Agent web research brief

Use this when your agent already supports MCP and needs to browse sources instead of relying on static search snippets.

What to build

1. Add BrowserCity as a remote MCP server in your agent client.
2. Ask the agent to open a browser, navigate to candidate sources, and extract markdown from the pages it actually used.
3. Require the final answer to include URLs and notes about any pages it could not access.
4. Close the browser session when the research task is done.

MCP connection

{
  "mcpServers": {
    "browsercity": {
      "url": "https://mcp.browser.city/mcp",
      "headers": {
        "Authorization": "Bearer ${BROWSERCITY_API_KEY}"
      }
    }
  }
}

Some MCP clients use a different environment interpolation syntax. Keep the endpoint and bearer header, then adapt the variable expression to your client.

Prompt scaffold

Use BrowserCity MCP browser tools to research: <topic>.

Workflow:
- Open a browser session.
- Visit at least three primary or high-quality sources.
- Use page markdown or snapshots to capture evidence.
- Keep notes with source URLs.
- If a source blocks access or needs login, say so instead of guessing.
- Close the browser session before you finish.

Return:
- Summary
- Source-by-source notes
- Open questions

Why this template is safe

The browser tools give the agent page context, but your prompt still controls the task boundary. Avoid asking the agent to bypass access controls, solve CAPTCHAs, or collect personal data without permission.

Cost and plan notes

This pattern is usually light: one browser session, a small number of pages, and modest traffic. If you scale it to many topics or parallel agents, watch session concurrency and use /pricing-calculator to estimate browser hours plus traffic.