whether Mastra is the right framework for your use case. The decision table in the platform guide covers that question more directly.
How often do these star counts change?
Significantly. Hermes went from 40,000 to 188,000 stars in roughly 8 weeks (April-June 2026). The counts in this roundup are sourced as of July 2026; they will be materially different in three months. For current counts, check the GitHub repos directly. AgentMarketCap and The Agent Report both track ecosystem metrics on a weekly cadence.