MCP is no longer a niche integration pattern for onchain agents — it's the largest single category in the builder library, and it isn't close. Of the 435 tools we're tracking this week, 141 are MCP servers. That's roughly a third of the entire stack, more than API/SDKs (55) and trading tools (36) combined.
Here's the state of the onchain agent economy, in the numbers we track — not the vibes.
The library, by category
MCP's dominance says something about where builder effort is actually going: less time writing bespoke integrations, more time exposing data and actions to an agent runtime through a standard interface. The rest of the library, ranked:
- ▸MCP — 141
- ▸API / SDK — 55
- ▸Trading Tool — 36
- ▸Wallet Infrastructure — 33
- ▸Data Tool — 33
- ▸Agent Framework — 30
- ▸Skill Repo — 30
- ▸Developer Tool — 22
- ▸DeFi Tool — 19
- ▸Security Tool — 15
- ▸Agent Marketplace — 7
- ▸Onchain Agent — 6
- ▸Agent Launchpad — 3
- ▸Research Tool — 2
- ▸Other — 2
- ▸Research Paper — 1
Notice how small "Onchain Agent" is next to everything above it. That's the taxonomy point we keep making: most of what gets built in this space is infrastructure *for* agents — frameworks, skills, wallets, data feeds — not deployed autonomous agents themselves. The stack is the product; the agent is what you build on top of it.
Where it's built
By chain coverage: Multichain tooling leads at 156, with Base (150) and Ethereum (118) close behind, then Solana (114), Arbitrum (60), Polygon (54), and BNB Chain (44). 40 tools are chain-unknown or chain-agnostic by design. Base and Solana sitting almost neck-and-neck with Ethereum is worth sitting with — this isn't a single-chain story.
Identity, onchain
ERC-8004 registered agents on Ethereum mainnet: 51,556 as of this week's check. That's a raw registry count, not an active-agent count — registration is cheap and some of those entries are dormant, duplicate, or experimental. What it tells you is that portable onchain identity for agents has moved from spec to habit: tens of thousands of entities have chosen to register rather than rely on a centralized allowlist.
Who's scoring well
The Sato Score measures how open, active, and verifiable a tool is — it is a transparency and liveness signal, not a safety or returns grade. This week's leaders:
1. Raydium — 95 (High) 2. OpenClaw — 92 (High) 3. Hermes Agent — 92 (High) 4. BlockRun MCP — 91 (High) 5. CoinGecko MCP — 89 (High) 6. Coinbase AgentKit — 88 (High)
All six clear the High tier (≥70) on transparent code, maintained repos, and documentation a builder can actually act on. None of that means "safe to point your agent's wallet at without checking the specifics" — read the methodology, not just the number.
What shipped this week
Six additions to the library since last week, several of them payment-rail infrastructure: Quant Flow (Trading Tool, added 2026-09-25), CoinOS Skills (AiCoin) (Skill Repo, 2026-09-25), 8004Swap (Trading Tool, 2026-09-24), Lightning pi (API/SDK, 2026-09-24), Aperture — Lightning Labs' L402 proxy (API/SDK, 2026-09-24), and Alby x402 Facilitator (API/SDK, 2026-09-24). Three of six landed the same day, all touching payment protocols (L402, x402) — a cluster worth watching rather than a coincidence to ignore.
Notable releases in the tracked repos: tari v6.0.1-pre.1, and two point releases of dexpaprika-mcp (v2.6.0 and v2.5.2) — an MCP server iterating fast enough to ship two versions in the same window. vultisig-sdk shipped v8.1.1. And outside the repos: CoinDesk reported Apollo's Torsten Slok warning that AI agents could drain cheap bank deposits — a reminder that the macro-finance conversation about autonomous agents moving money is already happening, whether or not the tooling underneath is ready for it.
What to watch
The MCP category's size relative to everything else suggests the next competitive edge isn't "who builds another framework" — it's who builds the sharpest, best-scored MCP servers on top of the frameworks that already exist. Watch the payment-rail cluster (L402, x402) for whether it's a trend or a blip next week.
Building your own onchain agent? Start with the stack, not a blank prompt — describe the agent at [satohub.ai/build](https://satohub.ai/build) and see what maps to it.
Sources
- ▸SatoHub Directory — https://satohub.ai/directory
- ▸Sato Score methodology — https://satohub.ai/sato-score
- ▸ERC-8004 verification — https://satohub.ai/resources/erc-8004