Any website, as a tool for your agent.
Wield calls a site's own internal API from your Chrome, with the session you already have. Your agent gets a typed tool. Your logins never leave your device.
Find me a train from Berlin to Munich on Friday morning, cheapest first.
- find_recipe"train Berlin Munich"→bahn.de/search-connectionsread-only · no login
- runbahn.de/search-connectionsfrom: "Berlin Hbf", to: "München Hbf", depart_at: Friday 08:00→connections[]in your Chrome, in a www.bahn.de tab
Here are Friday's connections after 08:00, sorted by price.
~/.wield/logs/executor.ndjsonevery request that left this laptop
- GETwield.run/api/v1/recipes/bahn.de/search-connections200
- GETwww.bahn.de/web/api/reiseloesung/orte200
- GETwww.bahn.de/web/api/reiseloesung/orte200
- POSTwww.bahn.de/web/api/angebote/fahrplan200
- wield.run saw one recipe download. Not the stations, not the results.
- www.bahn.de saw a normal search from your browser.
- Cookies and passwords never left Chrome.
Works withClaude DesktopClaude CodeCursorCodexAny MCP client
We never see your password.
Cloud agent platforms log in for you, so your passwords and cookies live on their servers. Wield runs the other way around. The recipe runs inside your own Chrome, in a tab of the site you're already logged into, and the site sees an ordinary request from your browser. We see which recipe you downloaded.
Your device
Runs the recipe
- Chrome extension. Sends the recipe's requests from a tab of the site, with your session.
- wield-mcp. The MCP server your agent talks to, over stdio. No port is opened.
- Bridge. Connects the two through a socket only your user account can open.
The site
Answers your browser
- Gets the recipe's requests from your browser, with your cookies, like a click you made yourself.
- Only the origins listed in the recipe can be called. Everything else is blocked.
wield.run
Serves recipes
- Public JSON, the same for everyone, each with a sha256 you can check.
- Counts the site requests you send us.
- If you opt in: ok or failed and duration per recipe. Nothing else.
- Sent to wield.run
- Recipe downloads. Site requests you choose to send. With telemetry on, and it is off by default: recipe id, version, ok or failed, duration.
- Sent to the site
- The recipe's requests, from your own browser with your own cookies, and only to the origins the recipe lists.
- Never leaves your device
- Cookies, tokens, passwords, page contents, the parameters you pass, and the results.
- Written to your disk
- Every outbound request goes into ~/.wield/logs: destination, method, path, status and time. No bodies, no query values. Read it any time.
- Writes ask first
- A recipe that books or sends something shows you exactly what it will send and waits for you to click Allow. No answer in 120 seconds means no.
Recipes are public. Runs are private.
Three parts. Only the first one is on our servers, and it holds nothing about you.
- 1 · wield.run
Registry
A recipe is a JSON file: the site's internal API calls for one task, its parameters and how to read the answer. Data, never code. No secrets, no cookies. Anyone can read every recipe.
GET /api/v1/recipes/bahn.de/search-connections - 2 · your Chrome
Executor
The extension runs the recipe as same-origin requests from a tab of the site, with the session you already have. You handle logins and 2FA the way you always do.
fetch → www.bahn.de only - 3 · your agent
MCP server
wield-mcp runs locally over stdio. Claude, Cursor and Codex call find_recipe and run like any other tool and get typed results back.
npx -y wield-mcp
Sites that work today.
11 recipes on 9 sites, starting with Europe. Each recipe is one task on one site. Health comes from opt-in run reports over the last 7 days.
Same tasks. Two ways to run them.
A browser-driving agent (Browser Use) and Wield recipes run the same tasks on the same sites. We publish success rate, median latency and cost per task, with the raw logs.
Browser Use and Wield recipes run the same tasks on the same sites. Success rate, median latency and cost per task go here, with the raw logs, when the run finishes. We publish the result whichever way it goes.
Which sites do you need?
Name the site and the task. The most requested sites get recorded first. Leave an email if you want to hear when yours works.
Most requested
Live countNo requests yet. Yours will be the first.