Zod to Json Schema

Looking for the exact opposite? Check out json-schema-to-zod

Summary

Does what it says on the tin; converts Zod schemas into JSON schemas!

• Supports all relevant schema types, basic string, number and array length validations and string patterns.
• Resolves recursive and recurring schemas with internal $refs.
• Supports targeting legacy Open API 3.0 specification (3.1 supports regular Json Schema).
• Supports Open AI strict mode schemas (Optional object properties are replaced with required but nullable ones).

Sponsors

A great big thank you to our amazing sponsors! Please consider joining them through my GitHub Sponsors page. Every cent helps, but these fellas have really gone above and beyond 💚:

Build AI apps and workflows with Retool AI retool.com

Cut code review time & bugs in half coderabbit.ai

Usage

Basic example

import { z } from "zod";
import { zodToJsonSchema } from "zod-to-json-schema";

const mySchema = z
  .object({
    myString: z.string().min(5),
    myUnion: z.union([z.number(), z.boolean()]),
  })
  .describe("My neat object schema");

const jsonSchema = zodToJsonSchema(mySchema, "mySchema");

Expected output

{
  "$schema": "http://json-schema.org/draft-07/schema#",
  "$ref": "#/definitions/mySchema",
  "definitions": {
    "mySchema": {
      "description": "My neat object schema",
      "type": "object",
      "properties": {
        "myString": {
          "type": "string",
          "minLength": 5
        },
        "myUnion": {
          "type": ["number", "boolean"]
        }
      },
      "additionalProperties": false,
      "required": ["myString", "myUnion"]
    }
  }
}

Options

Schema name

You can pass a string as the second parameter of the main zodToJsonSchema function. If you do, your schema will end up inside a definitions object property on the root and referenced from there. Alternatively, you can pass the name as the name property of the options object (see below).

Options object

Instead of the schema name (or nothing), you can pass an options object as the second parameter. The following options are available:

Definitions

The definitions option lets you manually add recurring schemas into definitions for cleaner outputs. It's fully compatible with named schemas, changed definitions path and base path. Here's a simple example:

const myRecurringSchema = z.string();
const myObjectSchema = z.object({ a: myRecurringSchema, b: myRecurringSchema });

const myJsonSchema = zodToJsonSchema(myObjectSchema, {
  definitions: { myRecurringSchema },
});

Result

{
  "type": "object",
  "properties": {
    "a": {
      "$ref": "#/definitions/myRecurringSchema"
    },
    "b": {
      "$ref": "#/definitions/myRecurringSchema"
    }
  },
  "definitions": {
    "myRecurringSchema": {
      "type": "string"
    }
  }
}

Error Messages

This feature allows optionally including error messages created via chained function calls for supported zod types:

// string schema with additional chained function call checks
const EmailSchema = z.string().email("Invalid email").min(5, "Too short");

const jsonSchema = zodToJsonSchema(EmailSchema, { errorMessages: true });

Result

{
  "$schema": "http://json-schema.org/draft-07/schema#",
  "type": "string",
  "format": "email",
  "minLength": 5,
  "errorMessage": {
    "format": "Invalid email",
    "minLength": "Too short"
  }
}

This allows for field specific, validation step specific error messages which can be useful for building forms and such. This format is accepted by react-hook-form's ajv resolver (and therefor ajv-errors which it uses under the hood). Note that if using AJV with this format will require enabling ajv-errors as vanilla AJV does not accept this format by default.

Custom Error Message Support

• ZodString
  • regex
  • min, max
  • email, cuid, uuid, url
  • endsWith, startsWith
• ZodNumber
  • min, max, lt, lte, gt, gte,
  • int
  • multipleOf
• ZodSet
  • min, max
• ZodArray
  • min, max

Additional properties

By default, Zod removes undeclared properties when parsing object schemas. In order to replicate the expected output of this behaviour, the default for behaviour of zodToJsonSchema is to set "additionalProperties" to false (although the correctness of this can be debated). If you wish to allow undeclared properties you can either:

• Set removeAdditionalStrategy to "strict". This will allow additional properties for any object schema that is not declared with .strict().
• Leave removeAdditionalStrategy set to its default value of "passthrough", and add .passtrough() to your object schema.

Removing the additionalProperties keyword using the allowedAdditionalProperties and/or rejectedAdditionalProperties options.

Some schema definitions (like Googles Gen AI API for instance) does not allow the additionalProperties keyword at all. Luckily the JSON Schema spec allows for this: leaving the keyword undefined should have the same effect as setting it to true (as per usual YMMV). To enable this behaviour, set the option allowedAdditionalProperties to undefined.

To exclude the keyword even when additional properties are not allowed, set the rejectedAdditionalProperties to undefined as well.

Heads up ⚠️: Both of these options will be ignored if your schema is declared with .catchall(...) as the provided schema will be used instead (if valid).

Expected outputs

Substitute true and false for undefined according to allowedAdditionalProperties and/or rejectedAdditionalProperties respectively.

override

This options takes a callback receiving a Zod schema definition, the current reference object (containing the current ref path and other options), an argument containing inforation about wether or not the schema has been encountered before, and a forceResolution argument.

