Configure your SDKs

Stainless automatically generates idiomatic yet flexible SDKs for your API. But every API is different. Create the perfect SDK for your API and your users by configuring the Stainless Generator.

When you first generate an SDK with Stainless, we create a reasonable draft configuration for you. You can see and edit it in the SDK Studio.

Resources

[reference]

The most important consideration when designing an SDK is having the correct names for your resources and CRUD operations. The Stainless configuration defines resources as models and methods, grouped by primitive.

Consider the following:

resources:
  accounts:
    models:
      account: '#/components/schemas/Account'
    methods:
      list: get /accounts
      create: post /accounts
      retrieve: get /accounts/{account_id}
      delete: delete /accounts/{account_id}
      update: put /accounts/{account_id}
    subresources:
      friends:
        models:
          friends: '#/components/schemas/Friends'
        methods:
          list: get /accounts/{account_id}/friends
          create: post /accounts/{account_id}/friends
          delete: delete /accounts/{account_id}/friends/{friend_id}

In this example, we have an accounts resource and an accounts.friends resource nested within. The resulting SDK usage is:

# `foo` is the instantiation of the library.
account = foo.accounts.create(...)
friends = foo.accounts.friends.list(...)

Depending on the constraints and how you want your users to reason about your API, you may want to restructure the resources to look like:

resources:
  accounts:
    models:
      account: '#/components/schemas/Account'
      friends: '#/components/schemas/Friends' # Moved here
    methods:
      list: get /accounts
      create: post /accounts
      retrieve: get /accounts/{account_id}
      delete: delete /accounts/{account_id}
      update: put /accounts/{account_id}
      # Below items were moved here and renamed
      list_friends: get /accounts/{account_id}/friends
      add_friend: post /accounts/{account_id}/friends
      remove_friend: delete /accounts/{account_id}/friends/{friend_id}

That resulting SDK usage is:

# `foo` is the instantiation of the library.
account = foo.accounts.create(...)
friends = foo.accounts.listFriends(...)

Which approach is better is ultimately up to how you want the user to reason about your API. In general, a good resource is a conceptually distinct, representational object in your API.

Models

[reference]

A model is a schema that Stainless generates as a distinct, named type. Consider the following configuration:

resources:
  accounts:
    methods:
      list: get /accounts
      create: post /accounts
      retrieve: get /accounts/{account_id}

By default, the Stainless generator creates methods with roughly the following signatures, even if they all reference the same #/components/schemas/Account in their response. (Note the return types.)

class Accounts extends APIResource {
    list(...): AccountListResponse[];
    create(...): AccountCreateResponse;
    retrieve(...): AccountRetrieveResponse;
}

// usage:
const account: AccountRetrieveResponse = client.accounts.retrieve()

info

We do not consider $ref to be part of a public-facing API. OpenAPI specs are difficult to deal with and $ref is often used to circumvent that complexity, which may not perfectly align with the API you want your users to have. For example, changing a $ref for internal organization purposes should never cause a backwards-breaking change to the users of your SDK.

Alternatively, you can configure a model to belong within a resource:

resources:
  accounts:
    models:
      account: '#/components/schemas/Accounts'
    methods:
      list: get /accounts
      create: post /accounts
      retrieve: get /accounts/{account_id}

That configuration creates these methods:

class Accounts extends APIResource {
    list(...): Account[];
    create(...): Account;
    retrieve(...): Account;
}

// usage:
const account: Account = client.accounts.retrieve()

Models need not only fall under #/components/schemas/*. You may specify any valid JSON Reference from the root.

Finally, the model configuration can also accept an object rather than just a string for even more options, if you need them. Please check out our full reference docs here.

Shared Models

In general, a model belongs under the most closely related resource, and each resource should have at least one model. All of our SDKs are designed so that you can put a model where it most logically belongs without worrying about dependency issues.

There are cases where a model doesn't belong under one resource. A common example of that might be something like #/components/schemas/Address or #/components/schemas/Contact. We have a special resource called $shared where these kinds of models should be defined.

resources:
  $shared:
    models:
      address: '#/components/schemas/Address'
      contact: '#/components/schemas/Contact'

Methods

[reference]

Methods most commonly define how an endpoint maps to a resource, by providing a string value:

resources:
  accounts:
    methods:
      list: get /accounts

You can configure the method in more detail by providing an object:

resources:
  accounts:
    methods:
      list:
        type: http
        endpoint: get /accounts
        deprecated: Use the /user endpoint instead

Authentication

The Stainless Generator defines your SDK's authentication using the #/security and #/components/securitySchemes particulars from your OpenAPI spec. By default, Stainless SDKs are set to authenticate using an environment variable:

client_settings:
  opts:
    auth_token:
      type: string
      # Whether this client option is required to instantiate the client:
      nullable: false
      # Whether this value should be read from an env:
      read_env: ORG_AUTH_TOKEN
      auth: { security_scheme: BearerAuth }

# optional, overrides the OpenAPI spec's top-level security key, required if it isn't present
security:
  - BearerAuth: []

For more complicated authentication schemes, or should the Stainless Generator fail to configure yours correctly, see the examples below.

