Frequently asked questions

What is AcruxCore?

AcruxCore is an LLMOps platform for managing applications built on large language models. It includes prompt versioning, an OpenAI-compatible LLM gateway, tracing, a tool catalog, and evaluation. The gateway records each model request that passes through it. If you do not use the gateway, your application can send its traces to AcruxCore directly.

AcruxCore is licensed under Apache 2.0, so you can self-host the platform.

What problems does AcruxCore solve?

AcruxCore lets a team change a prompt without redeploying the application. It also shows what happened inside each model call. Prompt versions live in one place that the whole team can edit.

The gateway records each request that passes through it, so you do not depend only on the traces your application reports.

The tool catalog stores tool definitions next to your prompts. Attach a tool to a prompt and the model can call it. The prompt then works as an agent, with no extra code to connect the tool and the prompt. See the connect a tool to a prompt guide.

What is the difference between a trace and an audit event?

A trace records traffic your application sent: one model call, its prompt, its tokens, its latency and its cost. An audit event records a change a person made, and names that person.

A trace answers why one model answer came out wrong. An audit event answers who changed the prompt on Tuesday, and what it said before. A model answer never becomes an audit event, and a role change never appears in a trace.

AcruxCore keeps traces and audit events, and both are on by default. See core concepts for how traces, prompts, versions and audit events relate.

Can I use AcruxCore without the gateway?

Yes. Call your model provider directly with your own API key while using AcruxCore for prompt versioning and tracing. Follow the gateway-free RAG tutorial for a complete example.

If your app already uses OpenTelemetry, it can send traces to the OTLP endpoint. Gateway features such as caching, budgets, virtual keys, and automatic fallbacks require traffic to pass through the gateway.

How does AcruxCore compare with Langfuse, Phoenix, Opik, Helicone, MLflow, and Laminar?

AcruxCore combines an LLM gateway with prompt management, tracing, and evaluation. The gateway applies routing, caching, budgets, and virtual keys before calls reach the provider. Helicone and MLflow also have gateways. Langfuse, Phoenix, Opik, and Laminar have no gateway, and collect traces from the calls your application makes.

AcruxCore versions prompts and tools in the same place. The gateway runs those tools, and tracing records each model call. AcruxCore is also the only one of these platforms with an audit trail that costs nothing extra.

The comparison matrix puts the seven platforms side by side on ten criteria, with a source for each cell. The hands-on test of nine platforms covers these seven, plus LangSmith and PromptLayer, in more detail.

What are AcruxCore’s key strengths?

AcruxCore’s first strength is its tool catalog. Each tool has a version history, and the gateway runs the tool when a model calls it. The second strength is prompt templates with conditions and loops. AcruxCore also has a prompt optimizer, user feedback linked to traces, and an audit trail that costs nothing extra.

The versioned tool catalog stores each tool with its version history, and the gateway runs the tool when a model calls it. No other platform in our comparison keeps versions of a tool and also runs it. Langfuse and Laminar store a tool schema inside a playground, and nothing runs it. Phoenix, Opik, and Helicone have no tool catalog. MLflow catalogs whole MCP servers, each a bundle of tools, rather than single tools.

Tools can connect to HTTP endpoints that the gateway calls for you. Add an existing endpoint as a tool without redeploying your application.

AcruxCore also supports prompt templates with conditional logic. Jinja2 if statements and for loops keep the branching inside the template. MLflow also supports conditions and loops. Langfuse, Phoenix, Opik, Helicone, and Laminar only substitute variables.

The prompt optimizer drafts rewrites and evaluates them against the same test cases as your production prompt. You choose which candidate to promote as a new version. Feedback links to a trace or to one span inside it. Each entry is labeled as an end-user rating or a developer note, so you can build a dataset from user feedback alone.

The team-wide audit trail records every change with the name of the person who made it. The trail covers API keys, members and roles, provider connections, virtual keys, budgets, secrets, prompts, and tools. Hosted or self-hosted, the trail is on by default. You can filter it by area, by event type, or by person.

A prompt optimizer and linked feedback are not unique to AcruxCore. Other platforms, Opik among them, also have optimizers and evaluation tools. The audit trail is different. AcruxCore is the only platform in our comparison where the trail costs nothing extra.

Which LLMOps platforms have an audit log of who changed what?

