Use W&B secrets - CoreWeave Docs

Use W&B Secret Manager to make API tokens and other credentials available inside a serverless sandbox. Your application passes secret names to the Sandbox API. When the API creates the sandbox, it resolves the values from your W&B team and injects them as environment variables.

CoreWeave Serverless sandboxes are in public preview.

Before you begin

W&B secret injection requires a sandbox authenticated with a W&B API key. It’s available for serverless sandboxes, but not for sandboxes on your own CoreWeave Kubernetes Service (CKS) cluster. CKS placement uses a CoreWeave API access token. You can’t combine that token with a separate W&B key to resolve secrets for a CKS sandbox.

Sandbox placement	Sandbox credential	W&B secret injection
Serverless	W&B API key	Supported. Use the `wandb` store.
Serverless	CoreWeave API access token	Not supported.
Your CKS cluster	CoreWeave API access token	Not supported. Adding `WANDB_API_KEY` doesn’t enable it.

To follow the examples, you need the following:

A W&B software-as-a-service (SaaS) account with serverless sandbox access.
A secret named HF_TOKEN in your W&B team’s Secret Manager. If the secret doesn’t exist, ask a W&B administrator to add the secret.
A W&B API key with access to that team.

Set WANDB_API_KEY in the environment of the process that creates the sandbox. Replace [WANDB-API-KEY] with your W&B API key:

export WANDB_API_KEY="[WANDB-API-KEY]"

WANDB_ENTITY is optional and defaults to the API key’s default entity. If the secret belongs to another team, set WANDB_ENTITY to that team. Replace [WANDB-TEAM] with the team name:

export WANDB_ENTITY="[WANDB-TEAM]"

Install the client for your language:

Python
TypeScript

Use Python 3.11 or later and install uv if needed. In your Python environment, install the client with its W&B authentication dependency:

uv pip install 'cwsandbox[wandb]'

Use Node.js 22 or later. Install the TypeScript client and tsx to run the examples:

npm install @coreweave/cwsandbox
npm install --save-dev tsx

Save the TypeScript example you want to run as secrets.mts, then run npx tsx secrets.mts.

Inject a secret

Use store="wandb" in Python or store: "wandb" in TypeScript. This store resolves secrets for the authenticated W&B team. You don’t need to register an organization secret store for this flow. The following examples inject HF_TOKEN and check that it’s present without printing its value. In the sandbox, the environment variable defaults to the secret’s name.

Python
TypeScript

Select W&B authentication explicitly with AuthStrategy.WANDB:

from cwsandbox import AuthStrategy, Sandbox, Secret

with Sandbox.run(
  auth=AuthStrategy.WANDB,
  placement_mode="serverless",
  container_image="python:3.11",
  resources={"cpu": "1", "memory": "1Gi"},
  secrets=[Secret(store="wandb", name="HF_TOKEN")],
) as sandbox:
  result = sandbox.exec([
    "python", "-c",
    "import os; assert os.environ.get('HF_TOKEN'); print('Secret is available')",
  ]).result()
  print(result.stdout)

Import the client from the W&B entry point. When you omit runnerIds, the client uses serverless placement:

import { createSandboxClientFromEnv } from "@coreweave/cwsandbox/wandb";

const client = createSandboxClientFromEnv();

const result = await client.withSandbox(
  async (sandbox) => sandbox.commands.run([
    "python", "-c",
    "import os; assert os.environ.get('HF_TOKEN'); print('Secret is available')",
  ]),
  {
    containerImage: "python:3.11",
    resources: { cpu: "1", memory: "1Gi" },
    secrets: [{ store: "wandb", name: "HF_TOKEN" }],
  },
);

console.log(result.stdout);

Both examples print Secret is available and stop the sandbox when the operation finishes. Code running inside the sandbox can read the value from HF_TOKEN. Pass the secret’s name, not its value, in secrets. The server resolves the value. Don’t print it or write it to logs.

Secret reference fields

Use these fields in a Python Secret or a TypeScript secrets entry:

Python field	TypeScript field	Required	Description
`store`	`store`	Yes	Secret store name. Use `wandb` for W&B Secret Manager.
`name`	`name`	Yes	Name of the secret in the selected W&B team’s Secret Manager.
`field`	`field`	No	Top-level key to select from a secret whose value is a JSON object. Omit it to inject the entire secret value.
`env_var`	`envVar`	No	Environment variable that receives the value. Defaults to `name`.

Use a custom environment variable name

To choose the environment variable that receives the secret, set env_var in Python or envVar in TypeScript. The following examples expose the HF_TOKEN secret as HUGGINGFACE_TOKEN:

Python
TypeScript

from cwsandbox import AuthStrategy, Sandbox, Secret

with Sandbox.run(
  auth=AuthStrategy.WANDB,
  placement_mode="serverless",
  container_image="python:3.11",
  secrets=[Secret(
    store="wandb",
    name="HF_TOKEN",
    env_var="HUGGINGFACE_TOKEN",
  )],
) as sandbox:
  result = sandbox.exec([
    "python", "-c",
    "import os; assert os.environ.get('HUGGINGFACE_TOKEN'); print('Secret is available')",
  ]).result()
  print(result.stdout)

import { createSandboxClientFromEnv } from "@coreweave/cwsandbox/wandb";

const client = createSandboxClientFromEnv();

const result = await client.withSandbox(
  async (sandbox) => sandbox.commands.run([
    "python", "-c",
    "import os; assert os.environ.get('HUGGINGFACE_TOKEN'); print('Secret is available')",
  ]),
  {
    containerImage: "python:3.11",
    secrets: [{
      store: "wandb",
      name: "HF_TOKEN",
      envVar: "HUGGINGFACE_TOKEN",
    }],
  },
);

console.log(result.stdout);

Both examples print Secret is available after they check that HUGGINGFACE_TOKEN is present.