The Stainless generator uses the top-level security as the security configuration supported by the SDKs, but you may want to specify a different combination of security than the one used in your OpenAPI spec. The security and security_schemes in the Stainless config overrides the values in the spec.

HTTP Bearer Authorization: Bearer <bearer-token>

The HTTP Bearer authentication method is configured like so:

# OpenAPI
components:
  security_schemes:
    MyBearerAuth:
      type: http
      scheme: bearer
      # optional, documentation purpose only
      bearerFormat: JWT

security:
  - MyBearerAuth: {}

# Stainless config
client_settings:
  opts:
    my_bearer_token: # or `token`, `bearer_token`, `api_key`, etc.
      type: string
      read_env: ORG_BEARER_TOKEN
      auth: { security_scheme: MyBearerAuth }

HTTP Basic Authorization: Basic <base64(username:password)>

The HTTP Basic authentication method is configured like so:

# OpenAPI
components:
  security_schemes:
    MyBasicAuth:
      type: http
      scheme: basic

security:
  - MyBasicAuth: {}

# Stainless config
client_settings:
  opts:
    my_username:
      type: string
      read_env: ORG_MY_USERNAME_TOKEN
      auth: { security_scheme: MyBasicAuth, role: 'username' }
    my_password:
      type: string
      read_env: ORG_MY_PASSWORD_TOKEN
      auth: { security_scheme: MyBasicAuth, role: 'password' }

API Key <Header>: <API Key>

An API key in a header authentication method is configured like so:

# OpenAPI
components:
  security_schemes:
    MyApiKeyAuth:
      type: apiKey
      name: My-Api-Key
      in: header

security:
  - MyApiKeyAuth: {}

# Stainless config
client_settings:
  opts:
    my_api_key: # or `token`, `auth_token`, etc.
      type: string
      read_env: ORG_API_KEY_TOKEN
      auth: { security_scheme: MyApiKeyAuth }

Optional Auth

To specify that you accept no authentication, declare a security configuration with no properties like so:

security:
  - BearerAuth: {}
  - {}

OAuth2

Stainless doesn't generate code to handle initial OAuth authorization or token management for you. To handle authorization and managing tokens, you typically want to point your users to one of the existing OAuth SDKs here. We expect to offer out-of-the-box codegen support for OAuth in the future. Please reach out if this is of interest.

Stainless does support endpoints that use OAuth-generated access tokens for authorization. In most cases, you just need to update your Stainless configuration to specify that authentication is happening via Bearer Authentication.

# openapi.yml

securitySchemes:
  OAuth2:
    type: oauth2
    description: OAuth2 security scheme
    flows:
      authorizationCode:
        authorizationUrl: https://my_auth_url.com/
        tokenUrl: https://my_token_url.com/
        scopes:
          read_only: read only
          write_only: write only

# Stainless config

client_settings:
  opts:
    access_token:
      type: string
      auth:
        security_scheme: BearerAuth
      read_env: MY_TEAM_ACCESS_TOKEN

security:
  - BearerAuth: []

security_schemes:
  BearerAuth:
    type: http
    scheme: bearer

Please reach out if you need help supporting your configuration.

README.md configuration

[reference]

We currently support configuration of the various code snippets in the README.md of a generated SDK.

readme:
  example_requests:
    default:
      type: request
      endpoint: post /cards
      params:
        type: SINGLE_USE
    headline:
      type: request
      endpoint: put /cards
      params:
        type: SINGLE_USE
        account_id: 123
    pagination:
      type: request
      endpoint: list /cards
      params:
        limit: 30

The headline example is the first usage of the API that your users see in the README.md, so it should be the most 'important' endpoint in your API.

The pagination example is required if you have configured pagination, and should point to a paginated endpoint.

The default example is inherited by all other example requests, though you can manually override them if there is a need to do so. Because of this, we suggest that you choose the most 'standard' endpoint in your API.

