# AcruxCore

> AcruxCore is an open-source platform for teams building on LLMs: version your prompts, route every model call through one OpenAI-compatible gateway, trace what happened, catalog the tools your models can call, evaluate the result, and read an audit trail of every change a person made. Apache 2.0, self-hostable, with published SDKs for Python and TypeScript.

The gateway sits in the request path, so a trace records the request that was actually sent rather than a copy reported alongside it. That is the design decision most of these pages come back to.

This documentation is Apache 2.0 and free to use, quote, and train on. Pages are written to be read in full: guides and tutorials each carry the real commands and the real output from a run against a live server, and every benchmark links the script that produced its numbers.

The product site is https://acruxcore.com. The source is at https://github.com/AcruxCore/AcruxCore.

## Start here

The shortest path to understanding what AcruxCore does and how it compares.

- [9 LLM Observability Tools Tested Hands-On (2026)](https://docs.acruxcore.com/blog/hands-on-llm-ops-comparison): We hands-on tested 9 LLM observability tools — Langfuse, Opik, MLflow, Laminar and more — comparing tracing, evals and measured latency, with screenshots of each.
- [Quickstart](https://docs.acruxcore.com/docs/getting-started/quickstart): Sign up, register a model, author a prompt, and make your first traced LLM call through the AcruxCore gateway in about ten minutes.
- [Core concepts](https://docs.acruxcore.com/docs/getting-started/core-concepts): The ideas behind AcruxCore: prompt versions and aliases, gateway model resolution, traces and spans, the tool catalog and its executors, and evaluation.
- [Full-cycle latency across six LLM-ops platforms, measured against real OpenAI](https://docs.acruxcore.com/blog/full-cycle-latency-benchmark): We timed prompt-fetch-plus-completion, not just the model call, across Opik, MLflow, Langfuse, Helicone, Phoenix, and AcruxCore's gateway and gateway-free BYOK modes, against a real OpenAI baseline — then ran it four times to see what's stable and what isn't.
- [Build a tool-calling agent in the dashboard (no code)](https://docs.acruxcore.com/docs/tutorials/build-a-tool-calling-agent-in-the-dashboard-no-code): Assemble a movie assistant that calls a live public API — two HTTP tools, a secret, and a model — entirely in the dashboard, with no code to run.

## Getting started

What AcruxCore is, the vocabulary, and a first traced call.

- [Introduction](https://docs.acruxcore.com/docs/getting-started/introduction): AcruxCore is one platform where you version prompts, route LLM calls through a gateway, trace every request, keep a catalog of tools, and evaluate quality.
- [Quickstart](https://docs.acruxcore.com/docs/getting-started/quickstart): Sign up, register a model, author a prompt, and make your first traced LLM call through the AcruxCore gateway in about ten minutes.
- [Product tour](https://docs.acruxcore.com/docs/getting-started/product-tour): Attach a tool, stream a response, inspect the trace, leave feedback, and turn that feedback into an evaluation. This page covers the steps that come after your first call.
- [Core concepts](https://docs.acruxcore.com/docs/getting-started/core-concepts): The ideas behind AcruxCore: prompt versions and aliases, gateway model resolution, traces and spans, the tool catalog and its executors, and evaluation.

## Guides

One AcruxCore capability per page, start to finish, with the exact requests and the real output.

