Briefing
Two years ago, if you wanted to build an AI agent, you had a few half-finished projects to choose from and a lot of guesswork. That has changed. By the middle of 2026, a developer starting fresh has hundreds of open-source options across frameworks, browser tools, memory layers, deployment runners and security scanners. The problem is no longer "is there a tool for this" but "which of the forty tools for this should I actually use".
That shift matters for any business team weighing whether to build on AI rather than just buy a subscription. The pieces are now mature enough to assemble into real internal systems, and most of them are free and open-licensed. The catch is that the landscape moves fast, the star counts and feature lists go stale within weeks, and a few of the loudest projects are louder than they are useful.
This is a map of where things stand. I've grouped the tools by what they actually do, kept the numbers the source material reported, and flagged where those numbers have already drifted. Treat the figures as a snapshot, not gospel, and click through to the repos before you commit.
Frameworks: The Foundation Layer
A framework is the scaffolding you build agent behaviour on top of. Here are the ones worth knowing.
OpenClaw (345,000 stars), The leading skills-based framework. It runs on Node.js under an MIT license, ships with 100+ built-in skills, and has its own ClawHub marketplace for community skills. Pick it if you work in JavaScript and want broad capability out of the box. Two caveats on the headline number: the live repo is closer to 379,000 stars now, and the "100+ skills" figure comes from the project's own marketing rather than an independent count (ClawHub itself hosts thousands of community-contributed skills).
Hermes Agent (reportedly 22,000 stars), Nous Research's learning agent. Python, MIT license, Honcho memory, 40+ tools, and a built-in learning loop. Good for Python developers building a personal assistant that improves over time. Be careful with the star count here: the article's 22,000 figure looks badly out of date, the live repo shows roughly 197,000 stars, nearly nine times higher. The 142-contributors figure couldn't be confirmed either.
AutoGen, Microsoft's multi-agent orchestration framework. Python, conversational agents, code execution, human-in-the-loop. The obvious choice for enterprise teams already living in the Microsoft stack.
CrewAI, The friendliest way into multi-agent work. Python, role-based agents, a clean API. Start here if multi-agent systems are new to you.
MetaGPT, Multi-agent software development teams in Python, coordinated through standard operating procedures. Built for code generation and software engineering tasks.
Visual Builders: No-Code and Low-Code
Langflow (146,000 stars), Drag-and-drop agent construction with 100+ components, and it exports to Python so you're not trapped in the GUI. Good for fast prototyping and for non-technical people who need to ship something. The star count is close to current sources (~147k, 148k); the component count is plausible but not independently confirmed.
Dify (136,000 stars), A full LLM app platform with visual orchestration, RAG, evaluation and deployment baked in. The better pick when you're heading to production rather than just trying ideas.
Web and Browser Tools
Firecrawl (130,000+ stars), A web context API that turns any website into clean Markdown. It's a top-100 GitHub repo globally, and the live count (~135k) backs up the figure.
Browser-use (reportedly 86,000 stars), Natural-language browser automation, for agents that need to click around real websites. The actual star count looks higher than 86,000, one 2026 source puts it past 97,000, so read that number as a floor.
Vercel agent-browser (reportedly 27,000 stars), Serverless browser automation for the Vercel ecosystem, written in Rust and built to run in Vercel Sandbox. The live repo is nearer 36,000 stars, so again the article's figure understates it.
Memory Systems
Mem0 (reportedly 52,000 stars), Standalone memory persistence with multi-layer storage. Model-agnostic, and the vendor claims sub-50ms retrieval. The star count is roughly right (a 2026 source says ~48,000), but treat the retrieval-speed claim as a marketing figure, not a benchmark.
Honcho, A dialectic memory system used by Hermes Agent. It tracks how an agent's knowledge changes over time and flags contradictions as they appear.
Local and Edge Deployment
LocalAI (44,000 stars), An OpenAI-compatible API for local models. It runs on CPU with no GPU required and supports a broad family of models. The live count (~47k) is close to the figure quoted.
Ollama, A developer-friendly local model runner. The CLI experience is the best in this category and it's well tuned for Mac.
Training and Education
nanochat (reportedly 55,000 stars), Karpathy's minimal LLM training stack. You can train a GPT-2-class model for about $48, which makes it the best hands-on way to learn how these models actually get built. Two notes: the live repo is closer to 42,900 stars (below the 55,000 quoted), and the project usually frames its headline cost as roughly $100, the $48 figure is the documented GPT-2-capability run (about two hours on 8x H100).
Security and Trust
Bumblebee (Perplexity), A supply-chain security scanner for AI projects, open-sourced by Perplexity in May 2026. It's a read-only scanner written in Go (Apache 2.0) that covers npm, PyPI, MCP configs, editor extensions and browser extensions, among others.
Developer Tools
awesome-claude-skills, 1,000+ production-ready skills for Claude Code, community-curated. Worth knowing this is a family of repos rather than one canonical list; the largest collection cited carries 1,200+ skills, and "quality-tested" is the maintainers' own framing.
LobeHub, A multi-agent chat UI with deep customisation and plugin support.
Pi Coding Agent, A Claude Code competitor (by Mario Zechner) with a minimalist take on agent-assisted development.
transitions.dev, Copy-paste CSS transitions for AI-generated UI, packaged as an installable agent skill. The article says 12 transitions; the current project actually lists eighteen, so that count is out of date.
developer-roadmap, Community-driven learning paths for AI and software development.
Emerging Areas
Agent Marketplaces: ClawHub for OpenClaw, Langflow's component library, and early standards for trading skills between projects.
Agent Standards: MCP (Model Context Protocol) is gaining ground as a universal tool interface.
Agent Safety: CVE databases, security audits and responsible-disclosure practices are starting to take shape around agents.
Agent Observability: Logging, monitoring and debugging tools built specifically for agent behaviour.
How to Choose
What you should reach for depends on what you're building:
- JavaScript developers: OpenClaw + Browser-use + Firecrawl
- Python developers: Hermes + CrewAI + Mem0
- Enterprise: AutoGen + Dify + LocalAI
- No-code: Langflow + Dify
- Education: nanochat + developer-roadmap
- Security-conscious: OpenClaw + LocalAI + Bumblebee
(One note: the source list named "OpenHuman" in that last row, but no project by that name exists in this category. It reads as a typo for OpenClaw, the framework named earlier, so I've used that here.)
The Bigger Picture
The thing that stands out about this landscape is how grown-up it has become. These aren't experimental toys anymore. They're production tools, used by real companies, and the star counts and contributor numbers point to genuine adoption rather than hype, even where the exact figures drift from week to week.
The ecosystem is also settling on shared standards. MCP is becoming the common tool interface, Mem0-style memory patterns are spreading across frameworks, and OpenAI API compatibility is now the default rather than a feature. That convergence makes it easier to mix tools together and harder to get locked into one vendor.
If you're a developer stepping into this space, the timing is good. The tools work, the docs are solid, and the communities will help you. Pick a framework, run the quickstart, and start building.