Research with sources
Collect public documentation and articles, then have your agent keep the original links alongside its notes. Revisit the source when a claim matters.
Proxies for AI Agents
Give the browser or script behind your agent a proxy connection. Research public pages, compare regional content and bring source material into your workflow.

Collect public documentation and articles, then have your agent keep the original links alongside its notes. Revisit the source when a claim matters.
Inspect the same public product page or your own landing page from a requested market. Record language, currency and the content actually returned.
Research a defined batch of public URLs. Let your own tools extract fields, remove duplicates and save a dated result.
An agent can call several tools, each with its own network connection. Set the proxy in the browser, HTTP client or script that fetches the page. Changing a chat app’s connection does not prove that its browsing tool uses the same route.
For Claude Code, Codex, Gemini or another host, first identify the tool doing the browsing. A local browser, a shell command and a remotely hosted browser may need different settings. Follow that tool’s documented proxy options.
| Layer | Its job |
|---|---|
| Your agent | Chooses actions and works with the results. |
| Browser or script | Opens pages, renders content when needed and extracts the fields you request. |
| Proxy | Routes the configured web requests through an exit IP. |
Start with a small set of public product pages. Ask for the product name, displayed price, currency, source URL and collection time. These are suggested output fields, not an automatic USM export.
Independent page checks can use changing exits. A sequence that depends on the same session needs continuity instead: changing IP between steps can interrupt a workflow.
A sticky session requests the same exit for a period; it is not a permanent IP or a guarantee against disconnection. Confirm the session options available for the USM product before relying on a particular duration.
USM catalogue prices use IP count or a GB allowance. Model calls, hosted browsers and extraction services have their own fees. One prompt may open many pages and load images or repeat requests, so prompt count is not a reliable traffic estimate.
Measure a representative run, then use its traffic per completed task to estimate a larger batch. Include the resources your task actually needs and leave room for retries; do not assume a fixed number of pages per GB.
This guide describes standard proxy configuration. USM does not currently offer a public scraping API, an installable Claude / Codex Skill or a public MCP service. Those release pages will identify supported operations and setup instructions when available.
USM supplies the proxy connection; it does not supply your model, render pages or return automatically parsed data. A proxy also does not grant access to a model subscription or to private website content.
CHOOSE THE CONNECTION
Public-page tasks that accept datacenter networks; compare the cost of the IP count you need.
Work requiring a residential network and a defined set of exits; check IP retention for multi-step tasks.
Independent research batches that need changing residential exits; estimate traffic from a sample run.
Prices are for planning; ordering is not open yet. Confirm locations, protocols, traffic limits and IP retention for the selected product before purchase. A requested country is not a stock reservation.
It changes the route used to reach a page. Your agent still needs to extract the right fields, distinguish current content from old information and cite its sources. Check important conclusions against the original pages.
The browsing tool must support the protocol and authentication provided for your proxy. Some cloud tools control their own network and do not accept a custom proxy. Check the specific tool and version before choosing a plan.
Inspect the target response and tool logs. A connection can work while a page needs JavaScript, authentication or a different extraction method. Send support the tool version, time, target URL and redacted error; leave out passwords and private page data.
YOUR NEXT STEP
Share your tools, requested locations and expected usage.