Important: if you don't want to override the current item you have to return the ignoreOverride symbol exported from the index. This is because undefined is a valid option to return when you want the property to be excluded from the resulting JSON schema.

import zodToJsonSchema, { ignoreOverride } from "zod-to-json-schema";

zodToJsonSchema(
  z.object({
    ignoreThis: z.string(),
    overrideThis: z.string(),
    removeThis: z.string(),
  }),
  {
    override: (def, refs) => {
      const path = refs.currentPath.join("/");

      if (path === "#/properties/overrideThis") {
        return {
          type: "integer",
        };
      }

      if (path === "#/properties/removeThis") {
        return undefined;
      }

      // Important! Do not return `undefined` or void unless you want to remove the property from the resulting schema completely.
      return ignoreOverride;
    },
  },
);

Expected output:

{
  "type": "object",
  "required": ["ignoreThis", "overrideThis"],
  "properties": {
    "ignoreThis": {
      "type": "string"
    },
    "overrideThis": {
      "type": "integer"
    }
  },
  "additionalProperties": false
}

postProcess

Besided receiving all arguments of the override callback, the postProcess callback also receives the generated schema. It should always return a JSON Schema, or undefined if you wish to filter it out. Unlike the override callback you do not have to return ignoreOverride if you are happy with the produced schema; simply return it unchanged.

import zodToJsonSchema, { PostProcessCallback } from "zod-to-json-schema";

// Define the callback to be used to process the output using the PostProcessCallback type:
const postProcess: PostProcessCallback = (
  // The original output produced by the package itself:
  jsonSchema,
  // The ZodSchema def used to produce the original schema:
  def,
  // The refs object containing the current path, passed options, etc.
  refs,
) => {
  if (!jsonSchema) {
    return jsonSchema;
  }

  // Try to expand description as JSON meta:
  if (jsonSchema.description) {
    try {
      jsonSchema = {
        ...jsonSchema,
        ...JSON.parse(jsonSchema.description),
      };
    } catch {}
  }

  // Make all numbers nullable:
  if ("type" in jsonSchema! && jsonSchema.type === "number") {
    jsonSchema.type = ["number", "null"];
  }

  // Add the refs path, just because
  (jsonSchema as any).path = refs.currentPath;

  return jsonSchema;
};

const jsonSchema = zodToJsonSchema(zodSchema, { postProcess });

Using postProcess for including examples and other meta

Adding support for examples and other JSON Schema meta keys are among the most commonly requested features for this project. Unfortunately the current Zod major (3) has pretty anemic support for this, so some userland hacking is required. Since this is such a common usecase I've included a helper function that simply tries to parse any description as JSON and expand it into the resulting schema.

Simply stringify whatever you want added to the output schema as the description, then import and use jsonDescription as the postProcess option:

import zodToJsonSchema, { jsonDescription } from "zod-to-json-schema";

const zodSchema = z.string().describe(
  JSON.stringify({
    title: "My string",
    description: "My description",
    examples: ["Foo", "Bar"],
    whatever: 123,
  }),
);

const jsonSchema = zodToJsonSchema(zodSchema, {
  postProcess: jsonDescription,
});

Expected output:

{
  "$schema": "http://json-schema.org/draft-07/schema#",
  "type": "string",
  "title": "My string",
  "description": "My description",
  "examples": ["Foo", "Bar"],
  "whatever": 123
}

Known issues

• The OpenAI target should be considered experimental for now, as some combination of options may break the compatibility.
• When using .transform, the return type is inferred from the supplied function. In other words, there is no schema for the return type, and there is no way to convert it in runtime. Currently the JSON schema will therefore reflect the input side of the Zod schema and not necessarily the output (the latter aka. z.infer). If this causes problems with your schema, consider using the effectStrategy "any", which will allow any type of output.
• JSON Schemas does not support any other key type than strings for objects. When using z.record with any other key type, this will be ignored. An exception to this rule is z.enum as is supported since 3.11.3
• Relative JSON pointers, while published alongside JSON schema draft 2020-12, is not technically a part of it. Currently, most resolvers do not handle them at all.
• Since v3, the Object parser uses .isOptional() to check if a property should be included in required or not. This has the potentially dangerous behavior of calling .safeParse with undefined. To work around this, make sure your preprocess and other effects callbacks are pure and not liable to throw errors. An issue has been logged in the Zod repo and can be tracked here.
• JSON Schema version 2020-12 is not yet officially supported. However, you should be able to use this library to obtain a compatible schema for the 2020-12 format just by changing the returned schema's $schema field to: "https://json-schema.org/draft/2020-12/schema#"

Versioning

This package does not follow semantic versioning. The major and minor versions of this package instead reflects feature parity with the Zod package.

I will do my best to keep API-breaking changes to an absolute minimum, but new features may appear as "patches", such as introducing the options pattern in 3.9.1.

Changelog

https://github.com/StefanTerdell/zod-to-json-schema/blob/master/changelog.md