Watch live smart-money trades on Solana and paper-trade them in real time: open a virtual position when a tracked KOL buys, close when they sell, and track PnL from the market-cap multiple. This tutorial walks the open-source kol-copytrade-bot-starter — ~300 lines, one dependency, free API key — and shows how to filter by KOL win rate, deployer tier, and multi-KOL confirmation before risking anything real.
AI trading agents fail for a boring reason long before strategy matters: they can't see. An LLM reasoning about a memecoin with no live data hallucinates market caps; an agent that can only read prices misses everything that actually predicts outcomes on Solana — who deployed the token, which smart-money wallets are entering, whether the first twenty buyers are a known dump cluster.
MadeOnSol exposes that intelligence three different ways, built for three different kinds of agent. This guide is the decision map.
The decision table
| Integration | Best for | Auth | Setup time |
|---|
| MCP server | Claude Desktop, Cursor, any MCP client — conversational research and copilots | Free API key | ~2 minutes |
| solana-agent-kit plugin | Autonomous TypeScript/JS agents that act on a loop | Free API key | ~5 minutes |
| x402 micropayments | Agents with their own wallet, no signup at all | The agent's wallet | zero — pay per call |
All three sit on the same data layer: 1,000+ labeled KOL wallets with trades indexed at swap-time market cap, 23,000+ scored Pump.fun deployers, coordination detection, transparent 0–100 rug-risk scoring, and an MEV-stripped all-DEX firehose.
One line before the paths, because it defines the architecture: MadeOnSol is data-only. It never executes trades, never holds funds, never takes custody. Your agent decides; the API informs. That separation is a feature — the component with the private key and the component with the market view should not be the same vendor.
Path 1 — MCP: give Claude or Cursor live market vision
The mcp-server-madeonsol server (listed on Smithery as madeonsol/solana-kol-intelligence) turns every endpoint into an MCP tool that Claude Desktop, Cursor, or any MCP-compatible client can call mid-conversation. Ask "is this mint safe?" and the model calls madeonsol_token_risk, gets the ten-factor breakdown, and reasons over it — with real numbers instead of vibes.
Tools cover the full surface: token risk (single and 50-mint batch), bundle-cohort holdings ("how much does the launch bundle still hold?"), almost-bonded discovery ranked by bonding velocity, KOL feed and leaderboard, deployer alerts with reputation history, wallet PnL, and signal-performance stats so the model can check how reliable a signal is before trusting it.
Setup is a one-liner via Smithery or a JSON block in your client config with a free API key (200 requests/day, every endpoint). This is the right path when a human is in the loop — research, screening, "explain this token to me" workflows.
Path 2 — agent-kit: autonomous agents in TypeScript
For agents that run unattended, the solana-agent-kit-plugin-madeonsol plugin drops the same capabilities into Solana Agent Kit as typed methods and LLM-callable actions: tokenRisk(), tokenRiskBatch(), walletPnl(), almostBonded(), tokenBundle(), deployerHistory(), kolConsensus() and the rest.
The patterns that work in production:
- Gate before acting — whatever the agent wants to buy,
tokenRiskBatch() scores up to 50 candidates in one call (one request against quota) and the explainable factors[] let your logs say why something was skipped.
- Self-throttle —
GET /me returns tier, quota state, and remaining requests, so the agent can adapt its cadence instead of hitting 429s.
- No look-ahead in backtests —
deployerHistory() returns a deployer's reputation as of each day, so "was this deployer elite when it launched?" is answerable honestly.
Start with the free key; an autonomous loop that proves itself graduates to PRO for the real-time WebSocket channels and batch endpoints.
Path 3 — x402: the agent's wallet is the API key
The strangest and most agent-native path: 18 endpoints under /api/x402/* require no account, no key, no signup. The agent hits the endpoint, receives an HTTP 402 challenge describing the price ($0.005–$0.02 per call in USDC on Solana), pays from its own wallet via the x402 protocol, and gets the data — settled in about a second, transaction fees sponsored by the facilitator, so the wallet needs only USDC.
This matters for a class of agent that's arriving fast: autonomous agents that manage their own budgets and can't fill in signup forms. Five dollars of USDC buys a thousand calls at the cheapest tier. Discovery is machine-readable too — GET /api/x402 lists every endpoint, price, and input schema, so an agent can find, understand, and pay for the data without a human touching anything. Full details on the x402 page.
Which one is yours?
Human-in-the-loop research → MCP. Autonomous JS agent with a subscription → agent-kit. Wallet-native agent, zero-signup, pay-as-you-go → x402. They compose, too: teams commonly prototype in Claude via MCP, ship the loop on agent-kit, and let their users' agents pay via x402.
If the agent becomes a product other people use, two follow-ups: the Telegram-bot SaaS data-layer guide covers per-user quota math, and the covers what happens when your customers start seeing the data.