- [Create a tool](https://docs.acruxcore.com/docs/guides/create-a-tool): Put a tool in the catalog two ways — a decorated function whose code owns the definition, or a catalog version your code only implements — and pick the owner that fits the project.
- [Evaluate a prompt against a dataset](https://docs.acruxcore.com/docs/guides/evaluate-a-prompt): Build a dataset from real user feedback or a CSV/JSON test set, run an experiment that sweeps prompt versions and models, and read the run report.
- [Invite a teammate to your team](https://docs.acruxcore.com/docs/guides/invite-a-teammate): Create a single-use invite link, choose the role it grants, and email it straight to a new teammate to get them into your AcruxCore team.
- [Route your app's LLM calls through the gateway](https://docs.acruxcore.com/docs/guides/route-calls-through-the-gateway): Add a provider credential, register a model, test it in the Playground, and call the OpenAI-compatible gateway from curl, the SDK, or the OpenAI client.
- [Trace and inspect an LLM call](https://docs.acruxcore.com/docs/guides/trace-an-llm-call): Every gateway call is traced automatically. Inspect spans, read dashboards, and report your own custom spans from app code with the SDK.
- [Version a prompt and ship it to production](https://docs.acruxcore.com/docs/guides/version-a-prompt): Create a prompt, commit versions, promote an alias to production, and render it from your app — the full prompt lifecycle in AcruxCore.
- [Chat, stream, and collect feedback with the SDK](https://docs.acruxcore.com/docs/guides/use-the-sdk-for-chat-and-feedback): Call the gateway from the SDK with chat() and streaming, run a traced tool-calling loop, then read traces back and record feedback — in Node and Python.
- [Connect a tool to a prompt](https://docs.acruxcore.com/docs/guides/connect-a-tool-to-a-prompt): Bind a catalog tool to a prompt by choosing the tool alias it follows, give one prompt alias its own tools, and see when the tool's code changed underneath you.
- [Improve a prompt from feedback](https://docs.acruxcore.com/docs/guides/improve-a-prompt-from-feedback): Turn replies your team disagreed with into candidate rewrites of your production prompt, score them against what is live, and promote the one that earns it.
- [Manage team roles and permissions](https://docs.acruxcore.com/docs/guides/manage-team-roles-and-permissions): See who's on your team, change an existing member's role, and remove their access — the three actions inviting a teammate doesn't cover.
- [Use conditional logic in prompt templates](https://docs.acruxcore.com/docs/guides/use-conditional-logic-in-prompt-templates): Write Jinja2-style {% if %} conditionals and {% for %} loops inside a prompt template, rendered by AcruxCore's nunjucks-based engine.
- [Using sessions and traces](https://docs.acruxcore.com/docs/guides/using-sessions-and-traces): Group related prompt calls into a session, follow one call into its trace, leave feedback, and jump back to the prompt that produced it.
- [Call a prompt's tools from the SDK](https://docs.acruxcore.com/docs/guides/call-a-prompts-tools-from-the-sdk): Every way to run the tools bound to a prompt from the SDK — the two-line shortcut, the loop by hand, and one raw completion — streaming or not.
- [Diff, export, and import your prompt library](https://docs.acruxcore.com/docs/guides/diff-export-and-import-your-prompt-library): Read a unified diff between two prompt versions, export one as a portable JSON file, and re-import it — including what import does and does not preserve.
- [Evaluate a prompt with conversation history](https://docs.acruxcore.com/docs/guides/evaluate-a-prompt-with-conversation-history): Carry a session's prior turns into a dataset example so a run, the judge, and the optimizer see the same conversation the model actually answered in.
- [Read the team audit trail](https://docs.acruxcore.com/docs/guides/read-the-team-audit-trail): See every recorded change in your team — API keys, members, gateway, secrets, prompts and tools — and filter it down to the one action you need to account for.
- [Scope access with virtual keys](https://docs.acruxcore.com/docs/guides/scope-access-with-virtual-keys): Create a gateway key restricted to one model, a rate cap, and a cache TTL, so a leaked key can't spend outside its job.
- [Tag and filter traces](https://docs.acruxcore.com/docs/guides/tag-and-filter-traces): Attach tags and metadata to traces, filter them by tag or metadata key/value, and discover what tags your team is using.
- [Look up the traces a prompt version produced](https://docs.acruxcore.com/docs/guides/look-up-the-traces-a-prompt-version-produced): Find every trace a specific prompt version powered, from the Versions tab or the API, and know when an empty result is expected.
- [Manage prompts via the SDK](https://docs.acruxcore.com/docs/guides/manage-prompts-via-the-sdk): Create a prompt, commit versions with templated messages, diff, promote aliases, export/import, and delete — the full lifecycle through curl or the SDK, no dashboard required.
- [Score live traffic with an evaluation rule](https://docs.acruxcore.com/docs/guides/score-live-traffic-with-an-evaluation-rule): Create a standing rule that judges matching production gateway calls automatically, with no dataset and no manual run.
- [Set spend limits with gateway budgets and rate limits](https://docs.acruxcore.com/docs/guides/set-spend-limits-with-gateway-budgets-and-rate-limits): Cap spend team-wide or per key by day, week, month, or total, and pair it with a request-rate cap — then prove both actually reject a call before it costs anything.
- [Version and track a tool](https://docs.acruxcore.com/docs/guides/version-and-track-a-tool): Commit a second tool version, move staging ahead of production, read real call volume, error rate and latency, and retire a tool you no longer call.
- [Evaluate a prompt that reads an image](https://docs.acruxcore.com/docs/guides/evaluate-a-prompt-that-reads-an-image): Store a prompt with an image part, fill the image from a dataset example, run it on several vision models, and read the scores a judge gives after looking at the same image.
- [Report a tool failure from your own code](https://docs.acruxcore.com/docs/guides/report-a-tool-failure-from-your-own-code): A tool call can fail while the HTTP status says 200. Return a ToolResult from your tool and the failure is recorded as a red span instead of a green one.
- [Retries and fallbacks: what the gateway does when a provider call fails](https://docs.acruxcore.com/docs/guides/automatic-model-fallbacks): Three separate things retry a failing LLM call, and each is set in a different place. Configure all three, then read two real traces to see which one fired.
- [Store prompts and tools via the API](https://docs.acruxcore.com/docs/guides/store-prompts-and-tools-via-api): Save a prompt and its tools over the REST API, confirm them in the dashboard, then fetch and stream them from the SDK.
- [View trace analytics](https://docs.acruxcore.com/docs/guides/view-trace-analytics): Read aggregate volume, error-rate, token, cost, and latency metrics over every trace your team has recorded, grouped by day, model, session, or prompt version.
- [Configure trace payload capture](https://docs.acruxcore.com/docs/guides/configure-trace-payload-capture): Read and toggle your team's default for whether a trace's request/response bodies get stored on ingest, and override that default for a single request.
- [Send an image to a model](https://docs.acruxcore.com/docs/guides/send-an-image-to-a-model): Pass an image to a vision model through the gateway, either as a link or as a file, and check what the trace stores.
- [Write a custom judge prompt](https://docs.acruxcore.com/docs/guides/write-a-custom-judge-prompt): Grade experiment runs and evaluation rules with your own rubric, including examples that have no criteria, and see exactly what the judge model receives.
- [Send OTel traces to AcruxCore with the SDK helper](https://docs.acruxcore.com/docs/guides/send-otel-traces-with-the-sdk-helper): Wire any OpenTelemetry-instrumented app to AcruxCore's OTLP endpoint in one call with register() — a bare pipeline, one framework, several frameworks, and session grouping, in Python or Node.

