Acrux Core

AcruxCore is an LLM-ops platform for engineering teams that combines five jobs that are often handled by separate products: prompt versioning, an OpenAI-compatible AI gateway, request tracing and observability, a versioned tool catalog, and evaluation. Its central idea is operational rather than model-centric: a team should be able to trace a bad answer back to the exact prompt and tool versions that produced it, change the prompt or tool, move a production alias and observe the next run without redeploying application code.¹²

AcruxCore’s strongest feature is that its components are designed around one operational loop rather than simply collected in one dashboard. Prompts are immutable versioned objects with aliases such as production and staging. A prompt can be rendered by alias at runtime, and moving the alias changes the version consumed by clients without changing application code.²

The gateway is OpenAI-compatible, which lowers migration friction for teams already using OpenAI-style clients. AcruxCore’s SDK documentation describes a single client surface for prompt rendering, gateway calls, tool loops, tracing and feedback. The same feature families are available through TypeScript, async Python and plain REST, so adoption is not tied to one application language.³⁴

The tracing model is particularly relevant for agentic applications. Gateway calls can record model, token, latency and cost metadata, while tool calls can be represented as spans in the same trace. Feedback can be attached to traces or individual spans and then used to build evaluation datasets. In principle, that creates a useful path from production failure to reproducible evaluation.³

None of this proves that AcruxCore executes these workflows better than established competitors such as Langfuse, Helicone, Phoenix or Opik. It does show that the product has a real systems thesis rather than being a thin UI around a single provider API.

AcruxCore treats prompts as versioned data. The current prompt-management page describes immutable versions, named aliases, server-side templating, diffs and an audit trail. The dashboard can edit and commit a prompt, attach tools, set a default model and promote or roll back an alias.²

For engineering teams, the important benefit is separation of prompt iteration from application deployment. A prompt change can be evaluated and promoted without rebuilding the calling service. That can shorten iteration cycles and improve reproducibility because the trace can point back to the precise version that was active.

AcruxCore does not resell model tokens. The hosted service requires users to bring their own provider credentials, and current pricing documentation says the provider bills the user directly at its own rates while AcruxCore adds no token markup.⁵

That design has two advantages. First, it avoids forcing a team onto a proprietary model marketplace. Second, it makes provider cost visible without embedding an additional per-token margin in the gateway.

The platform also supports a self-hosted route and OpenAI-compatible providers, which reduces lock-in compared with a closed hosted-only gateway. Prompt versions can be exported as JSON and traces are exposed through the API.⁵

The current SDK surface supports versioned tools and tool loops. In the TypeScript package, a tool can be declared with a schema and handler, synchronized to the catalog and then used in a model loop. The package documentation says multiple tools requested in a turn can execute concurrently and that traces can preserve the tool version that actually ran.⁶

This is useful because tool definitions are often a hidden source of agent regressions. Versioning the tool contract and connecting it to the trace gives a team more information than an opaque “agent failed” event.

The SDK is also evolving quickly. PyPI release history shows multiple releases from late July through early August, including a breaking API change around tool handling in version 0.5.0 before later 0.7.x releases.⁴ npm currently exposes a newer 0.8.0 TypeScript package.⁶ Rapid iteration is expected in beta, but it increases integration churn and makes version pinning important.

AcruxCore positions evaluation as the final part of the loop: collect feedback, build datasets, run prompt or session experiments and compare results. That is a sensible architecture because tracing without evaluation can become passive telemetry rather than an improvement system.¹

For a real pilot, teams should test at least dataset versioning, reproducibility, evaluator configuration, model/prompt comparison, failure attribution, exportability and the behavior of evaluation jobs under rate limits or provider errors.

The current official pricing and About pages explicitly say the full platform is open source under Apache 2.0 and can be self-hosted.⁵¹ AlternativeTo independently identifies the repository as Apache-2.0 and reports that it was updated in August, although the listing itself was initially added by AcruxCore and therefore is not wholly independent product validation.⁷

The SDK packages have their own licensing metadata. npm lists the TypeScript SDK as MIT, and PyPI lists the Python client as MIT.⁶⁴ These package licenses do not contradict the platform’s Apache-2.0 status; they apply to different artifacts.

Public repository activity is still tiny. AlternativeTo showed only a handful of GitHub stars in the current capture.⁷ Star count is not a quality metric, but at this stage it is a useful maturity signal: AcruxCore is open source in license and availability, not yet a broadly adopted open-source standard.

The current hosted beta is free. AcruxCore says every part of the platform is unlocked, there is no credit card requirement, prompt and tool versions are unlimited, team roles are included and users pay only their underlying model providers.⁵

Self-hosting is also available without a sales call under Apache 2.0. An Enterprise option is described as custom and aimed at longer retention, support agreements, procurement requirements, security questionnaires and roadmap needs.⁵

This is excellent short-term value for experimentation. It is not a stable long-term price signal. AcruxCore explicitly says post-beta pricing has not been set. A team evaluating total cost should therefore separate today’s zero platform fee from the future economics of hosted retention, support and enterprise controls.

The self-hosted option is strategically important because it gives teams an exit route if future hosted pricing becomes unattractive, although self-hosting transfers infrastructure, upgrades, backups and security operations to the customer.

AcruxCore’s privacy policy, updated August 1, 2026, states that the hosted service may collect account data, prompts, tool definitions, API keys/provider credentials, trace metadata, usage data and—when a team enables payload capture—full request and response payloads.⁸

The policy says provider credentials and secrets are encrypted at rest; traffic content is sent to the model provider selected by the customer; and AcruxCore does not use prompts or completions to train models. Full payload capture is opt-in rather than necessarily enabled for every trace.⁸

For self-hosted deployments, AcruxCore says the customer is the controller of data in its own deployment and the hosted-service privacy policy does not apply.⁸

AcruxCore’s security page says records are scoped by team with row-level access controls; API requests are authenticated and authorized per team; provider secrets are encrypted at rest and not returned in full through the API or dashboard; traffic is protected with TLS; and payload capture can be controlled by the team.⁹

These controls are appropriate for an LLM gateway, particularly because provider credentials and potentially sensitive prompts sit in its trust boundary. The self-hosted option can reduce exposure to the hosted service by keeping the database, keys and network inside the customer environment.⁵

The Terms also set a low default liability ceiling for the hosted service: the greater of fees paid in the previous 12 months or $100, subject to applicable law.¹⁰ That is common for an early free beta but relevant for production procurement because the technical role of the gateway can be operationally important.

AlternativeTo currently contains two comments/reviews and two likes. One five-star reviewer says they tested AcruxCore, liked having prompts, traces and evaluations in one place, and found prompt versioning useful, while also noting that there is a lot to explore.⁷

AcruxCore is most interesting for small and medium AI engineering teams that are currently stitching together prompt files, provider wrappers, logs and ad-hoc evaluation scripts and want one coherent control plane.

It is also attractive for teams that prioritize self-hosting and source availability, or that want a low-friction OpenAI-compatible gateway without buying model tokens from another intermediary.

The current beta is a sensible environment for pilots, internal tools and non-critical systems. Production teams with strict uptime, compliance or procurement requirements should validate scale, failure handling, support, backup/restore, access control and security evidence before placing AcruxCore in the critical request path.