Guides
Decisions and implementation questions about AI agents and websites, answered directly.
Comparisons
Local vs cloud browser agents: what actually differs?Cloud browser agents run in a fresh remote browser; local ones use your browser and session. How login, blocking, privacy and cost differ, with benchmark data.MCP server vs browser agent: which should you use?Use an MCP server for tasks it covers: structured data in about a second at no model cost. Use a browser agent for open-ended pages. Benchmark: 0.9 s vs 78 s.Official API vs internal API: what is the difference?An official API is documented and supported; an internal API is what a site's own pages call. Wield recipes replay the second, as you, in your browser.Recipes vs scraping: how is Wield different?Scraping parses pages from a server. A Wield recipe makes the page's own few requests in your browser, with your session, at human pace, with no model.When is a browser agent still the better choice?Browser agents win on sites with no recipe, one-off flows and visual tasks. Honest list of where Wield recipes don't help, plus where they do.
Cost and speed
How fast can an AI agent use a website?Median Wield recipe run: 0.9 s. Median browser-agent run: 78 s, about 85 times slower. Per-task numbers from our benchmark as of 7 Nov 2026.How we benchmark recipes against browser agents24 tasks, 21 sites, 3 queries, 3 runs per side, the same machine check. Setup, exclusions and caveats, so you can judge and reproduce the results.What does a browser agent cost per task?In our benchmark a Browser Use and Claude Sonnet 5.5 run cost $0.184 on average, with a median time of 78 s. A Wield recipe cost $0 in model tokens, at 0.9 s.
Decisions
How to choose a way to connect Claude to websitesFour ways to connect Claude to a website: an official API, a ready MCP server, a browser agent or recorded recipes. A decision table by login, cost and speed.Questions to ask any AI agent vendor about credentialsEight questions to ask before an AI agent touches your accounts: where cookies live, who sees results, what is logged, and how writes are approved.Read-only vs write tools: why agents should start read-onlyRead tools look things up; write tools change data. Wield recipes carry an effect flag, most are read, and every write stops for your Allow.Should I give my AI agent my password?No. An agent does not need your password when its tools run in your logged-in browser. Wield never copies or sends your cookies or passwords.
EU
Implementation
How an AI agent finds the right recipe for a taskThe agent calls find_recipe with the task, gets ranked recipes with params, then run. Here is the flow, with a real example and what happens when none fits.How to connect Claude Desktop to a site without an APIInstall the Wield .mcpb bundle, add the browser extension, then ask. Claude calls recipes that reuse the site's own requests, in your logged-in browser.How to contribute a public recipe to WieldThree ways to write a recipe, the automated checks it must pass, live verification and what gets accepted. Every public recipe is reviewed against site terms.How to find a website's internal API in DevToolsOpen DevTools, use the Network tab with a Fetch/XHR filter, do the task once, and read the JSON request. Then turn it into a Wield recipe or record it.How to give AI agents your internal tools with private recipesRecord an intranet or back-office task once. It becomes a private recipe your team's agents can run, signed with your org key and kept out of the public index.How to keep an AI agent from booking without askingMake confirmation structural, not a prompt. Wield stops before the first write request and shows every request and body. Deny is the default.How to record a site and turn it into a Wield recipeDo a task once in your browser. Wield sanitizes the recording on your device, the builder drafts a recipe, and your browser tests it. Paid plans only.How to set organization policies for AI agents with WieldAdmins can allow only listed sites, deny writes and force telemetry off for every member. A policy can only restrict; it never widens or skips Allow.How to test an MCP server before you rely on itTest an MCP server in four layers: schema, tool calls, error paths and live behavior. Examples use wield-mcp's doctor, status and run tools.How to write a Wield recipe, step by stepA recipe is a small JSON file with site, params, steps, output and verify. This guide covers each part, the 12-step and 40-request limits, and testing.What happens to AI agent tools when a website changes?Recipes are versioned and checked daily; a failing one is marked degraded and wield-mcp falls back to the last good version. Real examples of v2 fixes.Wield error codes and how to fix themEvery Wield error code with its cause and fix: executor_offline, tab_error, login_required, site_challenge, limit_exceeded, expect_failed and more.
Problems
Why can't my AI agent log in to websites?A fresh cloud browser has no login session, so an agent cannot log in there. Wield runs in your own logged-in browser, and no password or cookie is shared.Why do AI agents get stuck in CAPTCHA loops?Agents loop on CAPTCHAs because they retry a page that keeps challenging a fresh browser. Wield stops at the first check and hands it to you.Why do browser agents get blocked by websites?Cloud browser agents often run in a fresh headless browser with an empty profile. In our benchmark, 10 sites blocked that agent. Wield ran in your own browser.
Security
How to keep AI agent traffic polite to websitesPer-site pacing, a 40-request cap per run and 60 s timeouts keep an agent human-paced. Why it matters and what the benchmark paced at.How web pages can attack AI agents (prompt injection)A page can hide instructions for the agent that reads it. Wield marks results as website content and tells the agent never to follow instructions in them.Why signed recipes matter for AI agent safetyWield signs the registry index and every recipe with Ed25519 and refuses unsigned or changed ones. Recipes are data, not code. Here is what that blocks.