## Tutorials

End-to-end builds of something that exists outside AcruxCore — agents, assistants, analysts. Every one ships a runnable script and a notebook.

- [Start here](https://docs.acruxcore.com/docs/tutorials): Eight end-to-end agent builds, arranged from no-code to multi-agent — and a map for picking the one that matches how you want to work.
- [Build a tool-calling agent in the dashboard (no code)](https://docs.acruxcore.com/docs/tutorials/build-a-tool-calling-agent-in-the-dashboard-no-code): Assemble a movie assistant that calls a live public API — two HTTP tools, a secret, and a model — entirely in the dashboard, with no code to run.
- [Build a tool-calling agent in Python (no SDK)](https://docs.acruxcore.com/docs/tutorials/build-a-tool-calling-agent-in-python-no-sdk): Register a credential, model, prompt, and tool, then drive a client-side tool-calling loop from plain Python over the REST API — one trace, streaming and non-streaming.
- [Build a tool-calling agent in Python (SDK)](https://docs.acruxcore.com/docs/tutorials/build-a-tool-calling-agent-in-python-sdk): Set up a model, a client tool, and a prompt in the dashboard, then drive a text-to-SQL data analyst from Python with the async SDK's run_prompt_with_tools — one trace, real SQLite queries, and streaming.
- [Build a RAG agent without the gateway](https://docs.acruxcore.com/docs/tutorials/build-a-rag-agent-without-the-gateway): Answer questions over your own documents using your own OpenRouter key — calling the provider directly, while prompts, tools, and traces keep working.
- [Build a ReAct agent](https://docs.acruxcore.com/docs/tutorials/build-a-react-agent): A finance research agent that reasons step by step and calls two real tools, driven by a manual tool-calling loop straight against your own model provider — no gateway.
- [Build a configurable ReAct agent](https://docs.acruxcore.com/docs/tutorials/build-a-configurable-react-agent): A web-research agent whose model, persona, and search depth all change at call time — just by swapping one alias name, no redeploy and no code change.
- [Build a supervisor multi-agent system](https://docs.acruxcore.com/docs/tutorials/build-a-supervisor-multi-agent-system): A router that classifies an incoming request and hands it to one of three specialist subagents — finance research, general research, or writing — sharing one trace.
- [Build a Medical-Information QA Agent](https://docs.acruxcore.com/docs/tutorials/build-a-medical-information-qa-agent): A typed, cited, policy-aware medical-information agent that refuses off-label questions and escalates adverse events, enforced with the gateway's response_format feature — showing both dict and pydantic/zod forms.
- [Trace a CrewAI Trip-Planning Crew](https://docs.acruxcore.com/docs/tutorials/trace-a-crewai-trip-planner): Build a two-agent CrewAI crew that plans and revises a trip, and watch every agent, tool, and LLM call land in AcruxCore automatically over OTLP — no code change to the crew itself.
- [Trace an OpenAI Agents SDK Support-Triage System](https://docs.acruxcore.com/docs/tutorials/trace-an-openai-agents-sdk-triage-system): Build a support-triage system with the OpenAI Agents SDK's agent handoffs, and watch the full triage-to-specialist flow land in AcruxCore automatically over OTLP.
- [Build a travel planner agent](https://docs.acruxcore.com/docs/tutorials/build-a-travel-planner-agent): Give one prompt three tools and let the model decide which to call - flights, weather, currency, or nothing at all.
- [Trace a LangChain Research Agent](https://docs.acruxcore.com/docs/tutorials/trace-a-langchain-research-agent): Build a two-tool LangChain agent in Python or Node, and watch every chain, tool, and model call land in AcruxCore automatically over OTLP — no AcruxCore code inside the agent.

