The biggest open-source AI agent projects are harnesses, not models
A daily GitHub ranking of 576 agent projects shows the wrapper around the model is where the activity is. Here is what the top 20 says about the ecosystem.
The short version
- The ranking covers 576 GitHub agent projects with more than 500 stars, refreshed daily by AI-Sovereign.
- The top three are ECC (308K stars), DeepSeek Harness (285K) and Hermes Agent (265K), all harnesses.
- Ponytail, Taste Skill and Archify are add-on skills for existing coding agents, not agents themselves.
- Of the 576 projects, 182 are mainly Python and 177 TypeScript; JavaScript has 45, Rust 39 and Go 35.
- 19 of the top 20 use MIT or Apache 2.0 licenses; Firecrawl is the AGPL exception.
Why it matters
If you are choosing or building agent tools, this is a quick map of what already exists and where the community's attention sits. It suggests the layer around the model, with its tools, memory and rules, now matters as much as the model. Stars measure attention, not quality or safety, so treat the list as a starting point for checking, not a recommendation.
What the ranking shows
As dev.to reported, the AI-Sovereign directory tracks 576 agent projects on GitHub that have more than 500 stars, and it refreshes the numbers every day. The list published for 22 November 2026 ranks the 20 biggest by stars. This is a snapshot of attention, not a news event, but it is a useful picture of where open-source agent work is concentrated.
At the top sit ECC with 308K stars, DeepSeek Harness with 285K and Hermes Agent with 265K. Behind them come Firecrawl (206K), Ponytail (196K), Graphify (152K) and LangChain (150K). The list runs down to DeepSeek's neighbor in spirit, ByteDance's DeerFlow, at 86K.
Why the harness is on top
A harness is the layer of software that wraps an AI model with tools it can use, memory, packaged skills and rules. The model produces the thinking; the harness decides what it can touch and what it remembers. Per the ranking's write-up, all of the top three are harnesses rather than models or finished apps.
ECC is described as a performance system for agent harnesses covering skills, instincts, memory and security. DeepSeek Harness is DeepSeek's own, built so everything is a plugin. Hermes Agent is Nous Research's personal agent that grows with its user.
The ranking's authors draw a conclusion from this: the model is becoming a replaceable part, and the harness determines how good an agent feels to use. That is their reading of the list, not a measured result, but it fits what the stars show.
Skills, eyes and memory
Three entries are not agents at all. Ponytail, Taste Skill and Archify are skills: packaged know-how you drop into a coding agent you already use to change how it behaves. Ponytail pushes agents to write less code. Taste Skill gives them design judgment, including references and a browser check. Archify turns ideas and code into interactive diagrams. All three rank in the top 15.
Another group exists to give agents what a model lacks on its own. A model does not know what happened today or what you told it last week. Firecrawl, Crawl4AI and Agent Reach supply web and social data. claude-mem keeps memory across sessions, and Graphify turns a codebase, documents and schemas into a knowledge graph you can query.
Other entries fill out the picture: LangChain as an agent platform, Browser Use for agents that drive a real browser, Orca for running a fleet of coding bots on your own subscriptions, and OmniRoute as a free gateway to many model providers. Google's Gemini CLI, a terminal agent, is also in the top 20 at 109K.
Languages and licenses
Across all 576 tracked projects, Python and TypeScript split the field. 182 are written mainly in Python and 177 in TypeScript. JavaScript follows with 45, then Rust with 39 and Go with 35. The only Go project in the top 20 is RAGFlow, a retrieval engine for feeding documents to agents.
Licensing is mostly permissive. 19 of the top 20 use MIT or Apache 2.0, which are licenses that allow wide reuse. Firecrawl is the notable exception, using AGPL. If you plan to build a product on top of a project, that difference is worth reading before you commit.
How to read a list like this
The method is simple: collect GitHub repositories tagged with agent topics that have more than 500 stars, rank by stars, refresh daily. Stars are a rough signal of interest. They do not tell you whether a project is well built, maintained or safe.
That caution matters more for agents than for most software, because an agent project often runs code or holds credentials. A popular project can still be a poor fit, and a small one can be the right tool.
Used well, the list is a pre-build check. Before writing your own agent framework, tool connector or workflow runner, see whether a large project already does the job and has real users and an issue history you can read.
What to watch
- Whether harnesses keep outranking models and apps as the daily numbers build up.
- Whether skill projects like Ponytail, Taste Skill and Archify keep climbing as a category.
- How the 24-hour and 7-day star changes on project pages look once more history builds up.
Try it
Pick one project from the top 20 that matches a task you have, then check its license, last commit date and open issues before installing it.
Sources and links
- Reportingdev.to: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by dev.to · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
