OnchainAI

Crypto tools, unified. Discover, vet, and install MCP, CLI, SDK, API, x402, RWA, and AI-agent tools — from the web or straight from your agent.

Live: onchain-ai.xyz · Connect hub: onchain-ai.xyz/connect

Why?

Crypto tooling is fragmented across CryptoSkill, Smithery, npm, GitHub topics, and dozens of separate registries — and most of it is hard to judge: is this MCP server official? Is the install command safe? Does it charge x402 payments on every call?

OnchainAI auto-discovers tools from multiple sources, normalizes them into one 3-axis taxonomy (Function × Asset Class × Actor), scores trust and install risk, and exposes everything to both humans (web UI) and agents (MCP server, plugin, skill) — so "find me a safe tool to bridge USDC to Base" is one question, not an afternoon of research.

Use it from your agent (60 seconds)

Default MCP is a no-auth directory endpoint (discovery/metadata — not wallet custody). Free discovery on public /mcp (search, detail, compare, install guides, categories, …). Optional OnchainAI-owned premium tools may return HTTP 402 and settle via x402 ($0.01 USDC for export_toolkit / recommend_verified_tool / gap_audit; ~$0.001 USDC for check_endpoint_health). Website browse stays free.

https://www.onchain-ai.xyz/mcp

OKX marketplace integrators only: paid package path https://www.onchain-ai.xyz/mcp/okx (~$0.1 every tools/call when the OKX gate is active). Coding agents and the Claude plugin must use /mcp, not /mcp/okx.

{
  "mcpServers": {
    "onchainai": { "type": "http", "url": "https://www.onchain-ai.xyz/mcp" }
  }
}

Full per-client walkthroughs (Codex, Windsurf, Gemini CLI, …): docs/CONNECT.md or the live /connect page.

MCP tools (public POST /mcp)

Linking your account from a coding agent (/connect#agent-sync) unlocks three more (account link ≠ payment): save_to_toolkit, save_stack_to_blueprint, and link_status.

Full hybrid table (incl. /mcp/okx): docs/CONNECT.md.

Claude Code plugin

One command wires up the MCP server, a /find-tool command, and a crypto-tools skill:

/plugin marketplace add Coinyak/onchainai
/plugin install onchainai@onchainai

The bundle lives in plugin/onchainai/ — it connects only OnchainAI's own read-only endpoint, never auto-runs install commands, and always discloses x402 pricing before recommending a paid tool.

Agent skill

plugin/onchainai/skills/onchainai-crypto-tools/SKILL.md teaches an agent when to search the directory, how to rank trust signals, and the hard safety rules (never install critical-risk tools, always surface x402 costs). It ships with the plugin; you can also copy the skill directory into ~/.claude/skills/ or upload it to any agent runtime that supports Agent Skills.

Features

• Auto-discovery crawler — CryptoSkill, GitHub topics, npm, web3-mcp-hub, MCP registry, on cron schedules with dedup + relevance scoring
• 3-axis classification — Function × Asset Class × Actor, plus chain tagging
• Trust & safety pipeline — trust scores, identity-cluster checks (repo/npm/homepage), install-risk analysis with a hard block on critical commands, operator review queue, quarantine
• x402 awareness — paid tools carry price metadata, verification flags, and referral/attribution disclosure (policy below)
• MCP server for agents — the read-only tools above, rate-limited and sanitized
• Claude Code plugin + skill — one-command onboarding for agent users
• Auth — GitHub OAuth (primary); email magic link where configured
• Community layer — submissions, comments, upvotes, bookmarks, toolkit, compare, blueprints
• Admin dashboard — tool review, categories, crawler control, featured carousel, users, site settings

x402 & referral policy

OnchainAI is a tool directory MCP (discovery/metadata), not a custody wallet:

• Third-party x402 in the catalog is attribution and trust metadata only — we publish price/endpoint flags (payment_verified, x402_endpoint_verified, price_verified) and never proxy those payments.
• We record anonymous referral/attribution events (views, install-guide fetches) to support revenue-share agreements with tool owners.
• We never hold user funds, act as a third-party payment gateway, or invent undocumented referrer/split payment fields.
• OnchainAI-owned premium MCP tools (export_toolkit, recommend_verified_tool, gap_audit, check_endpoint_health) may settle x402 to our payee wallet when called — that is selling our own service, not custodying others.
• Unverified third-party x402 tools remain visible when they pass the normal public quality gate — verification is a badge, not a hiding mechanism.

Details: docs/X402_REFERRAL_SPEC.md, hybrid connect: docs/CONNECT.md.

Architecture

┌─ Vercel ──────────────┐      ┌─ Railway ─────────────────────────┐
│ Next.js frontend      │──────│ Rust binary (Axum)                │
│ /connect, /tools, ... │ /api │  ├─ REST API (/api/v2)            │
│ proxies /api /auth    │ /mcp │  ├─ MCP server (POST /mcp)        │
│ /mcp to the API       │      │  ├─ Auth (GitHub OAuth + JWT)     │
└───────────────────────┘      │  └─ Crawler (tokio-cron)          │
                               └────────────┬──────────────────────┘
                                            │ sqlx
                               ┌─ Supabase Postgres (RLS) ─┐
                               └───────────────────────────┘

• API/MCP: Rust — Axum + sqlx + tokio-cron-scheduler, one binary (cargo run --features ssr)
• Frontend: Next.js (App Router) on Vercel, rewrites /api, /auth, /mcp to the Railway API
• DB: Supabase Postgres with RLS; the public read gate is enforced in both SQL policies and server queries

Run it yourself

git clone https://github.com/Coinyak/onchainai
cd onchainai
cp .env.example .env        # fill in Supabase URL/keys, GitHub OAuth, JWT secret

# API (port 3000)
cargo run --features ssr

# Frontend (port 3001, proxying to the API)
cd frontend && npm ci && API_PROXY_TARGET=http://localhost:3000 npm run dev -- --port 3001

The crate's default feature set is empty so server deps don't leak into any WASM/client build. Plain cargo commands that touch server code need --features ssr.

Common checks:

cargo test --features ssr                      # tests
cargo clippy --features ssr -- -W clippy::all  # lint
cargo fmt --check                              # format
./scripts/agent-harness-check.sh               # agent/dev harness self-check

Repository layout