What is AcruxCore?

AcruxCore is an LLM-ops platform. You keep the prompts your app sends to language models in AcruxCore, and your app calls the models through the AcruxCore gateway. Every request is recorded, so you can see what happened. You can also measure whether a change made the answers better.

Everything in AcruxCore belongs to a team. When you sign up you get a team workspace, and you can invite teammates to it. You write prompts and inspect results in the web app. Your running application talks to AcruxCore through the REST API or one of two SDKs: @acruxcoreai/sdk for Node and acruxcore for Python. The two SDKs have the same feature set, so every feature in these docs works in Node and in Python.

The six building blocks

AcruxCore has six pieces. You can use each piece on its own or together with the others.

Block	What it does
Prompt management	A prompt is a set of message templates, with a new version for each change. A production alias points at one version. Move the alias to another version without redeploying your app.
LLM gateway	One OpenAI-compatible endpoint for every provider you connect: OpenAI, Anthropic, Gemini, or any other OpenAI-compatible endpoint you register. You bring your own provider keys. The gateway adds routing, fallbacks, cost tracking, and caching.
LLM observability	Every gateway call is recorded as a trace. The trace's spans hold the model, tokens, latency, and cost. If you call providers from your own code, you can send the same spans over OpenTelemetry instead.
LLM tool calling	Functions the model can call, versioned the same way as prompts. Declare a tool in your own code, and AcruxCore adds the tool to the tool catalog. Or define an HTTP tool in the catalog, and AcruxCore calls the tool's URL for you.
LLM evaluation	Build datasets from real feedback, by hand, or from a CSV or JSON file. Run experiments that compare prompt and model versions. Score live traffic with rules that keep running on new traces. The optimizer can draft the next prompt version for you.
Audit log	A list of every recorded change in the team, newest first. The list covers keys, members, gateway, secrets, prompts, tools, and evaluations. You can filter it by area, event, or the person who made the change.

How the pieces connect

Author a prompt  →  call it through the gateway  →  the call is traced
      (Prompts)              (Gateway)                   (Tracing)
                                  │
                          attach tools                collect feedback
                             (Tools)                  → build a dataset
                                                       → run an experiment
                                                          (Evaluation)

A concrete run looks like this:

1. You author support-reply in the Prompts UI and promote v2 to production.
2. Your app calls the Gateway with a reference to that prompt. The gateway renders the template, picks the model, and calls the provider.
3. The call shows up in Tracing with its model, token counts, and latency.
4. You declare a get_weather Tool in code. The SDK's tool-calling loop registers get_weather and gives the model the tool's schema. When the model asks for the tool, the loop runs your function and adds a span to the same trace.
5. Users rate the answers with a thumbs up or a thumbs down. You turn those ratings into a dataset and evaluate a new prompt version against the dataset.
6. The team's audit trail records the promotion in step 1, the tool you declared, and the creation of the API key your app uses. Each entry names the person who made the change.

Who it's for

• App developers who call LLMs from Node or Python and want prompts they can change without shipping code.
• Teams who want one path to every model provider, with the cost of each call shown and every change recorded.
• Anyone who keeps prompts as strings in code and debugs LLM features with print statements.

In the Quickstart you make your first call in a few minutes. Core concepts explains how the pieces fit together, if you want that first.