Skip to content

Codex CLI

Use vllm-responses as the model provider for Codex CLI.

Prerequisites

  • A running vllm-responses gateway. See Quickstart.
  • Codex CLI installed (codex on $PATH).

1. Gateway

Start the gateway with the --codex-approval-model flag so Codex's guardian auto-review feature works:

vllm-responses serve \
  --upstream http://<vllm-host>:<port>/v1 \
  --codex-approval-model Qwen/Qwen3.6-35B-A3B

vllm serve Qwen/Qwen3.6-35B-A3B \
  --responses \
  --responses-codex-approval-model Qwen/Qwen3.6-35B-A3B

Replace the model name with whatever your backend exposes.

2. Codex Config

Add the provider to ~/.codex/config.toml (or $CODEX_HOME/config.toml):

[model_providers.vllm-responses]
name = "vllm-responses"
base_url = "http://127.0.0.1:8457/v1"
wire_api = "responses"

3. Run Codex

# -c model_context_window is needed to ensure the codex auto compaction threshold follows the model context length
codex --disable image_generation \
  -c model_provider=vllm-responses \
  -m Qwen/Qwen3.6-35B-A3B \
  -c model_context_window=262144

Compaction

The gateway's implementation of POST /v1/responses/compact is not ready. However, Codex falls back to its own compaction method for custom providers: it sends a normal POST /v1/responses with "tools": [] asking the model to summarize prior turns. The gateway accepts this, so long interactive sessions still work.

References