AITOLL API Integration Skill

AITOLL is a unified OpenAI-compatible gateway that routes requests to multiple LLMs

(DeepSeek, GPT-5.2, Claude, Gemini) and image models through a single endpoint.

---

Workflow

When this skill is invoked, follow these steps:

1. Identify the task type:

- Text chat / Q&A → use a text model

- Image generation (text → image) → use gemini-3-pro-image-preview

- Image editing (image → modified image) → use gemini-3-pro-image-preview

- Image understanding (image → text description) → use a multimodal model

2. Identify the programming language the user wants (Python, JS, cURL, etc.).

Default to Python with requests if unspecified.

3. Check for API key: Remind the user to set AITOLL_API_KEY as an environment

variable if they haven't mentioned it.

4. Select a model using the decision tree below, then generate complete runnable code.

5. Explain the model choice and any important caveats (content format, streaming, etc.).

---

Model Selection

Use this decision tree to pick the right model:

| Use case | Preferred model | Alternative |

|---|---|---|

| Text chat, cost-sensitive | deepseek-chat | claude-haiku-4.5 |

| Text chat, balanced | glm-4.7 | glm-5 |

| Text chat, quality-first | gpt-5.2 | claude-sonnet-4.5 |

| Code generation | claude-haiku-4.5 | claude-sonnet-4.5 |

| Multimodal understanding | gemini-3-flash-preview | gemini-3-pro-preview |

| Image generation / editing | gemini-3-pro-image-preview | — |

See [references/models.md](https://aitoll.net/skills/aitoll-chat-completions/references/models.md) for the full model table with

streaming and multimodal capability notes.

---

Code Generation Guidelines

Always follow these rules when writing AITOLL integration code:

• Never hardcode API keys. Always read from os.environ.get("AITOLL_API_KEY").
• Always include error handling. Call response.raise_for_status() (Python) or

check HTTP status (other languages).

• Handle `content` as string OR array. Text models return content as a string.

Image models (gemini-3-pro-image-preview) return content as an array of objects

with type: "image_url". Write code that handles both.

• Use multimodal message format (content as array with type: "text" / type: "image_url")

for image generation, editing, or understanding requests.

• For streaming, set "stream": true in the request body and parse SSE chunks

(each line starts with data: , ends with data: [DONE]).

---

Output Format

When responding to the user:

1. Show complete, runnable code — no placeholders except AITOLL_API_KEY.

2. Include a brief explanation of which model was chosen and why.

3. Add a reminder: set export AITOLL_API_KEY="your-key" before running.

4. If the response may contain an image (base64), show how to save or display it.

---

References

Consult these files for detailed specs:

• [references/models.md](https://aitoll.net/skills/aitoll-chat-completions/references/models.md) — Full model table, streaming support,

multimodal capabilities, and selection guidance.

• [references/api-reference.md](https://aitoll.net/skills/aitoll-chat-completions/references/api-reference.md) — Base URL, auth headers,

request/response schema, content formats, error codes, and streaming details.

• [references/code-examples.md](https://aitoll.net/skills/aitoll-chat-completions/references/code-examples.md) — Complete working examples

in Python and cURL for all task types.