No agent opens a webpage. It integrates one of three ways — a paid API call, an MCP tool, or a LangGraph node — all built and working. It pays per answer with x402 on Base: $0.02 a derivation, $0.00 for anything reused. Pick a surface.
client = x402Client() # pays per call, on Base r = await http.post(".../answer") # x402 settles the quote dossier = r.json()["dossier"] # -> the verdict, as JSON
Run the exact call an agent makes — a real POST /answer, live against the running engine. You see the JSON the agent gets back.
# FastMCP server "rederive" @mcp.tool() def get_derivation(node): ... # get_derivation("s_verdict") @mcp.tool() def recall(q, limit): ... # keyword search the memory
Point any MCP client at the server; the agent’s model calls these tools directly — no glue code.
graph = build_langgraph(engine)
# Rederive becomes the memory node inside the agent's own workflowEvery answer settles gaslessly in USDC via x402 (EIP-3009). Reused derivations cost nothing — you only pay for the cone that actually re-derives.
| Call | What runs | You pay |
|---|---|---|
| Cold full run | 24 derivations × $0.02 | $0.48 |
| Warm re-answer (unchanged) | 24/24 reused · 0 recomputed | $0.00 |
| Per derivation | one node re-derives | $0.02 |
| Reused derivation | same hash · served from memory | $0.00 |
Four real gasless settlements on Base Sepolia — the price the engine quotes is the price actually paid.