We pulled live GitHub star counts for 42 of the most significant open-source AI agents — every category from personal assistants to security agents — and ranked them. The full interactive leaderboard lives on our AI agent rankings page; this post is about what the numbers actually say.
Combined, these 42 projects hold over 2.3 million GitHub stars. Here’s where they’re concentrated, and the five things that surprised us.
The Top 10, as of June 2026
| Rank | Agent | Stars | Category |
|---|---|---|---|
| 1 | OpenClaw | 377,973 | Personal assistant |
| 2 | Hermes Agent | 189,778 | Personal assistant |
| 3 | AutoGPT | 184,874 | Autonomous |
| 4 | OpenCode | 172,626 | Coding agent |
| 5 | Gemini CLI | 105,139 | Coding agent |
| 6 | Browser Use | 98,119 | Browser automation |
| 7 | Codex CLI | 90,203 | Coding agent |
| 8 | OpenHands | 76,394 | Coding agent |
| 9 | MetaGPT | 68,690 | Framework |
| 10 | Learn Claude Code | 65,882 | Coding agent |
Live counts, filters, and all 42 agents: findpicked.com/agent/
Surprise #1: OpenClaw Dethroned AutoGPT — Fast
AutoGPT spent three years as the most-starred agent on GitHub. It took OpenClaw months to blow past it — 377k stars and counting for the lobster-themed personal assistant that runs on any OS and talks over any messaging platform. For perspective: that’s more stars than React had after a decade.
The signal isn’t lobster memes. It’s that the center of gravity moved from “agents that demo well” (AutoGPT’s 2023 cohort) to agents that live in your life — persistent, always-on, hooked into your actual messages, calendar, and machines. AutoGPT proved the appetite; OpenClaw shipped the meal.
Surprise #2: The #2 Agent Is From a Model Lab You Might Not Expect
Hermes Agent (189k stars) comes from Nous Research — a lab best known for fine-tuning open models, not consumer software. Its pitch, “the agent that grows with you,” leans on persistent memory and skills that compound over time.
Hermes and OpenClaw together hold a quarter of all stars in our dataset. The personal-assistant category — eight agents, ~778k combined stars — is now the largest single category by adoption, which would have sounded absurd eighteen months ago when “AI agent” meant “coding tool.”
Surprise #3: Coding Agents Are the Deepest Bench
Personal assistants have the two giants, but coding agents have the depth: eleven projects in our ranking with ~720k combined stars — OpenCode (172k), Gemini CLI (105k), Codex CLI (90k), OpenHands (76k), Cline (63k), Goose (49k), Aider (46k), and more. (We compared the big four head-to-head in Claude Code vs Codex vs Gemini CLI vs OpenCode.)
Two reads on this:
- No winner-take-all. Eleven credible projects means developers are still experimenting — and that model lock-in is the thing they’re experimenting against. The most-starred coding agent (OpenCode) is precisely the one that works with every model.
- Even Big Tech open-sources here. Google’s Gemini CLI and OpenAI’s Codex CLI are both open source. When the two biggest AI labs decide openness is table stakes for a category, that category’s culture is set.
The oddball in the top 10: Learn Claude Code (65k stars) isn’t a product at all — it’s a teaching repo that rebuilds a Claude Code-style agent harness from scratch. A tutorial outranking funded startups says a lot about how many developers want to understand agents from first principles.
Surprise #4: Security Agents Arrived
A year ago “AI security agent” meant a research paper. Now Strix — open-source AI hackers that find and validate vulnerabilities in your apps — sits at 26k stars, and CAI (Cybersecurity AI) holds 9k more. Offensive security automation went from taboo to GitHub-trending in about two release cycles.
It makes sense in hindsight: pentesting is expensive, repetitive, and tool-driven — exactly the work profile agents eat first. Expect this category to grow faster than any other on the board this year.
Surprise #5: Python Still Runs the Show — But TypeScript Owns the Consumer Layer
Across our 42 agents, Python dominates (frameworks, research, security — 27 of 42 projects), but look at what’s written in TypeScript: OpenClaw, OpenCode, Cherry Studio, Cline, AionUi, elizaOS — the user-facing layer. Rust shows up exactly where you’d expect (Codex CLI, Goose — performance-critical CLIs).
The pattern: Python for the brain, TypeScript for the body. Frameworks and reasoning live in Python’s ML ecosystem; the agents that win consumers ship polished cross-platform experiences in TypeScript.
What Stars Actually Measure (and Don’t)
Honesty section: GitHub stars measure developer attention, not usage, revenue, or quality. They lag (AutoGPT’s 184k includes its 2023 hype) and they spike on virality. We treat them as the best available public proxy for momentum in open source — nothing more.
That’s also why our rankings page pulls counts from the GitHub API rather than hardcoding marketing numbers, and why we note the refresh date on every update.
The Takeaway
Three predictions the data supports:
- Personal assistants are the new platform war. OpenClaw and Hermes are competing for the slot your phone’s assistant never earned.
- Coding agents commoditize toward model freedom. The most-starred one is model-agnostic; the vendor ones are open source. Lock-in is losing.
- The next breakout categories are security and computer-use. Strix’s trajectory and the browser automation cluster (98k stars for Browser Use alone) point where attention flows next.
Explore all 42 agents — filterable by category, searchable, ranked live — on the FindPicked agent leaderboard. And if you’re building with any of them, our MCP servers directory and Claude Code skills directory cover the tool ecosystem they all plug into.
Frequently Asked Questions
How did you choose the 42 agents?
Curated for significance: actively maintained, genuinely agentic (plans and executes multi-step tasks), and meaningful adoption. Star counts come from the GitHub API; the ranking order is purely by stars.
Why isn’t Claude Code on the list?
The list covers open-source agents, and Claude Code’s CLI isn’t open source. It’s still the reference point for the category — see our coding agent comparison.
How often do the rankings update?
The /agent/ page shows its last-refresh date in the header badge. Star data comes from the GitHub API, so refreshes are cheap and frequent.
What counts as an “AI agent” here?
Software that uses an LLM to plan and execute multi-step tasks with tools — not chatbots, not autocomplete, not model weights. Frameworks for building agents get their own category.