Pagination

[reference]

Configuring pagination with Stainless SDKs generates a helper to fetch the next page and an auto-iterator to easily loop through items in your API. The helper makes it easy to manually paginate when needed and the auto-iterator makes consuming a list as natural as a for loop, with the iterator automatically fetching the next page when needed:

const iter: OffsetPage<Account> = await client.accounts.list()

for await (const account in iter) {
  if (account.name === 'Michael') {
    console.log(account.id)
  }
}

Pagination Scheme

To configure pagination, you need to first define a pagination scheme. A pagination scheme is closely related to the pagination class that is generated in the SDKs, and is made up of:

pagination:
  - name: <name of page class>
    type: <offset, cursor, cursor_id, hypermedia, etc.>
    request:
      <name of request parameter>: <schema of request parameter>
    response:
      <name of response field>: <schema of response field>

1. the type of pagination (offset, cursor, cursor_id, etc...),
2. the parameters in the request,
3. and the fields in the response that you expect to see on every paginated method of this type.

The request and response sections are used to generate the pagination classes which depending on the language is represented as a generic class, interface, or object. They should respectively contain keys of the request parameters and response JSON in your API, for example in the header, query, or the body.

You should include request parameters that are necessary for pagination, such as page_number, page_size, after, limit, and properties that should always be present in the response, like data or items (which contains the list of elements you have).

You may also include parameters that aren't strictly necessary for pagination to work. For example, it might be appropriate to add a request parameter like sort_by, which might be an enum of the ways you can sort across your API.

Pagination Type

The first step in configuring pagination is to define the type of pagination your API is using. The types of pagination Stainless supports are:

• offset: A simple pagination scheme that uses an offset (the number of entries to skip) and limit (number of elements to fetch) to paginate through a list of items.
• page_number: Similar to offset pagination, but instead of an offset, which indexes on the items, page_number indexes on chunks of elements to skip.
• cursor: A pagination scheme that uses a cursor (a string or a number to indicate the element to start from) to paginate through a list of items.
• cursor_id: Similar to cursor pagination, but the cursor comes from an id property of the response items.

note

If you haven't yet built pagination for your API and are unsure which type of pagination to use, we recommend using cursor_id pagination as it is typically more robust and performant than offset or page number pagination.

Cursor ID Pagination

pagination:
  - name: my_cursor_id_page
    type: cursor_id
    request:
      starting_after:
        type: string
        x-stainless-pagination-property:
          purpose: next_cursor_id_param
      ending_before:
        type: string
        x-stainless-pagination-property:
          purpose: previous_cursor_id_param
      limit:
        type: integer
    response:
      my_data:
        type: array
        items:
          type: object
          properties:
            id:
              type: string
              x-stainless-pagination-property:
                purpose: cursor_item_id
          required:
            - id

Example OpenAPI Specification

openapi: 3.1.0
paths:
  /accounts:
    get:
      parameters:
        - name: starting_after
          in: query
          schema:
            type: string
        - name: ending_before
          in: query
          schema:
            type: string
        - name: limit
          in: query
          schema:
            type: integer
      responses:
        '200':
          description: 'success!'
          content:
            application/json:
              schema:
                type: object
                properties:
                  my_data:
                    type: array
                    items:
                      type: object
                      properties:
                        id:
                          type: string

Cursor Pagination

pagination:
  - name: my_cursor_page
    type: cursor
    request:
      next:
        type: string
        x-stainless-pagination-property:
          purpose: next_cursor_param
      limit:
        type: integer
    response:
      my_data:
        type: array
        items:
          type: object
      next:
        type: string
        x-stainless-pagination-property:
          purpose: next_cursor_field

Example OpenAPI Specification

openapi: 3.1.0
paths:
  /accounts:
    get:
      parameters:
        - name: next
          in: query
          schema:
            type: string
        - name: limit
          in: query
          schema:
            type: integer
      responses:
        '200':
          description: 'success!'
          content:
            application/json:
              schema:
                type: object
                properties:
                  my_data:
                    type: array
                    items:
                      type: object
                  next:
                    type: string

Offset Pagination

info

Note: While creating APIs with offset pagination is often simpler to implement, they can often end up being a source of problems for both you and your users. Consider:

• If new resources are simultaneously being created or deleted that fall in the middle of the list based on the sort, you run the risk of missing or double scanning objects as you paginate through results.
• Many database engines have poor performance for large offset values, since they often have to scan through the entire table to get to the offset.

