Build a Self-Paying Research Agent in 20 Lines

By the end of this tutorial you’ll have a research agent that:

• Searches the web via SerpAPI
• Scrapes top results via Firecrawl
• Synthesizes a cited answer via GPT-4o-mini
• Pays for every single call out of its own USDC wallet on Base
• Costs you ~$0.04 per question

No SerpAPI account. No Firecrawl subscription. No manual API key juggling.

Prerequisites

bun add @axon/client @axon/vercel-ai @ai-sdk/openai ai zod

You need two environment variables:

• AXON_KEY — sign up for free $0.50 credit to get one
• OPENAI_API_KEY — for the LLM driving the agent

The whole agent

import { Axon } from '@axon/client';
import { axonTool } from '@axon/vercel-ai';
import { generateText } from 'ai';
import { openai } from '@ai-sdk/openai';
import { z } from 'zod';

const axon = new Axon({ apiKey: process.env.AXON_KEY! });

const search = axonTool(axon, 'serpapi', 'search', {
  description: 'Search the web.',
  parameters: z.object({ q: z.string() }),
  via: 'params',
});

const scrape = axonTool(axon, 'firecrawl', 'scrape', {
  description: 'Scrape a URL, return markdown.',
  parameters: z.object({ url: z.string().url() }),
  via: 'body',
});

const { text } = await generateText({
  model: openai('gpt-4o-mini'),
  tools: { search, scrape },
  maxSteps: 8,
  system: 'Search, scrape 2-3 top links, synthesize with source URLs.',
  prompt: process.argv.slice(2).join(' '),
});

console.log(text);

That’s the entire agent. Three moving parts.

What just happened

1. Axon wraps paid APIs. axonTool(axon, 'serpapi', 'search', ...) creates a Vercel AI SDK tool that, when called, hits Axon → SerpAPI → back. You never see a SerpAPI key. The cost is debited from your Axon wallet in USDC.

2. The LLM picks which tool to call and when. generateText with tools: and maxSteps: 8 lets the LLM decide: call search, look at results, call scrape on the most promising, summarize. This is the ReAct pattern you already know from LangChain — but with no local agent loop code.

3. The wallet makes it autonomous. Every call debits USDC atomically. If the agent tries to spend more than you deposited, Axon returns 402 insufficient_funds and the agent stops. You set a budget, the agent works within it.

Running it

AXON_KEY=ax_live_... OPENAI_API_KEY=sk-... \
  bun agent.ts "what are the top espresso bars in lisbon?"

Expected output:

1. Copenhagen Coffee Lab — reliably excellent beans, …
2. Hello, Kristof — minimalist space, high-end extraction, …
3. Fábrica Coffee Roasters — single-origin specialists, …

Sources:
- https://timeout.com/lisbon/coffee-spots
- https://eater.com/2025/lisbon-coffee-guide
- …

What you actually paid

Axon returns x-axon-cost-usdc on every call. Typical run:

• 1× search = $0.0055
• 3× scrape = $0.0165
• 4× LLM = $0.0080 (gpt-4o-mini is cheap)
• Total: ~$0.03

Run the same question an hour later → SerpAPI result comes from cache at 50% off, most scrapes from cache too. Second run: ~$0.015.

Compare:

• SerpAPI minimum plan: $50/mo
• Firecrawl minimum plan: $16/mo
• Combined monthly minimum: $66/mo

To break even on the subscription model you’d need to run ~2,200 of these agents per month. Most indie hackers don’t.

Safety rails

Two things stop the agent from burning your wallet by default, one upgrade for paranoid builders:

1. Your wallet balance is the hard ceiling. An agent with a $5 wallet cannot spend $6. Full stop.
2. Every response returns x-axon-cost-usdc. You can log it, alert on it, write circuit breakers.
3. (Pro tier) Policy engine: max_request_cost_micro, daily_budget_micro, per-API caps.

Going further

• Swap gpt-4o-mini for claude-sonnet-4-6 via @ai-sdk/anthropic — Axon bills both through the same wallet
• Add axon.call('voyage', 'embeddings', ...) to build a RAG layer on top
• Cache the whole agent loop: same question in 5 minutes = 80% cache hit = ~$0.005

The full template

Cloneable repo at axondevi/axon → templates/research-agent-ts. Runs out of the box.

If you build something on this, we’d love to see it — DM @axondevia.