Codex CLI

Checked against Codex CLI's docs on 2026-10-11

Codex CLI connects directly. Creekside also serves OpenAI's Responses API, so one provider entry in ~/.codex/config.toml is all it takes.

Base URL: https://api.creeksidecompute.ai/v1
API: OpenAI Responses, wire_api = "responses"
Model id: qwen/qwen3.8-27b, or qwen/qwen3.6-35b-a3b for speed, or meta/muse-glimmer-30b, which always reasons first
Context window: 131072 tokens
Key: In CREEKSIDE_API_KEY. Make one at creeksidecompute.ai/account

Set it up

Make a key. Sign in at creeksidecompute.ai/account with your email and make a key. Each request is paid from prepaid credit: top up by card from $10, or redeem a promotional code.
Put the key in an environment variable. Add this line to your shell profile (for example ~/.zshrc), then open a new terminal:
shell
```
export CREEKSIDE_API_KEY=<your key>
```
Add Creekside to Codex. Put this in ~/.codex/config.toml. The provider must be in this file in your home folder: a project's .codex/config.toml cannot set it.
~/.codex/config.toml
```
model = "qwen/qwen3.8-27b"
model_provider = "creekside"
model_context_window = 131072
web_search = "disabled"

[model_providers.creekside]
name = "Creekside Compute"
base_url = "https://api.creeksidecompute.ai/v1"
env_key = "CREEKSIDE_API_KEY"
wire_api = "responses"
```
model_context_window gives Codex the real window. Without it, Codex assumes 272,000 tokens and compacts too late. web_search = "disabled" says what is true anyway: no web search runs here.
Run Codex in your project: codex. To run one task and exit: codex exec "Summarise this repository".

Thinking

Thinking follows model_reasoning_effort in config.toml. Unset, it is off. low, medium and high turn it on; none and minimal keep it off. For one run: codex -c model_reasoning_effort=high.

Add show_raw_agent_reasoning = true to see the thinking. It comes back as the model's own reasoning text, not as a summary. The model id qwen/qwen3.8-27b:thinking is the same model with thinking on for every request. Thinking tokens are billed as output.

Muse Glimmer 30B always reasons before it answers, so model_reasoning_effort sets how long, not whether: unset, or at none, minimal or low, it reasons briefly, and medium and high make it reason longer. The model id meta/muse-glimmer-30b:thinking gives high.

What does not work

Chat Completions. Codex speaks only the Responses API, and wire_api = "chat" no longer loads.
Stored state. previous_response_id, conversations and background responses are refused. Codex sends the whole conversation on every turn, so it does not need them.
Hosted tools. Web search, file search, code interpreter, image generation and remote MCP run at OpenAI. MCP servers that you configure in Codex work.
Reasoning summaries and encrypted reasoning carried between turns.
The apply_patch tool. Codex offers it only to models in its own catalogue, so this model edits files with shell commands.
The model picker. /model shows the models of your OpenAI account. Choose ours in config.toml or with -m qwen/qwen3.8-27b.

Troubleshooting

Missing environment variable: CREEKSIDE_API_KEY: Codex did not find the key. Export CREEKSIDE_API_KEY in the shell that starts Codex.
401 invalid API key: The key is wrong, or Codex did not get it. Set CREEKSIDE_API_KEY in the shell that starts Codex, and restart Codex after you change it. To test the key alone, curl -s https://api.creeksidecompute.ai/v1/models -H "Authorization: Bearer $CREEKSIDE_API_KEY" returns the model list when the key is good.
404 model … is not served: The model id is wrong. Use qwen/qwen3.8-27b exactly, in lower case, with the qwen/ prefix. GET /v1/models lists every id we serve.
context_length_exceeded: The conversation is longer than the 131072-token window. Check that model_context_window = 131072 is set, so Codex compacts in time. Codex treats this error as a full context and compacts.
402 or 429: A 402 means your prepaid credit does not cover the request: add credit at creeksidecompute.ai/account. A 429 means the key's rate, or the number of requests your account may run at once, is used up; agents that work in parallel reach that sooner. Wait and retry. The API reference lists the limits.

Other tools

Sources, read on 2026-10-11: configuration reference, advanced configuration, Codex 0.162.1 source.