If you can, we'd recommend cursor-id-based pagination, which is more robust and performant.

pagination:
  - name: my_offset_page
    type: offset
    request:
      my_offset:
        type: integer
        description: The number of elements to skip.
        # this tells us to modify this param when getting the next page
        x-stainless-pagination-property:
          purpose: offset_count_param
      my_limit:
        type: integer
        description: The maximum number of elements to fetch.
    response:
      my_data:
        type: array
        items:
          type: object
      my_total:
        type: integer
        x-stainless-pagination-property:
          # total number of elements in the list
          purpose: offset_total_count_field
      my_count:
        type: integer
        x-stainless-pagination-property:
          # where to start the next page
          purpose: offset_count_start_field

Example OpenAPI Specification

openapi: 3.1.0
paths:
  /accounts:
    get:
      parameters:
        - name: my_offset
          in: query
          schema:
            type: integer
        - name: my_limit
          in: query
          schema:
            type: integer
      responses:
        '200':
          content:
            application/json:
              schema:
                type: object
                properties:
                  my_data:
                    type: array
                    items:
                      type: object
                  my_total:
                    type: integer
                  my_count:
                    type: integer

Page Number Pagination

note

Note: While creating APIs with page-number pagination is often simpler to implement, they can often end up being a source of problems for both you and your users. Consider:

• If new resources are simultaneously being created or deleted that fall in the middle of the list based on the sort, you run the risk of missing or double scanning objects as you paginate through results.
• Many database engines have poor performance for large offset values, since they often have to scan through the entire table to get to the offset.

If you can, we'd recommend cursor-id-based pagination, which is more robust and performant.

pagination:
  - name: my_page_number_page
    type: page_number
    request:
      page:
        type: integer
        x-stainless-pagination-property:
          purpose: page_number_param
      page_size:
        type: integer
    response:
      data:
        type: array
        items:
          type: object
      page:
        type: integer
        x-stainless-pagination-property:
          purpose: current_page_number_field
      last_page:
        type: integer
        x-stainless-pagination-property:
          purpose: total_page_count_field

Example OpenAPI Specification

openapi: 3.1.0
paths:
  /accounts:
    get:
      description: Example case for page_number pagination
      parameters:
        - name: page
          in: query
          schema:
            type: integer
        - name: page_size
          in: query
          schema:
            type: integer
      responses:
        '200':
          description: OK
          content:
            application/json:
              schema:
                type: object
                properties:
                  data:
                    type: array
                    items:
                      $ref: '#/components/schemas/MyModel'
                  last_page:
                    type: integer
                    description: The last page number
                  page:
                    type: integer
                    description: The page number

Multiple pagination schemes

The pagination section accepts multiple pagination schemes. We match each endpoint against the defined pagination schemes by making sure the relevant request parameters/response fields exist for that endpoint and have the correct type.

In cases where it's ambiguous or you want to explicitly assert that a method matches my_offset_page, you can provide paginated: my_offset_page and the generator reports an error if it doesn't match that specific page.

See the config reference for pagination for various examples and edge cases.

Other Pagination Properties

Retrieving pagination properties from headers

If your API returns pagination values in headers, you can use the from_header property to specify the header to read the value from.

pagination:
  - name: page_cursor_from_headers
    type: cursor
    request: ...
    response:
      my_cursor:
        type: string
        nullable: true
        x-stainless-pagination-property:
          purpose: next_cursor_field
          from_header: 'X-My-Cursor'

Retrieving pagination properties from nested items

If your API returns pagination values nested within the response, you can configure Stainless to read them from the nested items.

Example response:

{
  "data": {
    "items": [{}, {}]
  },
  "pagination_object": {
    "next_page": "next_page_cursor",
    "previous_page": "previous_page_cursor"
  }
}

pagination:
  - name: page_cursor_nested_items
    type: cursor_id
    request: ...
    response:
      data:
        type: object
        properties:
          items:
            x-stainless-pagination-property:
              purpose: items
            type: array
            items: {}
      pagination_object:
        type: object
        properties:
          next_page:
            type: string
            x-stainless-pagination-property:
              purpose: next_cursor_id_param
          previous_page:
            type: string
            x-stainless-pagination-property:
              purpose: previous_cursor_id_param

Top level arrays in response

You can use the is_top_level_array property to indicate that the response is a top-level array.

Example response:

[
  {
    "id": 1,
    "name": "Alice"
  },
  {
    "id": 2,
    "name": "Bob"
  }
]

pagination:
  - name: top_level_array
    type: offset
    request: ...
    response:
      items:
        type: array
        items: {}
        x-stainless-pagination-property:
          is_top_level_array: true

Miscellaneous

Extra client arguments

You can define extra client arguments, which generally appears as an extra argument on the client constructor of each SDK (in Go, it appears as an extra RequestOption). These are generally used for supplying values for authentication methods, but can be also used for extra headers and more.

