Skip to main content

Connect Codex to DIAL

Introduction​

From this tutorial, you will learn how to point Codex, the command-line coding agent, at DIAL and watch a DIAL-hosted model reply. Codex speaks only the OpenAI Responses API, so how it reaches DIAL depends on your deployment: one that exposes the Responses API natively connects directly, while a Chat-Completions-only deployment needs a small translating proxy. Step 1 shows which case you are in. By the end, codex exec sends a prompt and prints the model's reply, routed through your DIAL deployment with DIAL's roles, quotas, and cost tracking applied.

Prerequisites​

Step 1: Find out which path your deployment needs​

Ask DIAL for the deployment's metadata and read the responses_api flag:

curl -s -H "Api-Key: <DIAL_API_KEY>" \
  "https://<YOUR_DIAL_HOST>/openai/deployments/gpt-5.4-2026-03-05" \
  | grep -o '"responses_api":[^,]*'

On Windows (PowerShell), call curl.exe — PowerShell's curl is an Invoke-WebRequest alias — and pipe to findstr:

curl.exe -s -H "Api-Key: <DIAL_API_KEY>" `
  "https://<YOUR_DIAL_HOST>/openai/deployments/gpt-5.4-2026-03-05" | findstr responses_api

If curl is unavailable, Invoke-RestMethod reads the flag directly and prints True or False:

(Invoke-RestMethod -Headers @{ "Api-Key" = "<DIAL_API_KEY>" } `
  "https://<YOUR_DIAL_HOST>/openai/deployments/gpt-5.4-2026-03-05").features.responses_api
  • "responses_api":true — follow The direct path and ignore the proxy.
  • "responses_api":false — your deployment is chat-only; follow The proxy path.

Note: The responses_api flag only means anything on DIAL v0.45+ (refer to Check your DIAL version); on older DIAL every deployment is chat-only, so take the proxy path. Even on v0.45+, many deployments are still chat-only, so check the flag rather than assume. You can also test it functionally: a POST to https://<YOUR_DIAL_HOST>/openai/v1/responses returns 200 on a Responses-capable deployment and 503 on a chat-only one.

The direct path​

Follow this path when Step 1 reported responses_api: true.

Step 2: Tell Codex about DIAL​

Open ~/.codex/config.toml (on Windows, %USERPROFILE%\.codex\config.toml; create it if it does not exist) and paste the following. The model_provider and model_providers keys only take effect in this user-level file.

model = "gpt-5.4-2026-03-05"      # the DIAL deployment id; Codex sends it to DIAL as the model
model_provider = "dial"
disable_response_storage = true   # DIAL doesn't store responses, so this must be on

[model_providers.dial]
name = "DIAL (direct)"
base_url = "https://<YOUR_DIAL_HOST>/openai/v1"     # Codex appends /responses
wire_api = "responses"            # the only valid value — Codex's old "chat" wire was removed
env_key = "DIAL_API_KEY"
env_http_headers = { "Api-Key" = "DIAL_API_KEY" }   # DIAL authenticates on Api-Key
requires_openai_auth = false                         # don't enforce the sk- prefix check
query_params = { api-version = "2025-04-01-preview" }  # Azure-backed deployments require this

Note: Put the api-version in query_params, not in base_url. Codex appends /responses to base_url, so a query string there lands in the wrong place. Drop the query_params line if your DIAL host does not require an Azure API version — refer to Azure-backed deployment settings.

Step 3: Set your API key​

Bash or Zsh:

export DIAL_API_KEY="<DIAL_API_KEY>"

PowerShell:

$env:DIAL_API_KEY = "<DIAL_API_KEY>"

Step 4: Check the wiring​

codex doctor

Your config loads cleanly and a row appears for the dial provider — config.toml parse ok, auth ✓ with provider auth env var DIAL_API_KEY (present), and wire API responses. The model line names your deployment, for example gpt-5.4-2026-03-05 · dial.

Step 5: Run Codex​

codex exec "Reply with exactly: PONG"

You see PONG printed back, routed Codex → DIAL with no proxy in between. Run plain codex for the interactive session you will use day to day.

The direct path is complete. Skip to Additional Information.

The proxy path​

Follow this path when Step 1 reported responses_api: false. Put a small translating proxy between Codex and DIAL. The proxy (responses-proxy) converts Codex's Responses calls into the Chat Completions calls DIAL understands and forwards your Api-Key header to DIAL.

