Guide to Using Claude Desktop with ZenMux
Claude Desktop supports third-party inference, which routes model inference requests to a third-party platform. With ZenMux's Anthropic-compatible protocol, you can use the models aggregated by ZenMux in the Claude Desktop Cowork / Code experience instead of relying only on Anthropic's default service.
Compatibility Notes
Claude Desktop's third-party inference configuration is designed for Bedrock, Vertex AI, Azure AI Foundry, or an LLM gateway compatible with /v1/messages. ZenMux provides an Anthropic Messages-compatible endpoint and can be connected through Gateway mode.
ZenMux Anthropic-compatible Base URL:
https://zenmux.ai/api/anthropicFor more models that support the Anthropic Messages protocol, see the ZenMux model list.
Note
Claude Desktop third-party inference requires a recent desktop client version. Anthropic describes it as a third-party platform / enterprise deployment configuration mode. If your client does not show the relevant menu, update Claude Desktop first.
Prerequisites
Before you begin, make sure you have:
- The latest Claude Desktop installed
- A ZenMux API Key
- The model slug you want to use, such as
anthropic/claude-sonnet-4.6
Get a ZenMux API Key
ZenMux offers two types of API Keys. Choose based on your use case:
Best for: personal development, learning, and Vibe Coding
Key format: sk-ss-v1-xxx
Get it from: https://zenmux.ai/platform/subscriptionBest for: production, commercial products, and enterprise applications
Key format: sk-ai-v1-xxx
Get it from: https://zenmux.ai/platform/pay-as-you-goAPI Key Guidance
Subscription API Keys are better suited to personal development and learning. For production, commercial products, or enterprise applications, use a pay-as-you-go API Key.
Configure Claude Desktop
Step 1: Sign out or stay signed out
After opening Claude Desktop, sign out of any existing Claude account first, or complete the third-party inference setup while signed out.
According to Claude's official guidance, third-party inference can be configured without logging in to a Claude account. Once configured, Claude Desktop sends inference requests to your configured Gateway.
Step 2: Enable Developer Mode
In the Claude Desktop menu bar, select:
Help → Troubleshooting → Enable Developer ModeAfter you confirm, Claude Desktop restarts. Once it relaunches, a Developer menu appears in the menu bar.
Step 3: Open the third-party inference configuration
In the menu bar, select:
Developer → Configure third-party inferenceThis opens Claude Desktop's third-party inference configuration UI.
Step 4: Configure the Gateway connection
In the Connection section, choose Gateway, then fill in:
| Field | Value |
|---|---|
| Gateway Base URL | https://zenmux.ai/api/anthropic |
| Gateway API Key | Your ZenMux API Key |
| Gateway Auth Scheme | Bearer |
| Gateway Extra Headers | Usually leave blank. Add values only if your gateway requires tenant routing or enterprise headers. |
Relationship to Claude Code
If you have configured Claude Code before, this is the same value as ANTHROPIC_BASE_URL:
export ANTHROPIC_BASE_URL="https://zenmux.ai/api/anthropic"Claude Desktop does not require you to manually write shell environment variables. Enter the same Base URL in the third-party inference configuration UI.
Step 5: Configure the model list
Go to Identity & Models, or the model-related configuration section, and configure the models you want to use in Claude Desktop.
Claude Desktop treats the Gateway Base URL and the model list as separate configuration concerns:
- Gateway Base URL decides where inference requests are sent, such as
https://zenmux.ai/api/anthropic - Model List / Model Discovery decides which models appear in the model picker
Option A: Configure Model List manually
This is the most stable option if you want explicit control over which models appear in Claude Desktop.
Add model slugs to Model List. The model slug is sent to ZenMux as the actual model value in requests.
Examples:
anthropic/claude-sonnet-4.6
anthropic/claude-opus-4.6
openai/gpt-5.5
google/gemini-3.1-pro-previewUse exact model slugs
Use model slugs from the ZenMux model list that support the Anthropic Messages protocol. If a model name is misspelled, Claude Desktop may fail to start a session or return a model-not-found error at runtime.
Option B: Use Model Discovery
If Model Discovery is enabled, Claude Desktop attempts to fetch the model list from the Gateway model-list endpoint. For the ZenMux Anthropic-compatible endpoint, this is:
https://zenmux.ai/api/anthropic/v1/modelsThis option requires less manual setup and is useful for quick tests. However, in some Claude Desktop versions, discovered models may be filtered, alias-resolved, or reordered by the client. If the model picker shows fewer models than the ZenMux /v1/models response, use a manual Model List instead.
Option C: Bulk configure inferenceModels in JSON
If you need to import many models at once, click View as JSON in the configuration UI and write inferenceModels directly.
Example:
{
"modelDiscoveryEnabled": false,
"inferenceModels": [
{
"name": "anthropic/claude-sonnet-4.6",
"labelOverride": "Claude Sonnet 4.6"
},
{
"name": "qwen/qwen3.7-max",
"labelOverride": "Qwen: Qwen3.7-Max"
}
]
}Notes:
nameis the model slug actually sent to ZenMuxlabelOverrideonly changes the display name in Claude Desktop- When
modelDiscoveryEnabledis disabled,inferenceModelsis used as the fixed model list
Community tools
If you need to generate inferenceModels from the ZenMux /v1/models endpoint, you can refer to community projects such as Lovmux or other configuration management tools.
Step 6: Enable 1M Context only when supported
Claude Desktop's model configuration UI may expose advanced options such as 1M Context. Enable these only when the backend model and provider actually support the corresponding context length.
If you are unsure whether the model supports long context, keep the option disabled first. Complete the basic call verification, then adjust it later.
Step 7: Apply the configuration and restart
Click Apply locally at the bottom of the configuration UI. Claude Desktop writes the local configuration and restarts automatically.
After the restart, you should see entries such as Cowork / Code and the model picker should show the model list you configured. Choose a model to start using Claude Desktop through ZenMux.
Verify the Setup
Start with a simple message:
Introduce the model you are currently using in one sentence.If the request succeeds, test tasks closer to your real workflow, such as:
- Ask Cowork to summarize a local document
- Ask Code to open a project and explain its directory structure
- Switch between models to compare latency and output quality
You can also check the ZenMux request logs to confirm that the request went through ZenMux.
Troubleshooting
The Developer menu is missing
Check that:
- Claude Desktop has been updated to a version that supports third-party inference
- You enabled Developer Mode through
Help → Troubleshooting → Enable Developer Mode - You fully restarted Claude Desktop after enabling it
Which Base URL should I use?
For ZenMux, use:
https://zenmux.ai/api/anthropicDo not use the OpenAI-compatible endpoint https://zenmux.ai/api/v1. Claude Desktop Gateway mode expects an Anthropic Messages-compatible /v1/messages interface.
API Key or authentication errors
Check that:
- The API Key was copied completely, without extra spaces or line breaks
- Gateway Auth Scheme is set to
Bearer - The API Key is still valid and the account has available quota or balance
- If using a subscription key, the usage fits the subscription scope
Model unavailable or model list is empty
Check that the model slug in Model List is exact and that the model supports the Anthropic Messages protocol.
Copy model slugs from the ZenMux model list.
Calls fail after enabling 1M Context
1M Context only works for models whose backend explicitly supports long context. If calls fail after enabling it, disable the option first, verify the basic model call, then retest with a model that supports 1M context.