client_settings:
  opts:
    pet_store_version:
      type: string # can be a string, boolean, or a number
      nullable: true # makes this an optional argument
      default: v3 # the default pet store version to use

      read_env: PETSTORE_VERSION
      send_in_header: 'X-Petstore-Version'

const client = new Petstore({
  apiKey: '...',
  petStoreVersion: 'v2', // sends 'X-Petstore-Version: v2'
});

Default headers

Default headers are headers we add to every request made by the SDK. We send platform headers so that you can collect metrics on the languages and platforms your users use. For all SDKs, we send the following:

Headers	Description
X-Stainless-Arch	The architecture, such as x32, x64, arm, aarch64, or other:xxx.
X-Stainless-Lang	The language, such as node, python, java, kotlin, or go.
X-Stainless-OS	The OS, such as Android, MacOS, Windows, FreeBSD, OpenBSD, Linux, or Other:xxx.
X-Stainless-Package-Version	The package version, such as v2.3.1.
X-Stainless-Retry-Count	Which retry number the request is for, such as 0 for the first request, 1 for the first retry, 2 for the second retry, and so on.
X-Stainless-Runtime	The runtime, such as node or CPython.
X-Stainless-Runtime-Version	The runtime version, such as v14.8.0.

We also send some extra headers for each language:

Python Headers	Description
X-Stainless-Async	Whether or not the AsyncClient was used.

Java/Kotlin Headers	Description
X-Stainless-OS-Version	The OS version, such as 14.4.

Retries

Our clients retry connection errors (for example, a network connectivity problem), 408 Request Timeout, 409 Conflict, 429 Rate Limit, and >=500 Internal errors.

By default, our clients retry 2 times (so a total of 3 requests) using the exponential backoff strategy with an initial delay of 0.5s and a max delay of 8s. We also add a jitter of 25% to spread out requests.

This can be configured by you for your users:

client_settings:
  default_retries:
    # 5 retries are made, with the interval [1, 2, 4, 8, 10 (capped by the max)]
    # not accounting for jitter
    max_retries: 5
    initial_delay_seconds: 1
    max_delay_seconds: 10

Or it can be changed by the user in each language's SDK:

client = new Petstore({
  maxRetry: 3,
});

Each request includes an X-Stainless-Retry-Count header so you know how often clients retry. See Default headers for details.

Disabling retries

Retries can be disabled in two ways. Your users can disable this at at the client level by setting max_retries to 0. For example:

client = new Petstore({
  maxRetry: 0, // Doesn't retry at all
});

Alternatively, your API can direct our SDKs to not retry by sending the X-Should-Retry header in your responses.

Retry-After and Retry-After-Ms

Our SDKs also respect the Retry-After header sent by the API, which defines in integers how many seconds we should wait before making another request. We also support the Retry-After-Ms header which is less standard but gives more fine-grained control over timings in milliseconds.

The clients only respect values that are "reasonable" which are positive values less than a minute.

X-Should-Retry

Additionally, the clients support the X-Should-Retry header, which can be used to explicitly control whether a request should be retried. This header must be explicitly set to either X-Should-Retry: true or X-Should-Retry: false to be used as an override.

When set to true, it forces a retry attempt even if the request would not normally be retried based on the default retry rules. When set to false, it prevents a retry attempt even if the request would normally be retried. If the header is not present or set to any other value, the default retry behavior is used.

We recommend using the header over other methods to control the retry behavior on special status codes or other situations specific to your API, as it can be used by all consumers of the API and can be configured independently of the version of the SDK.

Timeouts

In addition to retries, our clients also have a default timeout of 60 seconds, which can be configured.

client_settings:
  default_timeout: PT60S # ISO8601 or number of milliseconds. This is 60 seconds

Idempotency key

Idempotency Keys can prevent errors where multiple retried requests are interpreted as separate requests. Our clients retry connection errors and certain status codes by default to create robust integrations, so we recommend that your API supports idempotency keys, especially for critical endpoints.

You can configure idempotency keys by specifying client_settings.idempotency like so:

client_settings:
  idempotency:
    header: 'Idempotency-Key' # or a header you prefer, like 'X-Request-Id'

We send the configured header with a value in the format stainless-retry-{random_uuid} on all non-GET requests. This header is also possible to override in every SDK that we generate.

Enterprise

OpenAPI transformations

If your OpenAPI is not perfect and/or hard to modify because it's generated, we can work with you to identify problems in your OpenAPI spec and fix issues on our end with OpenAPI transforms. Contact our sales.

SSE streaming

Configuring SSE streaming is supported by Stainless for enterprise customers. When configured, we will generate ergonomics bindings for your users to consume streams ergonomically and conveniently, for example in for await loops and more. Contact our sales.

We can also help write and design helpers to make your streamed APIs more powerful for your users.