Step 2: Build and run the proxy​

There is no published image, so build it from the repo — a one-time Rust build:

git clone https://github.com/chutesai/responses-proxy && cd responses-proxy

Warning: Do not use the project's install_codex.sh one-liner for DIAL. It installs chutes' own fork of Codex and points it at their hosted proxy (responses.chutes.ai, whose backend is chutes — not DIAL), and the config it writes omits both disable_response_storage and the Api-Key header DIAL needs. Build and run the proxy yourself, as below.

Point the proxy's BACKEND_URL at your DIAL chat deployment — the deployment id lives in the URL. If your DIAL's Chat Completions endpoint requires an Azure API version, append ?api-version=2025-04-01-preview to BACKEND_URL (refer to Azure-backed deployment settings). Start it either way:

# Option A — Docker Compose (the maintainers' path). --build compiles the image;
# naming the service skips the bundled Caddy/TLS sidecar you don't need locally.
BACKEND_URL="https://<YOUR_DIAL_HOST>/openai/deployments/gpt-5.4-2026-03-05/chat/completions" \
  docker compose up --build openai-responses-proxy
# Option B — plain Docker (build once, then run).
docker build -t responses-proxy .
docker run -d --name responses-proxy -p 8282:8282 \
  -e BACKEND_URL="https://<YOUR_DIAL_HOST>/openai/deployments/gpt-5.4-2026-03-05/chat/completions" \
  responses-proxy

On Windows (PowerShell), the inline VAR=… command form and the \ line-continuation do not work. Set the variable with $env: first and use a backtick to continue lines:

$env:BACKEND_URL = "https://<YOUR_DIAL_HOST>/openai/deployments/gpt-5.4-2026-03-05/chat/completions"
docker compose up --build openai-responses-proxy          # Option A
# or, after `docker build -t responses-proxy .` :
docker run -d --name responses-proxy -p 8282:8282 `
  -e BACKEND_URL="$env:BACKEND_URL" responses-proxy        # Option B

The proxy now listens on http://0.0.0.0:8282. Confirm it is up:

curl -s -o /dev/null -w "%{http_code}\n" http://0.0.0.0:8282/health

On Windows, use curl.exe, or Invoke-WebRequest if curl is unavailable:

curl.exe -s -o NUL -w "%{http_code}`n" http://0.0.0.0:8282/health
(Invoke-WebRequest -UseBasicParsing http://0.0.0.0:8282/health).StatusCode

You should see 200.

Step 3: Tell Codex about the proxy​

Open ~/.codex/config.toml (on Windows, %USERPROFILE%\.codex\config.toml; create it if it does not exist) and paste the following. The model_provider and model_providers keys only take effect in this user-level file.

model = "gpt-5.4"                 # the DIAL deployment id you set in BACKEND_URL above
model_provider = "dial"
disable_response_storage = true   # DIAL doesn't store responses, so this must be on

[model_providers.dial]
name = "DIAL (via responses-proxy)"
base_url = "http://0.0.0.0:8282/v1"   # the proxy; Codex appends /responses
wire_api = "responses"            # the only valid value — Codex's old "chat" wire was removed
env_key = "DIAL_API_KEY"
env_http_headers = { "Api-Key" = "DIAL_API_KEY" }   # forwarded through the proxy to DIAL
requires_openai_auth = false                         # don't enforce the sk- prefix check

Step 4: Set your API key​

Bash or Zsh:

export DIAL_API_KEY="<DIAL_API_KEY>"

PowerShell:

$env:DIAL_API_KEY = "<DIAL_API_KEY>"

Step 5: Check the wiring​

Before running a real prompt, ask Codex to inspect its own setup:

codex doctor

Your config loads cleanly and a row appears for the dial provider — config.toml parse ok, auth ✓ with provider auth env var DIAL_API_KEY (present), and wire API responses. The model line names your deployment, for example gpt-5.4 · dial.

Step 6: Run Codex​

codex exec "Reply with exactly: PONG"

You see PONG printed back, routed Codex → proxy → DIAL. Run plain codex for the interactive session you will use day to day.

Additional Information​

DIAL does not store responses or support previous_response_id, which is why disable_response_storage = true is required on both paths — without it, Codex breaks after the first turn.

  • Which path you need depends on the deployment. Re-run the Step 1 check for any new deployment — the flag can differ even between versions of the same model family.
  • Codex sends the key as an Authorization: Bearer token; the env_http_headers line is what makes it also send the Api-Key header that DIAL authenticates on. On the proxy path, the proxy forwards it.
  • On the proxy path, mind the proxy's own limits: only function tools are forwarded, file inputs must be inlined, and no session state is kept.