Web research agent (Steel + PRXY)

A minimal pattern for agents that read the live web without blowing the context window.

Architecture

Agent (Mastra / Anthropic SDK / curl)
    │
    ▼
api.prxy.monster  ──steel-browser──▶  Steel scrape API
    │                                      │
    ├─ structured-crusher ◀── tool_result ─┘
    ├─ ipc
    └─ patterns
    │
    ▼
Anthropic / OpenAI (BYOK)

1. Keys and pipeline

export ANTHROPIC_BASE_URL=https://api.prxy.monster
export ANTHROPIC_API_KEY=prxy_live_xxx
export STEEL_API_KEY=steel_xxx   # gateway env (cloud or local)
 
export PRXY_PIPE=steel-browser,mcp-optimizer,structured-crusher,code-crusher,ipc,patterns

In lair, pick the Web research (Steel) preset.

2. Single-shot research (curl)

curl https://api.prxy.monster/v1/messages \
  -H "Authorization: Bearer $ANTHROPIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-4-6",
    "max_tokens": 2048,
    "messages": [{
      "role": "user",
      "content": "Read https://docs.steel.dev/ and list three things Steel does that PRXY does not."
    }]
  }'

PRXY injects steel_scrape, fulfills it, auto-loops, returns the final answer. Check the receipt for steel-browser.scrapes and crusher savings.

3. Mastra agent

import { Agent } from '@mastra/core';
import { anthropic } from '@ai-sdk/anthropic';
 
export const researcher = new Agent({
  name: 'researcher',
  instructions: 'Use steel_scrape when you need live web content. Cite URLs.',
  model: anthropic('claude-sonnet-4-6'),
});

Env vars from Steel integration. PRXY handles tools — do not register a duplicate scrape tool in Mastra unless you bypass steel-browser.

4. When to add Steel SDK

Use steel-browser for read-only scrape. Use the Steel SDK when you need:

Playwright/Puppeteer sessions
Form fill, login, CAPTCHA
Computer-use screenshots