About AgentCrush

AgentCrush is protocol-neutral market intelligence for the agent economy.

AI agents are no longer just isolated tools that answer a prompt and disappear. Increasingly, they are persistent systems — connected to tools, memory, and execution environments. They expose endpoints, accept payments, join marketplaces, appear in multiple registries, and interact with other agents and services across the internet.

That shift is happening across several overlapping ecosystems at once: open-source GitHub projects, x402-enabled services, ERC-8004 on-chain registrations, A2A and MCP activity, agent marketplaces, and more. There is no single authoritative index. Each protocol sees part of the picture.

AgentCrush tracks across all of them.

What we track

For each indexed agent, AgentCrush collects evidence signals over time:

• GitHub repository activity — commits, releases, stars, forks
• Package usage — npm and PyPI download trends
• Ecosystem relationships — dependencies, integrations, co-mentions
• Discourse signals — Hacker News mentions, community threads
• Documentation quality — coverage, maintenance indicators
• Registry context — x402 payment endpoint declarations, ERC-8004 on-chain registrations
• Historical snapshots — daily rank and score history for trend analysis

Agents that accumulate enough verifiable evidence can appear in evidence-ranked leaderboards. All other indexed agents remain discoverable through the Explore index, sorted evidence-first.

Machine-readable intelligence

AgentCrush also exposes its data through machine-readable APIs and x402 micropayment endpoints. Agents, developers, and automated workflows can query trust summaries, ranking history, and verification status — without going through a web interface.

The x402 endpoints follow the open x402 payment protocol on Base mainnet, meaning any compatible agent or client can call them programmatically and pay per request in USDC.

What we are not trying to do

AgentCrush is not trying to declare one protocol the winner, or to position itself as the identity or reputation layer for any specific ecosystem. Rankings and scores are evidence-based signals, not endorsements. Labels such as evidence-ranked, ERC-8004 registered, or machine-callable are contextual descriptors — they reflect what the data shows, not a guarantee of quality, safety, or performance.

The goal is simpler: make the agent economy easier to observe, compare, and understand as it develops. There is a lot happening across a lot of ecosystems simultaneously. AgentCrush is a place to watch it clearly.

Background

AgentCrush was started in early 2026 as an experiment in tracking the emerging AI agent ecosystem. It began as a lightweight rankings and discovery surface and has grown into a multi-signal intelligence index. The project is actively developed and the methodology evolves as the agent economy itself evolves.

As of May 2026, AgentCrush runs four parallel scoring methodologies — one for each major category of agent: model families, tokenized agents, service agents, and developer agents. The full methodology is published at /methodology — every weight, every formula, every limitation. We also expose the index via an MCP server so AI clients can query AgentCrush as a live data layer.

For questions or to submit an agent, use the submission form or reach us at contact@agentcrush.xyz.