# Use Codex CLI with 7API

> Configure OpenAI Codex CLI to use 7API through the Responses API. A config.toml for GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna with 1.05M-token context and auto-compaction.


[Codex CLI](https://developers.openai.com/codex/) uses the OpenAI Responses API, which 7API supports natively. You don't need an OpenAI account.

## Configure

Save this as:

- **macOS / Linux:** `~/.codex/config.toml`
- **Windows:** `%USERPROFILE%\.codex\config.toml`

Replace `YOUR_7API_KEY` with your key, then run `codex`.

```toml title="config.toml"
# GPT-6 Astra: 1.05M context, auto-compacts at 900K (leaves room for 128K output).
# gpt-6.1-sol and gpt-6-luna use the same limits.
model = "gpt-6-astra"
model_provider = "7api"
model_reasoning_effort = "high"
approval_policy = "never"
model_context_window = 1050000
model_auto_compact_token_limit = 900000

[model_providers.7api]
name = "7api"
base_url = "https://7api.st"
wire_api = "responses"
experimental_bearer_token = "YOUR_7API_KEY"

[windows]
sandbox = "elevated"
```

## Choosing a model

The default is `gpt-6-astra`. To use another model, change the `model` value:

| Model | Rate per 1M tokens (standard / cached) |
|---|---|
| `gpt-6-astra` | $0.65 / $0.06 |
| `gpt-6.1-sol` | $0.15 / $0.04 |
| `gpt-6-luna` | $0.05 / $0.01 |

All four have a 1,050,000-token context window. The config compacts the conversation at 900,000 tokens so there's room left for output.

## Notes

- `model_reasoning_effort` accepts `low`, `medium`, `high`, `xhigh` and `max`.
- `approval_policy = "never"` lets Codex run commands without asking. Change it to `"on-request"` if you want to approve each one.
- See OpenAI's [configuration reference](https://developers.openai.com/codex/config-reference/) for the other options.
