The biggest number in this week's data is not ours. The ERC-8004 registry on Ethereum mainnet showed 56,540 registered agents when we checked on 2026-11-16. Meanwhile, the SatoHub library tracks 496 resources you can build agents with. Those two numbers measure different things, and the gap between them is the story: a lot of agents have an onchain identity, and a much smaller, curated set of tooling is what they get built from.
A registration is a claim of identity. It says an address registered an agent. It does not say the agent works, is maintained, or does anything useful. Counting registrations is easy. Counting what is live and checkable is the hard part, which is why we keep the two numbers separate.
The builder library: 496 tracked
Here is the category mix, straight from the directory:
- ▸MCP: 150, by far the largest category
- ▸API / SDK: 48
- ▸Trading Tool: 40
- ▸Developer Tool: 39
- ▸Data Tool: 36
- ▸Wallet Infrastructure: 35
- ▸Skill Repo: 32
- ▸Agent Framework: 31
- ▸Payment Rail: 24
- ▸DeFi Tool: 20
- ▸Security Tool: 17
- ▸Agent Marketplace: 8, Onchain Agent: 6, Agent Launchpad: 5
- ▸Research Tool (2), Other (2) and Research Paper (1) round it out
The take: MCP servers are the load-bearing layer of the agent stack right now. 150 of 496 is the biggest block by a wide margin. Wallet infrastructure (35) and payment rails (24) are smaller, but they are the pieces that let an agent do anything with money. Notice the other end, too: only 6 entries are classed as onchain agents themselves. Most of what the library holds is what you build *with*, not the finished agents. That matches how we frame it: a framework is not an agent, and a launchpad is not an agent.
Chains: Base leads the tag count
By chain tag: Base 176, Multichain 166, Ethereum 128, Solana 122, Arbitrum 64, Polygon 61, BNB Chain 47, and 46 with no chain recorded yet.
Do not add these up. A listing can carry more than one chain, so the tags overlap and the total is not the 496. Read each figure as "how many tracked resources mention this chain," not as a market share. Base leading the count says something about where builders are shipping tooling. It says nothing about where value sits.
Sato Score leaders
The top of the Sato Score this week:
| Resource | Sato Score | Tier | |---|---|---| | Raydium | 95 | High | | OpenClaw | 92 | High | | Hermes Agent | 92 | High | | NanoParse | 92 | High | | Coinbase AgentKit | 91 | High | | BlockRun MCP | 91 | High |
A reminder, because this is where people over-read: the Sato Score is a transparency and liveness score. It reflects what is open, active and checkable about a project. It is not a safety grade, not a quality grade, and not a returns signal. A high score means there is a lot of evidence to look at. It does not mean you should trust the thing with funds. Read the evidence itself before you wire anything up. Methodology here.
Newest in the library
Six recent additions, by date added:
- ▸SellTape MCP (MCP), 2026-11-16
- ▸Siebly Hyperliquid JavaScript SDK (API / SDK), 2026-11-12
- ▸Siebly Agent Skills (Skill Repo), 2026-11-12
- ▸402registry, 402cron and 402post (all MCP), 2026-11-09
Three of the six share a 402 prefix and landed on the same day, which reads like a family of tools built around the HTTP 402 payment pattern. We are inferring that from the names, so check each one's own docs before you assume what it does.
Moves worth a look this week
These are release notes we picked up. We are listing them, not endorsing them, and we have not reviewed what changed in any of them:
- ▸openclaw 2026.11.4, a new release from the project that also sits near the top of the score table
- ▸dstack v0.6.5 and a matching verifier release, from Dstack-TEE/dstack
- ▸opensea-skill v2.32.0, from ProjectOpenSea
- ▸DefiLlama-Adapters published a new
tvlModulesrelease
If you depend on any of these, read the changelog before you bump a version. A release tag tells you something shipped. It does not tell you the change fits your agent.
What to watch
- ▸Registrations versus real activity. 56,540 is a registry count. The interesting number is how many of those agents show live, checkable behavior. We will keep these figures apart until the data can join them honestly.
- ▸MCP concentration. If MCP stays the biggest category, the quality of tool descriptions and the clarity of permissions become the thing that separates a usable agent from a demo.
- ▸Payment rails catching up. 24 rail entries against 150 MCPs means agents can find tools faster than they can pay for them.
Build yours
The prompt is the cute part. The stack is where the work lives. Describe the agent you want, and Sato maps the skills, MCPs, wallets and payment rails from this same library. Build your own crypto agent at satohub.ai/build and join the agent economy.