AcruxCore has an audit trail that records every change with the name of the person who made it, whether hosted or self-hosted. The trail is on by default and costs nothing extra.

Each prompt and each tool has its own Audit tab. The tab lists when that prompt or tool was created or renamed, when a version was committed, and when an alias was promoted or deleted. Owners and admins also get one team-wide trail in the dashboard. That trail covers the whole team, including the changes that belong to no single prompt or tool.

The team-wide trail holds 39 event types across multiple areas. The areas are API keys, members and roles, invites, provider connections, virtual keys, budgets, gateway models, secrets, trace settings, datasets and evaluation rules, and prompt and tool changes. You can filter the trail by area, by one event type, or by person. Entries stay in the trail after the person who made them leaves the team.

Langfuse has an audit log, but it needs the Enterprise plan at $2,499 a month, even if you self-host. In Phoenix, Opik, Helicone, MLflow, and Laminar we found no audit log in the settings pages we opened.

The audit trail guide shows the dashboard screens, and the audit API returns the same events to your own code. Each cell in the comparison matrix has a check date. A platform can add an audit log after that date.

How much latency does the AcruxCore gateway add?

In our published benchmark, AcruxCore added about 42 ms of latency to each call compared with calling OpenAI directly. The confidence interval was +17 ms to +81 ms. Actual overhead depends on your setup and workload.

The benchmark write-up includes the latency distribution, the tail latency, and the test script, which you can run against your own workload.

When should I consider an AcruxCore alternative?

Pick something else if you need an organization or workspace above teams, or content filtering in the gateway. Also pick something else if you need more than one role per member, or a larger community. Langfuse has an organization above its projects, and Laminar has a workspace above its projects. AcruxCore has one team with nothing above it.

MLflow’s gateway can block content and personal data at each endpoint. AcruxCore has no such filter. AcruxCore also gives each member exactly one role at a time.

Langfuse and MLflow have larger GitHub communities than AcruxCore.

How can I verify AcruxCore’s comparisons?

Our comparisons link a source for each fact and include screenshots from our test sessions. We self-host each platform where we can, then build the same example prompt in that platform and in AcruxCore.

Every cell in the comparison matrix has a source link and a check date.

The latency test scripts are in the AcruxCore repository, so you can rerun the benchmark in your own environment.

How does AcruxCore’s open-source license compare?

AcruxCore uses Apache 2.0 for the whole platform, with no folder under a separate paid license.

Opik, Helicone, MLflow, and Laminar use Apache 2.0 on the same terms. Langfuse is MIT at its core, and some features sit in a folder under an enterprise license. Phoenix uses the Elastic License 2.0, a source-available license that the Open Source Initiative has not approved.

Can I self-host AcruxCore?

Yes. Run docker compose up to self-host the same code used by the hosted platform. There is no separate community edition and no folder under a separate paid license.

Does AcruxCore store my prompts and responses?

Hosted AcruxCore stores the request and response body of each call so you can inspect the trace. This stored body is called the payload, and you can turn payload capture off for the whole team or for single traces. The payload capture guide explains both options.

When you self-host, AcruxCore runs on your own servers. Self-hosting does not stop data going to the model providers you connect. Your own configuration decides what each call sends them. For hosted data handling, see the security page.

Can I export my prompts and traces?

Yes. Export your prompt library as JSON, including its version history, and import it again when needed. The diff, export and import guide explains the process. You can also retrieve traces through the REST API.

Is the AcruxCore audit log free, or does it need a paid plan?

It is free. The audit trail is part of the platform whether hosted or self-hosted. Self-hosting costs nothing beyond your own servers, because the whole platform is Apache 2.0 with no folder under a separate paid license.

AcruxCore is the only platform we compared whose audit trail costs nothing extra.

The trail has no CSV or JSON export yet, and you cannot set how long entries are kept. Events are never deleted. You read them in the dashboard or through the audit API.

How much does AcruxCore cost? Is there a token markup?

Self-hosted AcruxCore has no license fee, so you pay only for your own servers and for the model providers you use. The full platform is Apache 2.0, with no folder under a separate paid license.

Hosted AcruxCore is free during beta, with no credit card required at sign-up.

AcruxCore does not mark up model tokens. Connect your own provider credentials and pay the provider’s rates. The gateway records the cost of each call and adds nothing to it.

Hosted pricing after the beta is not announced yet. The pricing page will show it first, and existing accounts get notice before any price changes.