Every list like this one gets published as a ranking and read as a recommendation. This one is a snapshot with the method attached, so you can check it, disagree with it, or run it yourself tomorrow and get a different answer.
The list
| Repository | Stars | Created | License | What it is |
|---|---|---|---|---|
| affaan-m/ECC | 234,017 | January 18, 2026 | MIT | Harness that adds skills, memory and security around a coding agent |
| NousResearch/hermes-agent | 221,325 | July 22, 2025 | MIT | Agent runtime built to accumulate context about its user over time |
| firecrawl/firecrawl | 156,920 | April 15, 2024 | AGPL-3.0 | Turns websites into clean text an agent can actually read |
| langchain-ai/langchain | 142,704 | October 17, 2022 | MIT | The framework most agent tutorials still start from |
| browser-use/browser-use | 107,004 | October 31, 2024 | MIT | Lets an agent drive a real browser and click through a site |
| google-gemini/gemini-cli | 106,209 | April 17, 2025 | Apache-2.0 | Google's own terminal client for its models |
| Graphify-Labs/graphify | 97,006 | April 3, 2026 | Apache-2.0 | Turns a codebase, its docs and its schemas into a queryable graph, no vector store |
| DietrichGebert/ponytail | 90,275 | June 12, 2026 | MIT | Pushes an agent to write less code and reuse what exists |
| thedotmack/claude-mem | 88,725 | August 31, 2025 | Apache-2.0 | Captures what an agent did in a session and feeds it back into the next one |
| infiniflow/ragflow | 86,153 | December 12, 2023 | Apache-2.0 | Retrieval engine that assembles the context a model sees |
| nexu-io/open-design | 81,922 | April 28, 2026 | Apache-2.0 | Local desktop app that turns a coding agent into a design tool |
| bytedance/deer-flow | 77,966 | May 7, 2025 | MIT | Long-horizon harness with sandboxes, memory and subagents for multi-hour tasks |
Four of them did not exist in January
ECC was created on January 18 of this year. Graphify on April 3. Open-design on April 28. Ponytail on June 12. All four are now above 80,000 stars, and all four are less than seven months old.
That is a third of the list. If you made an architecture decision earlier this year on the assumption that the tooling layer would settle down, it did not settle down. “We standardized on X” currently has a half-life measured in months.
The practical response is not to guess which one wins. It is to decide, in advance, which parts of your stack are allowed to churn. Anything that touches your own data model gets built and owned. Anything that is a wrapper around somebody else's API gets borrowed, with the assumption that it will be replaced. The second list is allowed to move. The first one is not.
The four that name their agents all support more than one
Four of the twelve say out loud, in their own description, which agents they support. Not one of them names only one.
claude-mem: “Works with Claude Code, OpenClaw, Codex, Gemini, Hermes, Copilot, OpenCode + More” open-design: “20+ CLIs via BYOK” graphify: “Claude Code, Cursor, Codex, and Gemini CLI” ECC: “for Claude Code, Codex, Opencode, Cursor and beyond”
The teams behind these four projects are close to the problem and are refusing to pick a single winner. They build the layer and treat the agent underneath as a slot you swap. That is a useful signal, not proof that the other eight projects made the same architectural choice.
The practical lesson from this four-project subset is to own the layer that survives the pick: your data, your context, and the record of what your company already decided and already tried. The agent underneath can change without taking that knowledge with it.
What this list does not tell you
It is not a growth chart. These are star counts from one morning. A single reading cannot tell you what is climbing and what peaked in April. Anyone showing you “+40,000 stars this week” from a snapshot like this one is inventing the delta.
A star is not adoption. Starring costs one click and no commitment. It measures attention, not whether anything here runs in production at a company that depends on it.
A description is not a benchmark. When a README lists seven supported agents, that is a statement of intent. It is not evidence that the tool works equally well on all seven.
The query shaped the answer. Filtering on the ai-agents topic means projects that never applied that label are missing, no matter how big. This is the most-starred set inside one tag, which is not the same thing as the most important set.
Written while building RogerTeam, where people and agents work in the same room. If this is the kind of thing you want in your feed, follow Hugo Venda on LinkedIn →