Skip to main content

Troubleshooting

Fix common problems with the Sandbox MCP connection, sign-in, and work in a sandbox.

Problems fall into two groups. Connection and sign-in problems come from the Sandbox MCP in your AI app. Problems with results come from work in a sandbox. Find your question below and open it for the fix.

Connection and sign-in​

Sign-in or the first attempt to use the Sandbox MCP or its created sandbox fails. What do I do?

Remove the Sandbox MCP (elembio-sandbox) from your AI app, then install it again with the steps on Install the Sandbox MCP. You can also ask your agent to remove it for you.

My AI app does not list the Sandbox MCP. What do I do?

Your AI app loads the Sandbox MCP when a session starts. Start a new chat or session after you install the Sandbox MCP.

Sign-in fails with "dynamic client registration is disabled". What do I do?

The Sandbox MCP was added without the OAuth client ID. Remove the Sandbox MCP, then add it again with the client ID from the connection details.

I signed in, but every attempt to use the sandbox fails. What do I do?

Check that the Sandbox MCP server URL is exactly https://mcp.usw2.elembio.io/sandbox, with no trailing slash.

Claude Code reports "Callback URL mismatch". What do I do?

Add the elembio-sandbox MCP server again with the claude mcp add command from Install the Sandbox MCP. Keep --callback-port 8400 in the command.

Claude Desktop Team/Enterprise​

On Team and Enterprise plans, your organization's admins can control which plugins and connectors you can add in Claude Desktop. If a control blocks the plugin, add the Sandbox MCP manually as a custom connector instead.

Claude Desktop does not let me install the cloud-sandbox plugin. What do I do?

Your organization might limit plugins to approved marketplaces. You might see one of these:

  • Add from a repository is missing or unavailable in Customize > Plugins.
  • Claude Desktop refuses the Elembio/agent-skills repository.
  • Claude Desktop shows a message that your organization manages plugins.

Add the Sandbox MCP manually with the custom connector steps on Install the Sandbox MCP. On Team and Enterprise plans, an Owner adds the connector for your organization first.

The plugin installed, but the elembio-sandbox connector does not appear or connect. What do I do?

Your organization might allow only approved MCP servers from plugins, so Claude Desktop does not start the elembio-sandbox connector.

Add the Sandbox MCP manually with the custom connector steps on Install the Sandbox MCP, or ask your admin to allow https://mcp.usw2.elembio.io/sandbox.

I cannot add a custom connector. What do I do?

On Team and Enterprise plans, only an Owner can add a custom connector. Send your Owner the connection details, and ask them to:

  1. Open Organization settings > Connectors, select Add, then select Custom > Web.
  2. Enter the server URL.
  3. In Advanced settings, enter the OAuth client ID and leave the secret blank.
  4. Select Add.

After the Owner adds the connector, open Customize > Connectors, find elembio-sandbox, and select Connect.

What can my admin change to allow the plugin?

Claude Desktop admins can allow the plugin through managed configuration:

  • Add Elembio/agent-skills to the allowed plugin marketplaces (allowedPluginMarketplaces).
  • Add https://mcp.usw2.elembio.io/sandbox to the allowed plugin MCP servers (allowedPluginMcpServers).

Admins can also publish the plugin to their organization's own marketplace so members see it under the organization's plugins.

Working in a sandbox​

My agent is not consistently using the created sandbox. What do I do?

Using a sandbox requires an ElemBio Catalyst subscription.

Ask your agent to use the sandbox in your prompt, for example: run this analysis in an ElemBio Cloud sandbox.

If your agent keeps running code on your computer, start a new chat, sign in again if the client asks, and repeat the request.

My images and charts are not showing inline. What do I do?

Ask your agent to show the figure in the chat, where you can see it without opening a tool result. A file path, or a figure left inside a collapsed tool result, looks like a missing chart.

When the chat cannot carry the image, your agent returns a short-lived download link to the original file in the sandbox. Open that link.

If your agent reports that the read failed, ask it to fetch the figure again. A file that is too large, or a file type the chat cannot display, stays as that link. Ask your agent before it creates a smaller file, because a smaller file can drop detail.

A long analysis seems stuck. What do I do?

Long-running code keeps running in the sandbox in the background. Ask your agent to check the status of the run.

To stop the run, ask your agent to interrupt it. Interrupting keeps your loaded data and variables.

See also​

Last updated on