> ## Documentation Index
> Fetch the complete documentation index at: https://docs.ai-stats.phaseo.app/llms.txt
> Use this file to discover all available pages before exploring further.

# Set up coding agents with SKILL.md

> Pair the SDK with a SKILL.md file so local coding agents can use your AI Stats integration consistently.

This recipe is for teams building internal tools, CLIs, or coding-agent workflows on top of the gateway SDKs.

<Note>
  Reusable AI Stats skills live in [`AI-Stats/skills`](https://github.com/AI-Stats/skills). Install them first with:

  ```bash theme={null}
  npx skills add AI-Stats/skills --agent codex --skill '*'
  ```

  Then layer project-local instructions into the repository `SKILL.md`.
</Note>

The main idea is simple:

* keep your SDK setup inside the repository
* add a `SKILL.md` file that teaches the agent how the project expects requests to be made
* point the agent at that file during setup so request patterns stay consistent

## 1. Install the SDK

If you're using TypeScript or JavaScript:

<CodeGroup>
  ```bash npm theme={null}
  npm install @ai-stats/sdk
  ```

  ```bash pnpm theme={null}
  pnpm add @ai-stats/sdk
  ```

  ```bash yarn theme={null}
  yarn add @ai-stats/sdk
  ```

  ```bash bun theme={null}
  bun add @ai-stats/sdk
  ```
</CodeGroup>

If you're using Python:

```bash theme={null}
pip install ai-stats
```

## 2. Add a project-local skill file

Create `SKILL.md` in the repository root or the agent skill location your team uses.

```md theme={null}
# AI Stats Gateway

- Use `https://api.phaseo.app/v1` as the gateway base URL.
- Read the API key from `AI_STATS_API_KEY`.
- Prefer the Responses API for new text integrations.
- Use preset slugs when the repository defines shared request defaults.
- Log request ids and model ids for debugging.
- For async video jobs, store the returned id and poll status until terminal.
```

## 3. Keep the first integration tiny

For coding agents, smaller request helpers are easier to reuse than large wrappers.

```ts theme={null}
import AIStats from "@ai-stats/sdk";

export const gateway = new AIStats({
  apiKey: process.env.AI_STATS_API_KEY!,
});
```

Then keep one obvious call path per surface:

* text generation
* async video
* embeddings
* image generation

## 4. What the skill should encode

Your `SKILL.md` should answer the questions that waste the most agent time:

* which endpoint family to prefer
* which environment variable holds credentials
* which model ids or preset slugs are approved
* how request ids should be logged
* how async jobs should be tracked

## 5. Recommended repository pattern

* SDK helper in one shared module
* `SKILL.md` next to the integration code or at repository root
* one short example per major endpoint
* one note covering request logging and error handling

This gives both humans and coding agents the same operational entrypoint.

## Related guides

* [Codex](../guides/integrations/codex.mdx)
* [AI SDK integration](../guides/integrations/ai-sdk.mdx)
* [Examples](../guides/examples.mdx)