# Codex CLI (Mac/Linux)
> Install and configure the Codex CLI on Mac / Linux to connect to the LMU AI API, including how to enable the 1M-token long context in config.toml.
URL: https://docs.lmuai.ai/docs/tools/codex-cli-mac
After installing the Codex CLI, create two config files by hand (`config.toml` + `auth.json`) to connect to the LMU AI API — no third-party config tool needed.
***
## Step 1 — Install Node.js [#step-1--install-nodejs]
You need **Node.js 20+** — download: [https://nodejs.org/en/download](https://nodejs.org/en/download)
Click through the installer, then verify in the terminal:
```bash
node --version
# A version number (e.g. v23.7.0) means it installed correctly
```
## Step 2 — Install the Codex CLI [#step-2--install-the-codex-cli]
```bash
npm i -g @openai/codex
```
If you hit network issues, switch to a China mirror first:
```bash
npm config set registry https://registry.npmmirror.com
npm i -g @openai/codex
```
Verify the install:
```bash
codex --version
# A version number means it installed correctly
```
## Step 3 — Create the config files [#step-3--create-the-config-files]
The config directory is **`~/.codex/`** (create it if missing; overwrite any old files). Create/overwrite these two files in it:
### config.toml [#configtoml]
```toml
model = "gpt-5.4"
model_reasoning_effort = "xhigh"
disable_response_storage = true
sandbox_mode = "danger-full-access"
windows_wsl_setup_acknowledged = true
approval_policy = "never"
file_opener = "vscode"
model_provider = "codex"
web_search = "cached"
suppress_unstable_features_warning = true
[history]
persistence = "save-all"
[tui]
notifications = true
[shell_environment_policy]
inherit = "all"
ignore_default_excludes = false
[sandbox_workspace_write]
network_access = true
[features]
plan_tool = true
apply_patch_freeform = true
view_image_tool = true
unified_exec = false
streamable_shell = false
rmcp_client = true
elevated_windows_sandbox = true
[profiles.auto-max]
approval_policy = "never"
sandbox_mode = "workspace-write"
[profiles.review]
approval_policy = "on-request"
sandbox_mode = "workspace-write"
[notice]
hide_gpt5_1_migration_prompt = true
[model_providers.codex]
name = "codex"
base_url = "https://api.lmuai.ai"
wire_api = "responses"
requires_openai_auth = true
```
### auth.json [#authjson]
```json
{
"OPENAI_API_KEY": "your sk- key generated in the console"
}
```
## Step 4 — Launch the Codex CLI [#step-4--launch-the-codex-cli]
After saving the config files, reopen the terminal and run the following to launch and sign in automatically:
```bash
codex --dangerously-bypass-approvals-and-sandbox
```
Or the short form:
```bash
codex --yolo
```
> This launch command skips security approvals, so you do not approve each auto-executed step.
On a successful launch, the interface shows the current model as `gpt-5.4 xhigh` (the model set by the `model` field in `config.toml`).
If you accidentally switch to a different model in a session, change the `model` field in `~/.codex/config.toml` back to `gpt-5.4` and restart the terminal to restore it.
***
## Enabling the 1M context in Codex (optional) [#enabling-the-1m-context-in-codex-optional]
`gpt-5.4` supports a **1M-token long context window** (272K by default), useful for large-repo review, long-log analysis, and cross-file refactors. Codex does **not** enable the 1M window by default; you must declare it explicitly in `config.toml`.
### How to configure: add two lines to config.toml [#how-to-configure-add-two-lines-to-configtoml]
Edit `~/.codex/config.toml` and add the following at the **top level** (before any `[section]`, e.g. right after `model = "gpt-5.4"`):
```toml
model_context_window = 1000000
model_auto_compact_token_limit = 900000
```
Save, reopen the terminal, and re-enter Codex to apply it.
**Field reference**
| Field | Purpose |
| -------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `model_context_window` | Declares the model's usable context window in tokens; `1000000` is 1M. Unset, it uses the model default (272K) |
| `model_auto_compact_token_limit` | The token threshold that triggers auto-compaction of history; set it to about 90% of the window (`900000` for 1M). Values above 90% of the window are clamped by Codex, so a larger number is pointless |
### Verify it took effect [#verify-it-took-effect]
After re-entering Codex, type `/status` and check whether `context window` shows around 1M (some versions show about 950K — a reserved 5% safety margin from Codex). You can also use `/statusline` to add context usage to the status bar for live monitoring.
### If you use CC Switch [#if-you-use-cc-switch]
If you manage the Codex config with [CC Switch](/docs/tools/cc-switch), add the same two lines in its Codex config editor and apply — the effect is identical:
* **Must be at the top level**: these two fields do not take effect inside a `[profiles.xxx]` section (a known Codex limitation); also put them in the user-level `~/.codex/config.toml`, since a project-level `.codex/config.toml` may be ignored
* **Model support**: `gpt-5.4` and `gpt-5.5` support 1M over the API channel; models like `gpt-5.3-codex` are capped at 272K, and no config value can exceed the model's own limit
* **Higher cost**: a longer context uses more tokens per request; the portion above 272K is billed at the higher long-context tier, so do not leave it on for anything but very long tasks
* **Version differences**: some newer Codex versions read the window size from the model catalog first; if `/status` still shows the old value, upgrade to the latest (`npm i -g @openai/codex@latest`) and retry