Skip to main content
Run OpenCode against a repository in a CoreWeave sandbox. Send a task from your terminal, check the resulting files, and save the workspace for later. The agent process, workspace, and commands run in the sandbox. Model requests go to your configured provider.

​
Prerequisites

Before you begin, you need the following:
  • A CoreWeave API access token with the SANDBOX_USER Identity and Access Management (IAM) action for sandbox access. See Choose and get credentials.
  • Access to a model through W&B Serverless Inference or another OpenCode provider. Provider access and billing are separate from sandbox access.
  • A public repository URL. Private repositories require Git credentials inside the sandbox.
  • cws-agent installed by following its installation instructions.
In your local terminal, replace [API-ACCESS-TOKEN] with your CoreWeave API access token. To use W&B Serverless Inference, also replace [WANDB-API-KEY] with your W&B API key:

​
Launch OpenCode with cws-agent

Use cws-agent to install OpenCode in a sandbox, clone your repository into the /workspace/project directory, and configure a snapshot mount for file system snapshots. Choose one of the following provider options. Replace [SANDBOX-NAME] with a session name containing 1 to 40 lowercase letters, digits, or hyphens. Start the name with a letter or digit. Replace [REPOSITORY-URL] with your public repository URL. W&B Serverless Inference is the preferred provider for this guide. You can use it to access hosted open models. You can use any model and provider supported by OpenCode instead. Follow Use another model provider. Complete the W&B Serverless Inference prerequisites, including a W&B project, and check your available inference credits. This example selects GLM 5.2 explicitly:
The --wandb preset configures W&B’s inference endpoint, forwards WANDB_API_KEY into the sandbox, and limits OpenCode to that provider. The agent uses this key for inference, and cws-agent uses your CoreWeave API access token for sandbox access. To choose a different W&B model, replace zai-org/GLM-5.2 with an ID from the available models list.

​
Use another model provider

To use another provider, omit --wandb and --wandb-model when you create the sandbox. Export that provider’s API key locally. For example, choose one of these alternatives and replace the key placeholder:
Replace [PROVIDER-ID]/[MODEL-ID] in both fields with your OpenCode model identifier, using the anthropic/ or openai/ provider prefix as appropriate. The small_model field selects the model for lightweight tasks such as title generation. See OpenCode models.
If they’re set, cws-agent forwards ANTHROPIC_API_KEY and OPENAI_API_KEY. For providers whose keys aren’t forwarded automatically, follow their OpenCode provider configuration and use --env-passthrough [VARIABLE-NAME] for additional environment variables. Both launch options print session ready. after setup. --detach leaves the sandbox running without opening the terminal interface or importing local skills and Model Context Protocol (MCP) configuration. The 2-hour lifetime is a wall-clock limit, including setup. Stop compute when you finish.

​
Run a task and check the result

From your local terminal, ask OpenCode to write and read a file:
This runs OpenCode without an interactive terminal. --permission-mode native keeps OpenCode’s configured permissions. Without it, cws-agent defaults to accepting approval requests. If your OpenCode policy requires approval for the task, use an interactive session to approve it. See OpenCode permissions. Check the actual file independently of the agent’s response:
The output should be Hello from CoreWeave. Before you stop compute, retrieve a local copy:

​
Optional: Work interactively

From an interactive terminal on macOS or Linux, open OpenCode in the same workspace:
The --no-config-sync flag skips local skills and MCP imports. Send prompts and approve actions in OpenCode, then use /exit to return to your local terminal. When you exit OpenCode, the sandbox continues to run. Repeat the connection command to open OpenCode again. It doesn’t automatically select a previous conversation. For conversation history and resume commands, see cws-agent sessions.

​
Save, stop, and restore

After the agent finishes, snapshot the workspace and stop compute:
down saves the /workspace directory before stopping the sandbox. If the snapshot fails, the sandbox remains running. Inspect the error and retry after resolving it. To stop without waiting for an explicit snapshot, run:
Don’t rely on this command to save changes. Copy results elsewhere or confirm a completed snapshot before stopping. Restore requires a completed snapshot. To return to a workspace saved with W&B inference, make sure WANDB_API_KEY is still exported in your local terminal, then run:
Use the same --wandb-model value you launched with. If you omit it while using --wandb, restore selects the preset’s default model. For another provider, re-export its key and OPENCODE_CONFIG_CONTENT, then run:
Restore creates a new sandbox from the saved workspace. Supply environment-only credentials again. Snapshots can contain saved OpenCode logins and conversation data, but they don’t preserve running processes. Run the file check again to confirm the sandbox-proof.txt file survived. Continue with cws-agent run or the interactive connection command. Stop the restored sandbox with cws-agent down when you finish.

​
Troubleshoot

If setup or a task fails, use the following checks:
  • Sandbox authentication fails: Check your CoreWeave API access token and its SANDBOX_USER IAM action.
  • W&B model requests fail: Check inference access, credits, and the model ID. Sandbox access alone doesn’t establish inference access.
  • Another provider isn’t available: Start a new session without --wandb. The W&B preset is saved in the workspace, so omitting --wandb on restore doesn’t remove it. For an existing workspace, see W&B preset configuration.
  • A headless task needs approval: Connect interactively and use OpenCode’s approval prompts before continuing.

​
Next steps

For more information, see these guides:
Last modified on October 5, 2026