# Maniac: Continually optimizing models from your LLM telemetry and evals.

Maniac is an enterprise AI platform that makes it easy to replace existing LLM API calls with fine-tuned, task-specific models. Drop in Maniac with one line of code to:

* Capture and structure production **LLM traffic**
* Automatically **fine-tune** and **evaluate** Small Language Models (SLMs) on your tasks
* **Replace over-generalized LLM calls** with higher performance, lower latency models built for just what you need
* **Focus engineering time where it matters most:** building and refining high-quality model evaluations—not managing infrastructure, hyperparameters, or bespoke fine-tuning pipelines

All with virtually no changes to your existing codebase.

## Getting started

{% stepper %}
{% step %}
**Sign up for Maniac**

Head over to <https://app.maniac.ai/auth/register>
{% endstep %}

{% step %}
**Grab your API key**

From Project Settings or on the Containers page.
{% endstep %}

{% step %}
**Set the environment variable**

```bash
export MANIAC_API_KEY=your-api-key
```

{% endstep %}
{% endstepper %}

{% embed url="<https://youtu.be/gTcfZSXy4hM>" %}

***

## Dropping Maniac into your Codebase

For an agentic setup, copy this prompt and give it to your preferred coding agent:

```
Fetch and follow https://raw.githubusercontent.com/ManiacIncorporated/sb-monorepo/main/docs/gitbook/agent-setup.md to instrument this repo for Maniac telemetry.
```

#### Install the library

{% tabs %}
{% tab title="Python" %}

```bash
pip install maniac
```

{% endtab %}

{% tab title="TypeScript" %}

```bash
npm install maniac-js
```

{% endtab %}
{% endtabs %}

#### Initialize client

{% tabs %}
{% tab title="Python" %}

```python
from maniac import Maniac

maniac = Maniac(api_key="your-maniac-api-key")
```

{% endtab %}

{% tab title="TypeScript" %}

```typescript
import Maniac from "maniac-js";

const maniac = new Maniac();
```

{% endtab %}
{% endtabs %}

#### Create a container

Containers log inference and automatically build datasets for fine-tuning and evaluation. <mark style="color:red;">`initial_model`</mark> sets the model used in that container until a Maniac model is deployed in its place.

{% tabs %}
{% tab title="Python" %}

```python
container = maniac.containers.create(
  label = "my-container",
  default_system_prompt = "You are a helpful math tutor.",
  initial_model = "openai/gpt-5.2"
)
```

{% endtab %}

{% tab title="TypeScript" %}

```typescript
const container = await maniac.containers.create({
  label: "my-container",
  initial_model: "openai/gpt-5.2",
});
```

{% endtab %}
{% endtabs %}

#### Log Completions

Now that you've made a container, let's add some data to it. Note that the inputs and outputs of items are both required to be in openai chat completions format. The output field is not required.

{% tabs %}
{% tab title="Python" %}

```python
response = maniac.chat.completions.register(
    container = "my-container",
    items = [{
        "input": {
            "messages": [{"role": "user", "content": "Hello!"}]
        },
        "output": {
            "choices": [{
                "message": [{"role": "assistant", "content": "Hello! How can I help?"}]
            }]
        }
    }]
)
```

{% endtab %}

{% tab title="TypeScript" %}

```typescript
await maniac.chat.completions.register({
  container: "my-container",
  items: [{
    input: {
      messages: [{ role: "user", content: "Hello!" }],
    },
    output: {
      choices: [{
        message: { role: "assistant", content: "Hello! How can I help?" },
      }],
    },
  }],
});
```

{% endtab %}
{% endtabs %}

{% embed url="<https://youtu.be/5ygzMi4okJ8>" %}

***

## Optimizing your model

The inference logs in your container now serve as training data for a new SLM—fully yours, lower latency, most cost effective, and optimized specifically for your task.

{% stepper %}
{% step %}
**Create an&#x20;**<mark style="color:green;">**Eval**</mark>

Evaluations define the optimization target. They can be implemented as arbitrary code or defined using judge prompts.

From the **Evals** tab inside a container, **Add Eval**.

{% tabs %}
{% tab title="Judge Eval" %}

<div data-with-frame="true"><figure><img src="https://2748330800-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FIDQ6PndRGkOnW2GcVyH5%2Fuploads%2Fgit-blob-9d8892c08aa7b8387e9005ce222f05a5c6b3a9ad%2FScreenshot%202026-01-06%20at%2012.43.14%E2%80%AFPM.png?alt=media" alt=""><figcaption></figcaption></figure></div>
{% endtab %}

{% tab title="Code Eval" %}

<div data-with-frame="true"><figure><img src="https://2748330800-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FIDQ6PndRGkOnW2GcVyH5%2Fuploads%2Fgit-blob-54351ad3efb2650c92c168dfc453732b93acdcbe%2FScreenshot%202026-01-06%20at%201.15.10%E2%80%AFPM.png?alt=media" alt=""><figcaption></figcaption></figure></div>
{% endtab %}
{% endtabs %}
{% endstep %}

{% step %} <mark style="color:green;">**Optimization**</mark>**&#x20;happens automatically**

Once your telemetry hooks and evals are in place, Maniac automatically optimizes a model for your task — no manual configuration required.
{% endstep %}
{% endstepper %}

***

## Deploy.

Optimized models can be be deployed into a container from the **Models** tab. Once deployed, you can chat with your generated models, and inference requests are now routed through the Maniac model instead of the <mark style="color:$info;">`initial_model`</mark>.

<div data-with-frame="true"><figure><img src="https://2748330800-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FIDQ6PndRGkOnW2GcVyH5%2Fuploads%2Fgit-blob-35bfc5634ad448e3e3370137e699d493cd83c583%2FScreenshot%202026-01-06%20at%202.05.29%E2%80%AFPM.png?alt=media" alt=""><figcaption></figcaption></figure></div>

### Need help?

:e-mail: Email us at <helpme@maniac.ai>

We'll get back to you within a day.