Stars lie: 669 open-source AI agent repos ranked by what they actually do

AgentoolRank — Thu, 01 Oct 2026 06:29:37 +0000

Disclosure: I build AgentoolRank, the directory these numbers come from. The data is refreshed daily from the GitHub API and the full tables are free to reuse (CC BY 4.0).

I kept picking AI agent frameworks the way most of us do: sort by stars, skim the README, pip install. Then, weeks later, I'd find the repo hadn't merged anything in months.

So I started tracking 669 open-source tools for building, running and evaluating AI agents — frameworks, coding agents, MCP servers, RAG and memory layers, eval tools — and ranking them by activity instead of stars. Here's what that looks like.

1. A third of the ecosystem has gone quiet

213 of 669 tools (32%) have had no commit on their default branch in six months or more. That includes 68 projects with 5,000+ stars:

Project	Stars	Last commit
MetaGPT	70.7k	Jan 2026
gpt-engineer	55.1k	Nov 2024
Grok-1	52.2k	Mar 2024
Quivr	39.6k	Jun 2025
FastChat	39.6k	Jun 2025
Langchain-Chatchat	38.7k	Nov 2025
AgentGPT	36.3k	Apr 2025

Stars are a record of past attention. They say very little about whether issues get answered or whether the thing works with this month's model APIs.

2. The fastest growers add 20k+ stars a month

Star pace over our daily snapshots, scaled to 30 days:

Project	Stars	+30 days
Skills	179k	+26.3k
LangChain	147k	+23.5k
ECC	270k	+22.5k
hermes-agent	250k	+20.9k
DeepSeek Harness	241k	+20.1k
MarkItDown	188k	+15.3k
Firecrawl	187k	+14.1k

3. Some repos commit hundreds of times a day

Commits on the default branch in the last 90 days (I double-checked the top ones against the GitHub REST API):

hermes-agent — 32,240
OpenHuman — 21,133
elizaOS — 19,953
DeepSeek Harness — 19,798
LiteLLM — 13,153

That's 150–360 commits a day. A good chunk of that is almost certainly agents committing to their own repos — which is its own interesting signal about where this ecosystem is heading.

4. Where the growth is, by category

Category	Tools	Stars added / 30 days
Agent frameworks	329	+380k
Tool integration & infrastructure	128	+165k
Memory & knowledge	127	+150k
Coding agents	42	+92k
Enterprise agent platforms	57	+75k
Observability & evaluation	70	+68k

MCP-first tools are now a category of their own: 29 projects, about 580k stars combined.

5. How the ranking works

Four signals, refreshed daily from the GitHub API:

Star pace — stars gained per 30 days, from daily snapshots (not the lifetime total).
Commits in the last 90 days on the default branch.
Releases in the last 6 months.
Issue response — median time to close recent issues.

They're combined into a percentile score. None of them is perfect on its own (commit counts reward bots; stars reward launch hype), which is why the site shows all four side by side, plus an alternatives page and head-to-head comparisons for every tool.

6. Query it from your own agent

If you use Claude, Cursor or any MCP client, you can point it at the data directly:

{
  "mcpServers": {
    "agentoolrank": { "url": "https://agentoolrank.com/api/mcp" }
  }
}

Tools: search_tools, get_tool, get_alternatives — and submit_tool if you want to list something you've built. There's also a plain JSON API; details at agentoolrank.com/agents.

The full, daily-updated tables are at agentoolrank.com/report. What would you weight more heavily than stars when picking a framework — release cadence, issue response, something else?