# Deep Code

> Run DeepSeek's Deep Code agent on Tchavi by setting BASE_URL, and pay in FCFA by mobile money.

- Canonical: https://tchavi.com/en/docs/coding-agents/deep-code

---


Part of the [coding agents](/docs/coding-agents) guides.

Deep Code is DeepSeek's terminal agent. It reads an OpenAI-compatible `BASE_URL`, so it points at Tchavi with three lines — and the same file configures the Deep Code VS Code extension.

## Configure it

Create `~/.deepcode/settings.json`:

```json
{
  "env": {
    "MODEL": "deepseek-v4-pro",
    "BASE_URL": "https://tchavi.com/api/v1",
    "API_KEY": "sk-tch-your-key"
  },
  "thinkingEnabled": true,
  "reasoningEffort": "max"
}
```

That is the whole setup. Start `deepcode` in a project and it runs on Tchavi credits.

## Choosing the model

`MODEL` takes any Tchavi chat model, not only DeepSeek's:

| Model               | Credits / 1K in | Why                                           |
| ------------------- | --------------- | --------------------------------------------- |
| `deepseek-v4-pro`   | 3               | Closest to Deep Code's native behaviour       |
| `deepseek-v4-flash` | 0.72            | The budget option for routine edits           |
| `claude-sonnet-4-6` | 7               | Stronger on large refactors                   |
| `gpt-5.4`           | 6               | A different model family for a second opinion |

`reasoningEffort` and `thinkingEnabled` are Deep Code's own settings. They are passed to the model as sent; a model that does not support extended thinking simply ignores them.

<Callout type="warning" title="Where the key lives">
  A project-level settings file overrides the user-level one. Keep the key in the user-level file at
  `~/.deepcode/settings.json` and leave credentials out of any per-project copy you might commit.
</Callout>

---

Next: [Codex](/docs/coding-agents/codex) · Back to [coding agents](/docs/coding-agents)