What AgentRank is.
AgentRank (agentrankhq.com) is a benchmark of what AI coding agents actually install, build with, and keep. We run the real agent CLIs against real repos, hundreds of sessions per category, and measure the four things that decide a developer tool's standing with agents: how often it gets named, installed, compiled, and kept.
Why it exists.
AI coding agents already decide a growing share of tool adoption. When a developer says "add auth" and the agent picks a provider, installs it, and ships it, that selection never touches a search results page. Existing visibility tools measure what AI assistants say. AgentRank measures what agents do: the install, the build result, and whether the integration survives the next refactor. Actions, not mentions.
Who pays us, and what that buys.
- Standing can't be bought.There is no sponsored placement, no paid listing, no way to pay for a better public number. A score only moves when a re-run of the same benchmark says it moved.
- What vendors buy is measurement, not rank.A private scorecard of their own funnel, the failing runs behind it, and re-tests that prove whether a fix worked. We never sell being recommended.
- The public numbers are the same for everyone.Public releases follow one published policy for every vendor, customer or not: leaders and aggregates with sample sizes and intervals, customer or not. Per-vendor detail stays private to that vendor, whether they pay us or ignore us.
- The detection can't flatter anyone.Outcomes are read deterministically from the repo: the dependency diff, the build exit code, the code that survived. Never an LLM grading an LLM, never a vendor-reported number.
How we measure.
Real agent CLIs, real repos, tasks worded the way developers word them, and a pre-specified classifier scoring every outcome. The methodology is public and versioned; the first citable release will include the complete release bundle. Check the numbers: the methodology. The public instruments built on it: the Agent Default Index and the Agent Default Report. Controlled runs are synthetic benchmarks, reported as an exploratory release with sample sizes shown.
Not to be confused with other projects that share the name: AgentRank (agentrankhq.com) is not a directory of AI agents, an AI-agent leaderboard, or an MCP-server ranking.