Connect DeepSeek Harness to Flex AI

DeepSeek Harness (dsh) is an open-source agent harness developed by DeepSeek AI. Its local Web UI provides session and workspace management, agent presets, plugins and custom model providers. Flex AI (Flex247) can be connected through the OpenAI Responses protocol.

• Official GitHub repository
• Official README
• Official model-provider guide

Developer preview

DeepSeek Harness is still a developer preview. Later releases may change its interface, configuration fields or behavior. The screenshots and steps on this page reflect the current version.

Do not expose it to the public Internet

Do not publish DeepSeek Harness through public port forwarding, an open reverse proxy or a public bind address. It may have access to API keys, session content, code and local tool permissions.

Keep the default service on 127.0.0.1:3080. For access from another device, use Tailscale or a similar trusted private-network solution and allow only your own trusted devices.

1. Install and start DeepSeek Harness

Install Node.js, then run the command provided by the project:

npx @deepseek-ai/dsh web

Open the local Web UI at:

http://127.0.0.1:3080/

Keep the terminal that runs this command open. If the page is unavailable, confirm that the command is still running and port 3080 is not already in use.

2. Create a dedicated Flex247 key

Sign in to Flex247, open API Keys, and create a key dedicated to DeepSeek Harness, such as DeepSeek-Harness-Windows.

Protect the key

An API key grants access to your account usage. Never place a real key in screenshots, public documentation, chats or source repositories. The API-key field is intentionally blank in this guide's screenshot.

3. Open Add Custom Provider

In DeepSeek Harness:

1. Click Settings in the lower-left corner.
2. Open Models.
3. Click Add Custom Provider.

Figure 1: Open Settings → Models and click Add Custom Provider. The screenshot shows the Chinese interface.

4. Enter the Flex AI provider

Use these values:

DeepSeek Harness field	Value
Provider ID	flexai
Display name	Flex AI
API URL	https://api.noflex.work/v1
API protocol	openai-responses
API key	Your dedicated Flex247 key
Model catalog	Click Add Model and enter an exact model ID from Model Square

The model ID must exactly match Model Square. For example, start with gpt-5.6-luna and use GPT 5.6 Luna as its display name.

Figure 2: Example Flex AI custom provider. The API-key field is intentionally blank.

Keep the URL and protocol separate

Set the API URL only through /v1. Do not append /responses or duplicate the suffix as /v1/v1. The openai-responses selection controls the request protocol.

Treat the Provider ID as permanent

The Provider ID is referenced by requests, sessions, model defaults and credentials. To change it, create a new provider and remove the old one. The display name, URL, protocol, key and models remain editable.

Currently known incompatible models

Based on current compatibility testing, do not use grok-4.5-free or grok-4.6-free in DeepSeek Harness. Because their architecture is incompatible with the harness, these models trigger an infinite loop. This is not an official DeepSeek Harness compatibility statement.

If either model is selected accidentally, stop the current task immediately, switch to a compatible model and test again in a new session.

5. Create the provider and test

1. Add at least one model, then click Create Provider.
2. Return to the session view and click Select Model below the prompt box.
3. Select Flex AI and the model you added.
4. Start a new session and send a short, read-only test request.
5. Confirm the successful request and expected model ID in Flex247 Usage Logs.

Provider and model routing changes take effect on the next request; DeepSeek Harness does not need to be restarted.

After entering the URL and key, you can also click Get Available Models before saving. If the upstream does not expose a compatible GET /models endpoint, add the exact model ID manually.

Troubleshooting

Get Available Models returns 401

Verify that the API key is complete and active, with available quota and permission for the target model.

404 or Responses error

• The API URL must be https://api.noflex.work/v1.
• Select openai-responses as the API protocol.
• Do not append /responses to the API URL.

The model is missing from the selector

Edit the Flex AI provider and confirm that the model catalog contains at least one exact model ID. Use Add Model when automatic discovery is unavailable.

The session repeats indefinitely

Stop the task immediately, confirm that neither grok-4.5-free nor grok-4.6-free is selected, and start a new session with a compatible model.