How an AI agent finds the right recipe for a task

#find_recipe

find_recipe takes a task of 1 to 300 characters in German or English, an optional site to restrict the search, and a limit from 1 to 10, with a default of 5.

Each result includes:

• recipe_id, such as bahn.de/search-connections;
• version, site_id and task;
• effect, either read or write;
• requires_login;
• params_schema, the JSON schema the params must match;
• description and a score.

Wield ranks results by task words, keywords, site names and synonyms. It prefers sites in your country, which you set in the config. A health field appears only when a recipe is degraded or broken.

Calling find_recipe also makes the extension open or keep a background tab for the top site, so the run does not wait for a page to load.

#run

run takes the recipe_id, the params and an optional version. Before anything is sent, Wield checks the params against the schema and fills in defaults. Then it loads the recipe from its cache, checks the signature and sends the recipe to the extension. The extension runs it in a tab of the site, so requests carry your normal browser session.

The result returns the output, the duration, the number of requests and the path to the executor log. If the latest version is degraded, Wield can run an older good version and reports a fallback.

For the full sequence, see how Wield works.

#An example

A user asks Claude for a train. The agent makes these calls:

1. find_recipe with task set to "train from Berlin to Munich tomorrow morning" and limit set to 3. The top result is bahn.de/search-connections, with from, to and date in its schema.
2. run with recipe_id set to bahn.de/search-connections and params from: "Berlin Hbf", to: "München Hbf" and date: "2026-11-08". The agent turns "tomorrow" into an absolute date before it fills the params.
3. The agent reads the returned connections and answers with the departure times, duration, transfers and fare.

The train search needs no login, so both calls run right away.

#When no recipe fits

If nothing matches, find_recipe returns no results and a hint that suggests request_recipe. Two modes are available:

• request adds the site and the task text to the public request list on wield.run. Only the site and the task are sent.
• record builds a private recipe for your organization. You do the task once in your browser, review a sanitized recording, and the builder drafts a recipe. The draft is tested in your browser before it is saved. Record mode needs a Pro or Team plan. The new recipe_id can run right away.

Use list_sites to see which sites already have recipes. Pass site to find_recipe when a task names one.

#FAQ

#Does the agent see the recipe's code?

No. Recipes contain no code. They are JSON files that describe the requests, the params and the shape of the result, and Wield's interpreter runs them.

#What if two recipes match?

The agent gets both, with scores. It can choose the higher score, pass site to narrow the search, or ask you which one you mean.

#Can I pin a recipe version?

Yes. The version input of run pins one version. By default Wield runs the latest version, or the last good one when the latest is degraded.

#Does the search leave my computer?

No. Ranking runs over the cached index on your computer. The registry receives a request only when a recipe is downloaded or when you send a request for a new site.