Local vs cloud browser agents: what actually differs?
A local agent runs in your own browser, with your session, and your results stay on your device. A cloud browser agent runs in a fresh browser on a remote machine, with no session. In our benchmark, the cloud-style setup was blocked by 10 sites. The local recipe side passed on 7 of those 10. The setup also changed the time and cost of each task.
Comparison table
| Factor | Local (Wield in your browser) | Cloud-style browser agent (benchmark setup) |
|---|---|---|
| Browser | Your everyday Chromium browser | Fresh headless Chromium, empty profile |
| Session | Your existing logins | None |
| Sites that blocked it in our runs | None flagged in the benchmark | 10 of 21 sites |
| Median time per run | 0.9 s | 78 s |
| Model cost per run | $0 | $0.184 |
| Browser must be open | Yes | No |
| Where results go | Extension, then wield-mcp, then your AI app. Never through wield.run | Depends on the provider |
Login and sessions
A local recipe uses the session in your browser. If you are logged in, the recipe can use it. If you are not, the run stops with login_required, and you log in yourself. A cloud browser starts with an empty profile, so it has no session. See why an agent can't log in for the details.
Blocking and bot checks
In our runs, bahn.de blocked the browser agent. Its success was 0%, and one run took 675.9 s. The Wield recipe for train search passed 9 of 9 runs, with a median of 2.1 s.
Local does not mean unblockable. When a site shows a check, Wield stops with site_challenge and asks you to handle the check in your own tab. It does not solve checks. The benchmark also recorded recipe misses. The Zalando recipe failed 9 of 9 runs, so local is not a guarantee.
Privacy: where credentials and results go
Wield keeps passwords and cookies on your computer. It never reads, copies or sends them. Results go from the extension to wield-mcp and then to your AI app. They never pass through wield.run. Logs stay on your computer and hold no bodies or query values. Telemetry is off by default.
For a cloud browser agent, the data path depends on the provider. Read that provider's data policy before you send it credentials or results.
Cost and speed
Local recipes cost no model tokens, because no model runs during a recipe. The benchmark median was 0.9 s for recipes and 78 s for the browser agent. The browser agent averaged $0.184 per run in model cost. See what a browser agent costs per task.
FAQ
Is a local agent always better?
No. A recipe covers only the sites that have one. A browser agent can try any site, although it was blocked on 10 of the 21 sites in our benchmark. Choose by site and task.
Does my browser have to be open?
Yes, for Wield. Recipes run in a tab of the site, so the browser must be running. The window can be in the background. Wield opens background tabs and reuses tabs you already have.
Does local mean my data stays private?
Wield keeps passwords and cookies on your computer and does not send them to wield.run. Results go to your AI app, not to Wield. Your AI app has its own data policy, which applies to what it receives.
Which browsers does Wield support?
Chrome, Edge, Brave, Arc, Vivaldi, Opera and Chromium, on macOS, Windows and Linux. Firefox and Safari are not supported, because Wield needs Chromium's native messaging and extension APIs.