@naulon/wayfarer

The paying agent. Give it a topic and a budget; it discovers tolled articles, decides which are worth citing, pays only those, and answers from what it bought.

The point isn't automation for its own sake — it's that the agent makes a real budget decision at each step instead of paying for everything. It ranks candidates by relevance-per-dollar, buys greedily under the budget down to a relevance floor, reuses anything already paid for, and logs why for every candidate. The reasoning is the artifact.

Run

npm run -w @naulon/tollgate dev &                    # the gate to pay
npm run -w @naulon/wayfarer start -- "payment and passage" # the agent

decisions:
  [PAY]  the-naulon — relevance 1.00 @ $0.005; $0.095 left
  [SKIP] on-stillness — relevance 0.00 below floor 0.35
  ✓ paid $0.005000 for the-naulon

It runs offline against mock settlement. Set OPENAI_API_KEY for LLM appraisal and answer synthesis instead of the keyword heuristic; set PAYMENT_MODE=gateway with a funded buyer wallet to pay for real over Circle Gateway on Arc Network.

What's inside

• discover.ts / discovery.ts — find tollable candidates from a gate or feed.
• appraise.ts / decide.ts / allocation.ts — relevance, the pay/skip call, budget allocation.
• buyer.ts / gateway.ts — the x402 payment loop (mock and Circle Gateway).
• agent.ts / lib.ts — the end-to-end research run, also importable as a library.

Exposed as an MCP server by @naulon/wayfarer-mcp.

MIT.