Skip to main content
In Claude Desktop, routing traffic through an LLM gateway is called third-party inference. The setup differs depending on whether you are configuring a single machine or rolling out the gateway to an organization:

Prerequisites

Before you can configure Claude Desktop to use the Connect AI LLM Gateway, you must register a provider, create a routing rule, create a project, and issue a virtual key. See LLM Gateway for details.
A Connect AI Personal Access Token (PAT) is not accepted by the LLM Gateway. Only a virtual key authenticates requests.
You must also run a recent version of Claude Desktop. Older versions do not include the third-party inference settings. If you do not see Enable Developer Mode under Help > Troubleshooting, update Claude Desktop and relaunch it.

Limitations

After you configure Claude Desktop to use a gateway, your connectors, artifacts, and projects are no longer available in Claude Desktop. Chat, Cowork, and Code sessions are limited to the models and routing rules of the gateway.
These limitations are inherent to how Claude Desktop handles third-party inference. They apply to any LLM gateway and are not specific to the Connect AI LLM Gateway. To restore these features, remove the gateway configuration and restart Claude Desktop.

End-User Configuration

Use these steps to configure Claude Desktop on a single machine. If your organization manages your device, your IT admin may already have configured the gateway for you. See IT Admin Configuration.
1
Open Claude Desktop.
2
Enable Developer Mode by selecting Help > Troubleshooting > Enable Developer Mode from the top menu. On Windows, open the application menu (☰) in the top-left corner first.
Enable Developer Mode
3
Select Developer > Configure Third-Party Inference.
4
On the Connection tab, set Inference provider to Gateway.
5
Configure the gateway connection as follows:
  • Credential kind–Static API key.
  • Gateway base URL–https://cloud.cdata.com/gateway. Do not include /v1. Claude Desktop appends the path automatically.
  • Gateway API key–Your virtual key from the Virtual Keys tab of the Connect AI LLM Gateway page.
  • Gateway auth scheme–x-api-key.
6
Click Apply Changes, then click Save & Restart.
After Claude Desktop restarts, Chat, Cowork, and Code sessions route through the Connect AI LLM Gateway. The gateway selects the provider and model based on your project’s routing rule.

IT Admin Configuration

For an organization, do not ask each user to follow the end-user steps. Instead, push a managed configuration to your devices through your device management (MDM) tool. This is the approach Anthropic supports for Claude Desktop. IT configures it through your MDM tool, not in the Claude admin console.
1
On one admin workstation, complete steps 1 through 5 of End-User Configuration to enter the gateway settings. Use a virtual key issued for the project that your organization shares.
2
In the Configure Third-Party Inference window, click Export and select the format for your device management tool:
  • .mobileconfig–macOS, for Jamf, Kandji, or Mosyle.
  • .reg–Windows, for Intune or Group Policy.
  • Bootstrap server JSON–Also available.
3
Push the exported profile to your devices through your MDM tool.
4
Push the Claude Desktop installer.
5
Optional: Give each user their own virtual key instead of the shared one. See Optional: Per-User Virtual Keys.
Push the profile before the installer. Users then land directly in gateway mode the first time they open Claude Desktop.
On managed devices, the Configure Third-Party Inference window opens read-only. Users cannot change the gateway, and they do not need to enable Developer Mode.

Optional: Per-User Virtual Keys

The exported profile the admin creates carries one virtual key. Every device shares that key, so the gateway sees all traffic as a single user. Virtual keys are tied one-to-one to users so that you can track usage per person. To keep that tracking, give each user their own key using one of the methods below. Every method requires you to create one virtual key per user in Connect AI.
Last modified on November 9, 2026