Use ChatGPT subscription models in DeepSeek Harness

Give DeepSeek Harness access to OpenAI models from your ChatGPT subscription through Coskey's model gateway, with a dedicated client, provider setup and verification.

This guide uses the official model gpt-6.1-terra as an example and connects DeepSeek Harness to the ChatGPT subscription models available on the Coskey host. DeepSeek Harness receives only a ckg_… client token issued by Coskey. It never receives the ChatGPT OAuth token or another provider’s API key; the model gateway forwards the actual request.

← Back to Model gateway

1. Before you begin

  • Coskey is running and the model gateway is enabled;
  • the Coskey host has a valid ChatGPT subscription login; an API-key login is not a substitute for a ChatGPT subscription login;
  • DeepSeek Harness is installed on the same computer;
  • note the Base URL shown on Coskey → Model gateway. A normal installation usually shows http://127.0.0.1:15721/v1.

The screenshots in this guide come from an isolated instance: 17978 is that instance’s assigned port. Use the address displayed on your own gateway page, not the port in the screenshots.

2. Create a dedicated client in Coskey

Open Coskey → Model gateway and click Create client in the clients section:

The model gateway and the Create client entry point

Use these values:

Field Example Purpose
Client ID deepseek-harness A gateway client dedicated to Harness
Scope This computer Use this when Harness and Coskey run on the same machine
Allowed models gpt-6.1-terra Expose only the ChatGPT subscription model needed here

Saving displays a ckg_… token once. Copy it now and use it as the API key in DeepSeek Harness. Only the plaintext token is kept in the system keychain; the Coskey configuration stores its SHA-256:

The deepseek-harness client and its masked token

The command-line equivalent is:

coskey gateway enable
coskey gateway client add deepseek-harness \
  --scope local \
  --model gpt-6.1-terra

gateway client add also displays the token only once. To add another model later, edit the client row and select the model there. You do not need to recreate the client or replace the token; rotate it only if the token is lost.

3. Add Coskey to DeepSeek Harness

In DeepSeek Harness, open Settings → Models → Add model provider, choose Custom model API, and enter:

Field Value
Provider ID coskey
Display name Coskey
API address The Base URL shown by the Coskey model gateway
API protocol OpenAI Responses
API key The ckg_… client token from step 2

Provider ID coskey is DeepSeek Harness’s local identifier; it is not the Coskey gateway client ID deepseek-harness. Click Get available models, select gpt-6.1-terra and add it:

Entering the Coskey gateway in DeepSeek Harness

The model ID must exactly match the ID returned by the Coskey gateway. When the client allows only gpt-6.1-terra, that should be the only model in the list. You can also add the ID by hand.

4. Start a session and verify

Save the provider, start a new session and confirm that the model picker shows gpt-6.1-terra. Send a minimal test message:

Reply only with “Coskey gateway test passed”.

A matching reply confirms that the Coskey client, model gateway, ChatGPT subscription login and DeepSeek Harness are connected correctly:

DeepSeek Harness receiving a reply through Coskey

5. Subscription login and security boundary

gpt-6.1-terra is an official model. The Coskey gateway reads the ChatGPT subscription login in the host’s CODEX_HOME/auth.json to make the upstream request. It reads the login without refreshing it. The client submits only its ckg_… token; the host OAuth token does not appear in the DeepSeek Harness configuration, gateway API responses, chat history or gateway logs.

Official and third-party models use their respective quotas: an official model uses the ChatGPT account subscription quota on the host, while a third-party model uses its provider’s quota. When you no longer need Harness, disable or revoke the deepseek-harness client in Coskey; other agents’ clients are unaffected.

6. Common questions

Symptom What to check
401 / invalid token Copy a token from the current client; after rotation, enter the new token in Harness
403 If Harness connects over the LAN, the client scope cannot be This computer only
Empty model list Check whether the client allows gpt-6.1-terra, save the client, then fetch the model list again
host_login_required Check that the Coskey process running the gateway uses the expected CODEX_HOME and that it contains a ChatGPT subscription login
UNSUPPORTED_REASONING_EFFORT In Harness, switch to another model and back to gpt-6.1-terra, making sure no high reasoning level remains
Harness connects but requests fail Return to the Coskey gateway page and use Test connection to verify the health endpoint, authentication and a minimal request

If the problem remains, run coskey gateway status and coskey doctor in Coskey. The gateway log is ~/.coskey/logs/gateway.log; it never contains tokens, upstream keys, prompts or completions.

Related: Model gateway, Status and settings.