How to get started

Check if this is the right service for you

The universe of APIs connected to AI is vast.

The ultimate aim of this project is to allow you to monitor costs anywhere in that universe.

We are not there yet. We may, in fact, never be there. The universe is expanding rapidly.

However, we've built something useful, so we've decide to release it while we continue development.

What we are good at

We are good at handling direct API responses. We don't expect you to do any normalisation, or add any fields of your own. Just forward what you received, and we'll do the calculations. The only important thing we ask is that you redact any large assets, to save our bandwidth. Please also redact anything you want to keep private. This can be achived with a few simple statements like response['output']="REDACTED" before sending it to us. See our redaction-guide for a list of recommended redactions.

If you are using AI to generate assets (eg images, video, sound files, blocks of text), and your workflow is send prompt -> receive asset, we have you covered. We support all the major models and endpoints, including batch modes.

Similarly, if you are using AI to process data (eg image analysis, voice transcription, text analysis, image editing), and your workflow is send data -> receive result, you're going to have a good time here.

What we are NOT good at

Working with some SDKs. If you use a SDK between your workflow and the AI API, it must have a faithful "show raw response" feature, so you can get the raw response to send us. Most of them do this, but see our sdk-guide for more details if you are using a SDK.

Streaming APIs, where you send a prompt and receive a stream of data back, are not supported yet. They will be soon though. There are no technical barriers, just coding effort required.

Steerable APIs, like GPT-6 Astra, are not supported yet.

How your workflow design can help or hinder

Models are generally designed around a single native modality (eg text, image, audio). If you send a prompt like "Give me an image of a cat" to a model which is primarily designed for text, that model will have to use an image generation model in the background. The usage report you get back with your cat image will then contain a "tool cost" related to that. Sometimes this is problematic. For example, when OpenAI models do this, they don't tell you which image generation model they used. They'll tell you how many tokens were used by the image generation model, but not which model it was. All the image generation models have different token costs, so we can't tell you what the actual cost was. So if you place your image generation request directly to an image generation model of your choosing, you will get a more accurate cost report.

Playground

There's a playground where you can see what the API does. You can paste in your own example, or try a pre-loaded one.

Get API Key

Sign up for an account here.

No credit card is required, just a valid email address where you can receive a confirmation email.

Pricing

Pricing and conditions will likely change in future.

Pricing is based on monthly subscription, with rate limits.

Start on the free tier. This gives you 10 calls per minute, 100 calls per 24 hours.

Paid tiers unlock higher rate limits. See pricing for details.

Pricing, and conditions will likely change in future.

Enterprise agreements are available by negotiation.