Docs

recipe MCP tool

Reusable playbooks.

MCP name recipe dispatches to renderer recipe.

Does not use hosted quota. Local editor operation.

Reusable playbooks. action search finds recipes by query. action get loads one recipeId. action create writes a private recipe. action update patches an existing private recipe. Editor must be open. Does not use hosted quota. Local editor operation.

Advertised MCP input

This tool is registered with passthrough. The fields below are what listTools advertises. The renderer still validates the child fields for the chosen action or kind. Extra advertised-only values are documented under Reachable behavior.

Field	Type	Required	Constraints	Default	Description
`action`	enum("search", "get", "create", "update")	yes	—	—	action for recipe
`resultStart`	integer	no	min 0; max 9007199254740991	0	First character offset to return. Use 0 for the first call.
`resultEnd`	integer	no	min 1; max 9007199254740991	8000	Exclusive end offset. resultEnd - resultStart must be at most 8000. Increment resultStart by 8000 to page through large results.

Reachable behavior

Advertised action values match the renderer children: search, get, create, update.

Renderer child fields

These are the authoritative fields for each action. Send them on the same MCP call. Grouped MCP tools use passthrough, so these fields are not all listed in the advertised schema.

`action=search`

Renderer tool: searchRecipes.

Search semantic document playbooks that can raise output quality. Requires AI Search to be ready (embedding model + recipe index). In Agent Mode, try this before tackling substantive user tasks so existing recipes can guide the work. If unavailable because AI Search is still warming up, continue without recipes — do not retry until ready.

Field	Type	Required	Constraints	Default	Description
`query`	string	yes	minLength 1	—	Natural language description of the needed reusable playbook.
`visibility`	enum("all", "private", "public")	no	—	"all"	—
`limit`	number	no	min 1; max 20	5	—

`action=get`

Renderer tool: getRecipe.

Retrieve one recipe playbook as structured context and record basic usage. Use before applying a recipe.

Field	Type	Required	Constraints	Default	Description
`recipeId`	string	yes	minLength 1	—	Recipe ID returned by searchRecipes.
`visibility`	enum("all", "private", "public")	no	—	"all"	—

`action=create`

Renderer tool: createRecipe.

Create a private reusable recipe playbook. Requires AI Search to be ready (embedding model + recipe index). Before creating, searchRecipes should be used to avoid duplicates and updateRecipe should be used for a close match. If AI Search is still warming up, skip creating a recipe.

Field	Type	Required	Constraints	Default	Description
`title`	string	yes	minLength 1; maxLength 200	—	—
`summary`	string	yes	minLength 1; maxLength 2000	—	—
`whenToUse`	string	yes	minLength 1; maxLength 8000	—	—
`steps`	object[]	yes	minItems 1; maxItems 80	—	—
`steps[].title`	string	yes	minLength 1; maxLength 200	—	—
`steps[].detail`	string	yes	minLength 1; maxLength 8000	—	—
`acceptanceCriteria`	string[]	yes	minItems 1; maxItems 80	—	—
`tags`	string[]	no	maxItems 40	[]	—
`document`	string	yes	minLength 1; maxLength 80000	—	—

`action=update`

Renderer tool: updateRecipe.

Update an existing private recipe when a new reusable pattern overlaps it. Requires AI Search to be ready (embedding model + recipe index). Use after searchRecipes finds a close match. If AI Search is still warming up, skip updating recipes.

Field	Type	Required	Constraints	Default	Description
`recipeId`	string	yes	minLength 1	—	—
`title`	string	no	minLength 1; maxLength 200	—	—
`summary`	string	no	minLength 1; maxLength 2000	—	—
`whenToUse`	string	no	minLength 1; maxLength 8000	—	—
`steps`	object[]	no	minItems 1; maxItems 80	—	—
`steps[].title`	string	yes	minLength 1; maxLength 200	—	—
`steps[].detail`	string	yes	minLength 1; maxLength 8000	—	—
`acceptanceCriteria`	string[]	no	minItems 1; maxItems 80	—	—
`tags`	string[]	no	maxItems 40	—	—
`document`	string	no	minLength 1; maxLength 80000	—	—

Result window

Every local MCP tool accepts resultStart and resultEnd so large payloads can be paged.

resultStart — first character offset. Default 0.
resultEnd — exclusive end offset. Default 8000.
resultEnd - resultStart must be at most 8000 characters.
Increment resultStart by 8000 to read the next window.