## Benchmarks and comparisons

Measured write-ups and hands-on comparisons against other LLM-ops platforms. Every number comes from a script committed to the repository, and every competitor screenshot from a real run on that competitor.

- [Braintrust alternative: what changes when the platform runs your tool code](https://docs.acruxcore.com/blog/acruxcore-vs-braintrust): The same prompt and the same tool run on both platforms, with real screenshots. A tool was pushed and executed on Braintrust's runtime, and running their Python SDK turned up a templating bug.
- [Laminar alternative: SQL over traces vs a prompt gateway](https://docs.acruxcore.com/blog/acruxcore-vs-laminar): A Laminar alternative tested hands-on — one example on self-hosted Laminar and AcruxCore, with real screenshots from both and a 100-round latency benchmark.
- [Full-cycle latency across six LLM-ops platforms, measured against real OpenAI](https://docs.acruxcore.com/blog/full-cycle-latency-benchmark): We timed prompt-fetch-plus-completion, not just the model call, across Opik, MLflow, Langfuse, Helicone, Phoenix, and AcruxCore's gateway and gateway-free BYOK modes, against a real OpenAI baseline — then ran it four times to see what's stable and what isn't.
- [MLflow alternative: two gateways, one with guardrails](https://docs.acruxcore.com/blog/acruxcore-vs-mlflow): Considering an MLflow alternative for LLM work? We timed the same prompt through both gateways — MLflow's added +135 to +225 ms, ours +21 to +43 ms.
- [Helicone alternative: two request-path proxies, measured](https://docs.acruxcore.com/blog/acruxcore-vs-helicone): Weighing a Helicone alternative? Both sit in the request path. We rebuilt one prompt on each, and timed 300 benchmark rounds through both gateways.
- [Opik alternative: guardrails, PII and online evaluation](https://docs.acruxcore.com/blog/acruxcore-vs-opik): An Opik alternative tested hands-on — one prompt rebuilt on self-hosted Opik and AcruxCore, with real screenshots, an SDK trace and a latency benchmark.
- [Arize Phoenix alternative: OTel tracing vs a gateway](https://docs.acruxcore.com/blog/acruxcore-vs-phoenix): An Arize Phoenix alternative tested hands-on — one prompt rebuilt on Phoenix and AcruxCore, with real screenshots, an SDK trace and a latency benchmark.
- [Langfuse alternative: org hierarchy vs a flat team](https://docs.acruxcore.com/blog/acruxcore-vs-langfuse): A Langfuse alternative tested hands-on — the same prompt rebuilt and run on both, with real screenshots, an SDK trace and a measured latency benchmark.
- [Exact-match gateway caching: latency and cost saved](https://docs.acruxcore.com/blog/exact-match-gateway-caching): We measured the gateway's exact-match response cache against the calls it's built for — a ~100x latency drop on a hit, and zero cost, from a live run.
- [One gateway, three providers: request shapes compared](https://docs.acruxcore.com/blog/one-gateway-three-providers): The same chat request, translated into Anthropic, Gemini, and OpenAI's own shapes by the gateway's adapters — headers, params, and tool defs, side by side.
- [9 LLM Observability Tools Tested Hands-On (2026)](https://docs.acruxcore.com/blog/hands-on-llm-ops-comparison): We hands-on tested 9 LLM observability tools — Langfuse, Opik, MLflow, Laminar and more — comparing tracing, evals and measured latency, with screenshots of each.
- [LangChain vs AcruxCore SDK: the same tool-calling agent](https://docs.acruxcore.com/blog/tool-agent-sdk-langchain-vs-acruxcore): We rewrote the same tool-calling agent with LangChain's create_agent and the AcruxCore SDK's run_tool_loop, ran both live, and compared the code and traces.
- [Tool-call traces: LangSmith vs AcruxCore](https://docs.acruxcore.com/blog/tool-calling-traces-langsmith-vs-acruxcore): We built the same weather tool twice — a traced Python function on LangSmith, a versioned Tool Catalog entry on AcruxCore — ran both live, and counted the code.
- [A hands-on walkthrough of the AcruxCore dashboard](https://docs.acruxcore.com/blog/acruxcore-hands-on-walkthrough): A step-by-step walkthrough of prompt versioning, tracing, and evaluation in AcruxCore — the baseline for our competitor comparisons.
- [Hands-on with Langfuse: prompts, tracing, and datasets](https://docs.acruxcore.com/blog/langfuse-hands-on-walkthrough): A step-by-step walkthrough of Langfuse's prompt management, trace view, and dataset/experiment flow, with screenshots and honest observations.
- [Hands-on with LangSmith: a real walkthrough](https://docs.acruxcore.com/blog/langsmith-hands-on-walkthrough): A step-by-step, screenshot-backed walkthrough of creating a prompt, running a trace, and building an eval dataset in LangSmith's hosted product.
- [A hands-on walkthrough of PromptLayer](https://docs.acruxcore.com/blog/promptlayer-hands-on-walkthrough): Logging into hosted PromptLayer, creating and versioning a prompt, running it, and checking what its trace view and evaluation tooling actually show.
- [How much overhead does an LLM gateway add?](https://docs.acruxcore.com/blog/llm-gateway-overhead): We benchmarked five ways to reach the same OpenAI model — gateway, BYOK, and raw. The gateway's software costs about 42 ms; the rest is network distance.
- [LangSmith alternative: prompts, gateway and tracing](https://docs.acruxcore.com/blog/acrux-core-vs-langsmith): A LangSmith alternative compared feature by feature — prompt versioning, a request-path gateway, tracing, tools and evals, open source and self-hostable.