Select a field from a structured secret

To inject one value from a JSON object, set field to its top-level key. Nested paths aren’t supported. For a JSON string, the injected value has no surrounding JSON quotes. For this example, add a team secret named DB_CREDENTIALS whose value is a JSON object such as {"password":"[DB-PASSWORD]"}. Replace [DB-PASSWORD] with the password before you save the secret. The examples inject its password field into DB_PASSWORD:

Python
TypeScript

from cwsandbox import AuthStrategy, Sandbox, Secret

with Sandbox.run(
  auth=AuthStrategy.WANDB,
  placement_mode="serverless",
  container_image="python:3.11",
  secrets=[Secret(
    store="wandb",
    name="DB_CREDENTIALS",
    field="password",
    env_var="DB_PASSWORD",
  )],
) as sandbox:
  result = sandbox.exec([
    "python", "-c",
    "import os; assert os.environ.get('DB_PASSWORD'); print('Secret is available')",
  ]).result()
  print(result.stdout)

import { createSandboxClientFromEnv } from "@coreweave/cwsandbox/wandb";

const client = createSandboxClientFromEnv();

const result = await client.withSandbox(
  async (sandbox) => sandbox.commands.run([
    "python", "-c",
    "import os; assert os.environ.get('DB_PASSWORD'); print('Secret is available')",
  ]),
  {
    containerImage: "python:3.11",
    secrets: [{
      store: "wandb",
      name: "DB_CREDENTIALS",
      field: "password",
      envVar: "DB_PASSWORD",
    }],
  },
);

console.log(result.stdout);

Both examples print Secret is available without printing the password. Selecting a field fails if the secret isn’t a JSON object, the key doesn’t exist, or its value is null.

Reuse secret references

For Python sessions, put shared secret references in SandboxDefaults. For TypeScript, reuse a secrets array when you create sandboxes:

Python
TypeScript

from cwsandbox import AuthStrategy, SandboxDefaults, Secret, Session

defaults = SandboxDefaults(
  auth=AuthStrategy.WANDB,
  placement_mode="serverless",
  secrets=(Secret(store="wandb", name="HF_TOKEN"),),
)

with Session(defaults) as session:
  sandbox = session.sandbox()
  result = sandbox.exec([
    "python", "-c",
    "import os; assert os.environ.get('HF_TOKEN'); print('Secret is available')",
  ]).result()
  print(result.stdout)

The Python client merges session defaults with per-sandbox secrets and ignores exact duplicates. If different references target the same environment variable, the client raises ValueError before it creates the sandbox. For the Python type signature, see the Secret reference.

import { createSandboxClientFromEnv } from "@coreweave/cwsandbox/wandb";

const client = createSandboxClientFromEnv();
const secrets = [{ store: "wandb", name: "HF_TOKEN" }];

const result = await client.withSandbox(
  async (sandbox) => sandbox.commands.run([
    "python", "-c",
    "import os; assert os.environ.get('HF_TOKEN'); print('Secret is available')",
  ]),
  { containerImage: "python:3.11", secrets },
);

console.log(result.stdout);

Troubleshoot secret injection

If the sandbox fails to start with a secret reference, use the error text to check authentication, the selected team, and placement:

Error text	What to check
`secret store "wandb" not found`	Check whether the client selected CoreWeave authentication. For serverless placement, select `auth=AuthStrategy.WANDB` in Python or the `@coreweave/cwsandbox/wandb` entry point in TypeScript. Setting `WANDB_API_KEY` alone doesn’t switch a CoreWeave-authenticated client to W&B. You don’t need to register a store.
`one or more W&B secrets not found`	Check the secret name and the selected W&B team. Set `WANDB_ENTITY` if the secret belongs to a team other than the API key’s default entity, and verify that your W&B account can access it.
`entity not found`	Check `WANDB_ENTITY` for a typo and verify that the team exists and your W&B account can access it.
`no eligible runner is available on this replica right now; retry shortly`	If you selected a CKS runner with W&B authentication, this is an unsupported combination. Retrying won’t enable it. Use W&B authentication with serverless placement for secret injection. For a serverless request, this error can indicate temporary runner unavailability.

W&B secret injection isn’t available for CKS placement, including when you set both a CoreWeave API access token and a W&B API key.

​Before you begin

​Inject a secret

​Secret reference fields

​Use a custom environment variable name

​Select a field from a structured secret

​Reuse secret references

​Troubleshoot secret injection

Before you begin

Inject a secret

Secret reference fields

Use a custom environment variable name

Select a field from a structured secret

Reuse secret references

Troubleshoot secret injection