Use Your Subscription-Authenticated Cloud LLMs in DeepSeek Harness — Without BYOK

Architecture

We want to create this path:

DeepSeek Harness
       │
       │ OpenAI-compatible HTTP
       ▼
Local Codex Proxy
       │
       │ Codex app-server
       ▼
Official Codex Runtime
       │
       │ OAuth
       ▼
ChatGPT / Codex Subscription
       │
       ▼
GPT-5.6 Sol

DeepSeek Harness therefore does not need to understand Codex OAuth itself. It simply sees another OpenAI-compatible model endpoint.

The proxy used here delegates authentication to the official Codex runtime rather than copying or managing your Codex OAuth tokens itself.

1. Install DeepSeek Harness

If you do not already have Harness running, the simplest local launch is:

npx @deepseek-ai/dsh web

The Web UI runs locally on port 3080 by default.

DeepSeek Harness is currently in developer preview, so expect its configuration and extension APIs to continue changing.

2. Install the official Codex CLI

Install or update Codex:

npm install -g @openai/codex@latest

Then authenticate:

codex login

Codex opens the official authentication flow and associates the CLI with your ChatGPT/Codex account.

OpenAI currently requires Codex CLI 0.144.0 or newer for GPT-5.6 access, and availability of Sol depends on your account and rollout status.

You can confirm your version with:

codex --version

3. Verify GPT-5.6 Sol first

Before involving Harness or the proxy, make sure the model works directly through Codex:

codex exec \
  --model gpt-5.6-sol \
  --sandbox read-only \
  --ephemeral \
  --skip-git-repo-check \
  "Reply exactly: SOL_OK"

You should receive:

SOL_OK

If this fails, fix the Codex authentication or model-access issue first.

The model identifier currently used by Codex is gpt-5.6-sol, and recent Codex clients can execute it directly when the account has access.

4. Install codex-proxy

For this example we will use the open-source codex-proxy project.

If you have GitHub CLI installed:

gh repo clone mehdic/codex-proxy
cd codex-proxy

Then install and build it:

npm install
npm run build

codex-proxy wraps the official codex app-server and exposes OpenAI-style endpoints including:

/v1/chat/completions
/v1/responses
/v1/models

By default it binds only to 127.0.0.1:3466, which is exactly what we want for a local integration.

5. Start the proxy with GPT-5.6 Sol

Start it with Sol as the default and pre-warmed model:

CODEX_PROXY_DEFAULT_MODEL=gpt-5.6-sol \
CODEX_PROXY_PREWARM_MODELS=gpt-5.6-sol \
npm start

Your OpenAI-compatible base endpoint is now:

http://127.0.0.1:3466/v1

The proxy's public model catalog may lag behind the newest Codex models, but it passes requested model names through to codex app-server. Therefore we can manually configure gpt-5.6-sol in Harness when necessary.

6. Keep it alive with tmux

If you close the terminal running the proxy, the process will normally stop.

A simple solution is to run it inside tmux.

On macOS:

brew install tmux

Create a dedicated session:

tmux new -s codex-proxy

Then start the proxy inside that session:

cd codex-proxy

CODEX_PROXY_DEFAULT_MODEL=gpt-5.6-sol \
CODEX_PROXY_PREWARM_MODELS=gpt-5.6-sol \
npm start

Detach from the session with:

Ctrl+B
D

The proxy continues running. Later, reconnect with:

tmux attach -t codex-proxy

This is particularly useful for long-running Harness agent sessions or remote development machines.

7. Verify the proxy

First check its health:

curl http://127.0.0.1:3466/health

Then make an actual model request:

curl http://127.0.0.1:3466/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-sol",
    "messages": [
      {
        "role": "user",
        "content": "Reply exactly: SOL_PROXY_OK"
      }
    ]
  }'

You should receive a normal OpenAI-compatible response containing SOL_PROXY_OK.

At this point the entire lower half of the architecture works:

HTTP Request
    ↓
codex-proxy
    ↓
Codex app-server
    ↓
Codex OAuth
    ↓
GPT-5.6 Sol

8. Connect the proxy to DeepSeek Harness

Now open DeepSeek Harness → Settings → Models → Add a custom provider.

Harness officially supports custom providers using a provider ID, base URL, API protocol, credential and model catalog.

Configure:

Provider ID:
codex-proxy

Base URL:
http://127.0.0.1:3466/v1

Protocol:
openai-completions

Model:
gpt-5.6-sol

If Harness requires a credential value, use a local placeholder such as:

local

This is not an OpenAI API key.

The authentication chain still lives here:

Codex CLI
   ↓
Codex OAuth
   ↓
ChatGPT account

Rather than here:

DeepSeek Harness
   ↓
OPENAI_API_KEY

If /v1/models does not advertise Sol yet, simply enter gpt-5.6-sol manually in the Harness model catalog. Harness explicitly supports manually declared models for custom providers.

9. Select GPT-5.6 Sol in Harness

After saving the provider, return to a Harness session and select:

codex-proxy / gpt-5.6-sol

Send a simple test:

Explain which model provider you are using and return the text HARNESS_SOL_OK.

The complete request path is now:

DeepSeek Harness Agent
          │
          ▼
codex-proxy
          │
          ▼
Codex app-server
          │
          ▼
Codex OAuth Session
          │
          ▼
GPT-5.6 Sol

You have effectively turned an authenticated cloud-model runtime into a model provider that DeepSeek Harness can consume.

Security: keep the proxy local

Do not casually expose port 3466 to your LAN or the public Internet.

codex-proxy deliberately binds to 127.0.0.1 by default and delegates authentication to your existing Codex session. Its own documentation specifically recommends keeping the service private because Codex is an agent runtime, not merely a text-generation endpoint.

For a local workstation, keep:

CODEX_PROXY_HOST=127.0.0.1

Also remember that the proxy exposes Codex-native capabilities according to its configured sandbox policy. Its conservative default is read-only; use broader filesystem or shell permissions only when you understand the consequences.