Version prompts, route LLM calls, trace and evaluate — one platform.

AcruxCore is an open-source, Apache-2.0, self-hostable LLMOps platform for engineering teams. It sits between your app and the model providers you connect — so you ship, change, and measure LLM features without redeploying to move a prompt.

import acruxcore from '@acruxcoreai/sdk';

const hub = new acruxcore({ apiKey });

// move 'production' between versions

// — no redeploy of your app

const { messages, tools } = await hub.prompts.render(
  'support-agent',
  'production',
  { ticket },
);

Watch

The whole platform in two and a half minutes.

One thread, end to end: a versioned prompt, the gateway call that renders it, the trace that call produced, an agent traced over OpenTelemetry without the gateway, and a thumbs-down that becomes a scored fix — shipped by moving an alias. Recorded against a live instance, not a mockup.

The round-trip

From a bad answer to a fixed prompt — without leaving the app.

Follow one thread across the whole platform. Open the trace behind a bad answer, widen to the session it belonged to, rate the span that missed, and jump straight to the exact prompt version that produced it. Fix it — or let the optimizer draft the fix — score it against the cases it got wrong, then move the production alias. The promotion is recorded, so the change has a name against it. No redeploy, no context-switch, no leaving AcruxCore.

The platform

Six parts of the LLM stack, one control plane.

Each piece works on its own and composes with the rest. Adopt what you need, no rip-and-replace — prompts, tracing and tools do not need your traffic routed through us at all.

How it works

Wired in three steps.

Why AcruxCore

Why teams switch.

• Change prompts without a deploy. Move the production alias: the next render returns it, and SDK callers pick it up within their cache window.
• Drop-in gateway. OpenAI-compatible, so your current client and code paths stay put.
• TypeScript and Python SDKs. First-class async clients for both — same prompts, gateway, tool loops, and tracing — plus a plain REST API.
• Every call is a trace. Spans for model, tokens, latency, and cost — automatic through the gateway, or over OpenTelemetry.
• Prove a change before you ship it. Score prompt versions and models against a dataset, or let the optimizer draft the rewrites and rank them.
• Guardrails on spend and failure. Budgets and per-key rate limits, an ordered fallback chain per model, and caching you switch on per key.
• Every change has a name against it. A team-wide audit trail of keys, members, gateway, secrets, prompts and tools — filter it by area, event or person.
• Own your keys and data. Bring your own provider keys, or self-host the whole platform.
• Open source. Apache License 2.0 — read the code, self-host it, or send a pull request.

import acruxcore from '@acruxcoreai/sdk';

const hub = new acruxcore({ apiKey });

// rendered prompt + its attached tools

const { messages, tools } =
  await hub.prompts.render('support-agent', 'production', { ticket });

// gateway routes, prices & traces the loop

const result = await hub.gateway.runToolLoop({
  model: 'gpt-4o', messages, toolDefs: tools, dispatch,
});

console.log(result.content);   // final answer
console.log(result.traceId);   // spans + cost