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

The platform is available both as a hosted beta and as a self-hosted open-source deployment. AcruxCore says the complete platform is Apache-2.0 licensed, while its TypeScript and Python SDK packages are separately published under MIT licenses.³⁴⁵ That distinction resolves the earlier ambiguity in AiToolMap’s database: the current product itself is explicitly open source, not merely accompanied by open client libraries.

This is a desk review. AiToolMap did not deploy AcruxCore, route production traffic through its gateway, benchmark latency, test failure modes, audit the source code or validate its security controls independently.

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.⁷⁵

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.⁷

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.

The main risk is organizational rather than conceptual. Once prompts become mutable production configuration, access control, approval processes and change auditing matter. AcruxCore says team roles and audit records exist, but AiToolMap did not test whether those controls are sufficient for regulated or high-assurance deployments.

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.³

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.³

The limitation is that AcruxCore sits in a sensitive position in the request path. In the hosted configuration, model traffic can pass through its gateway and provider credentials are stored by the service. A gateway outage, routing defect or security failure can therefore affect every model call behind it. The current Terms explicitly do not guarantee uninterrupted operation, and AiToolMap found no independent availability history or published service-level statistics.⁸

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.⁴

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.⁵ npm currently exposes a newer 0.8.0 TypeScript package.⁴ 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

However, the fixed panel contains no independent AcruxCore evaluation benchmark, and AiToolMap did not execute a dataset experiment. There is therefore no basis to claim that its evaluator, experiment statistics or scoring workflows are superior to established LLM observability platforms.

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.³¹ 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.⁶

The SDK packages have their own licensing metadata. npm lists the TypeScript SDK as MIT, and PyPI lists the Python client as MIT.⁴⁵ 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.⁶ 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.³

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.³

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.

All 50 mandatory members of `ai-review-panel-2026-08-v4` were audited under methodology v1.8. No fixed-panel source produced current exact-product evidence for AcruxCore. That result is not treated as negative evidence. AcruxCore is a new, specialist infrastructure product and most panel members concentrate on frontier models, mainstream AI applications or mature software categories rather than small LLM-ops projects.

The absence of panel coverage does matter for confidence. There is no current G2, Capterra, Gartner Peer Insights, TrustRadius, PeerSpot or equivalent review population to corroborate ease of deployment, reliability or support. Model benchmarks such as Artificial Analysis, LM Arena, SWE-bench, HELM and ARC Prize are also inapplicable: AcruxCore does not supply a model whose capability can inherit those scores.

AiToolMap therefore expanded beyond the fixed panel to the current AcruxCore site and documentation, npm and PyPI package registries, and AlternativeTo. The latter provides only a tiny user signal and much of its product description appears vendor-supplied, so it is used cautiously.⁶

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.⁹

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.⁹

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.⁹

This is a relatively clear disclosure, but enterprise buyers still need to inspect actual deployment configuration, subprocessors, retention settings, backup behavior and any data-processing agreement. AiToolMap did not find independent privacy certification evidence in the fixed panel.

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.¹⁰

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.³

The caveat is evidence maturity. AiToolMap found no current fixed-panel validation of AcruxCore’s security, no independent penetration-test report in the reviewed evidence and no basis to claim SOC 2 or ISO 27001 certification. The current security claims should be treated as first-party statements, not audited guarantees.

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.⁸ 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.⁶

That is positive but far too small a sample for a numeric quality conclusion. The page also states that AcruxCore itself added the listing and much of the detailed product description is vendor-authored. AiToolMap therefore treats the user comment as qualitative early evidence only and does not use a 5/5 value as a rating input with meaningful weight.

There is no current Trustpilot signal to use even qualitatively, and no G2/Capterra population. Under v1.8, absence stays absence: it is not converted into a low user-satisfaction score.

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.