How to choose a way to connect Claude to websites
To connect Claude to a website, choose one of four routes: an official API, a ready MCP server, a browser agent, or recorded recipes. Decide by whether the task needs your login, how often you will run it, and how fast you need the answer. For read tasks on a site with a recipe, a recipe is usually fastest and cheapest.
Decision table
| Route | Best when | Needs your login | Model cost per run | Speed | Main limit |
|---|---|---|---|---|---|
| Official API | The site publishes an API for your task | Usually an API key or OAuth | Only the model tokens you use | Depends on the API | Often limited to the vendor's scope |
| Ready MCP server | Someone already built one for the site | Depends on the server | Model tokens | Depends on the server | Quality depends on its author |
| Browser agent | No API and no recipe, and the task is one-off | Only if you share a session | About $0.18 per run in our benchmark | Median 78 s | Slow, costly, and blocked by some sites in our runs |
| Recorded recipe (Wield) | A recipe exists, or you can record one | Runs in your own logged-in browser | $0 | Median 0.9 s | Covers only the tasks someone has written |
The browser agent cost and speed are from our benchmark as of 7 Nov 2026, with Browser Use 0.13.14 driving Claude Sonnet 5.5. Official API speed and cost depend on the API and were not measured.
If the site has an official API
Start here when the vendor offers one for your task. Read its terms, use the narrowest scope it offers, and store the key outside your prompts. An official API is the most stable route when it covers what you need. It is not always complete. Many sites have no public API for private data, such as your own orders or appointments.
If a recipe already exists
Ask your agent to look up the task. With Wield, the agent calls find_recipe with the task in your words, then run with the parameters. Runs happen in your browser, at the speed in the table above. Wield's registry has public recipes for 29 sites as of 7 Nov 2026. Ask your agent for list_sites to see the current list.
If nothing exists: request or record
You have two options when the site has no recipe:
- Request it. Use
request_recipefrom your agent, or add the site at wield.run/sites. The most-requested sites are built first. This route is free. - Record it. Show Wield the task once in your browser, and the recipe builder drafts a recipe that your browser then tests. Recording and the builder are paid-plan features. See how to record a site.
If you need a result today and nothing else works, a browser agent is the fallback. Expect it to be slower and to cost model tokens.
Questions to settle first
- Does the task need your login? If yes, prefer a route that runs in your own browser.
- Will you run the task once or many times? Repeated runs favour a recipe or an API.
- Does the task change data, such as a booking or a message? Use a route that asks for your approval before any write.
- Is the site's bot protection likely to stop an automated cloud browser? Our benchmark found that some sites blocked cloud browser agents. Wield stops with
site_challengewhen a site shows a check, and asks you to solve it in your own tab.
FAQ
Which route should I try first?
Try a recipe first if one exists, because it is the fastest and needs no model tokens per run. If there is no recipe, request one. Use a browser agent only for a one-off task that you cannot wait on.
Does Claude need a special app to use Wield?
No. Wield is an MCP server, so any MCP client can use it. The setup command npx -y wield-mcp init configures Claude Desktop, Claude Code, Cursor and Codex. See AI apps that work with Wield.
Is Wield an agent?
No. Wield is the layer that agents call. There is no model in the loop when a recipe runs, so a run costs no model tokens.