## SDK reference

Every method and parameter in the two published SDKs.

- [Node SDK reference](https://docs.acruxcore.com/docs/sdk-reference/node): Every method and parameter in @acruxcoreai/sdk, the Node/TypeScript SDK for AcruxCore — render, chat, tool loops, tracing, feedback, and the tool catalog.
- [Python SDK reference](https://docs.acruxcore.com/docs/sdk-reference/python): Every method and parameter in acruxcore, the Python SDK for AcruxCore — render, chat, tool loops, tracing, feedback, and the tool catalog.

## Optional

The REST API, one page per domain. Every endpoint here was verified with curl against a running server, and each entry shows the exact request and the exact response it returned.

- [API Keys](https://docs.acruxcore.com/api-reference/api-keys): Create, list, and revoke personal API keys — the Bearer tokens every other endpoint authenticates with.
- [API Reference](https://docs.acruxcore.com/api-reference): The AcruxCore REST API — every endpoint is curl-verified. Authentication, prompts and versions, the gateway, tracing, tools, and evaluation.
- [Audit](https://docs.acruxcore.com/api-reference/audit): Read the paginated, newest-first audit trail of changes made to a prompt, a tool, or the whole team.
- [Authentication](https://docs.acruxcore.com/api-reference/authentication): Generate a Bearer token in the dashboard, then use it on every endpoint in this reference.
- [Datasets](https://docs.acruxcore.com/api-reference/datasets): Create and manage example sets used by experiments, evaluations, and the optimize loop.
- [Evaluation Rules](https://docs.acruxcore.com/api-reference/eval-rules): Score live production traffic automatically with a standing LLM-judge rule.
- [Experiments & Runs](https://docs.acruxcore.com/api-reference/experiments): Sweep a dataset across prompt versions and models, then read back per-cell results and reports.
- [Export](https://docs.acruxcore.com/api-reference/export): Download a prompt version as a portable JSON file for backup or re-import.
- [Gateway Budgets](https://docs.acruxcore.com/api-reference/gateway/budgets): Cap spend team-wide or per virtual key, by day, week, month, or total.
- [Gateway Chat Completions](https://docs.acruxcore.com/api-reference/gateway): Call the OpenAI-compatible gateway endpoint with raw messages or a stored-prompt reference.
- [Gateway Connections](https://docs.acruxcore.com/api-reference/gateway/connections): Store BYOK provider credentials (OpenAI, Anthropic, Gemini, OpenAI-compatible), encrypted at rest.
- [Gateway Models](https://docs.acruxcore.com/api-reference/gateway/models): Register a public model name that maps to a credential and an upstream provider model.
- [Gateway Usage & Lineage](https://docs.acruxcore.com/api-reference/gateway/usage): Read gateway request logs and usage analytics, including which prompt version produced each call.
- [Gateway Virtual Keys](https://docs.acruxcore.com/api-reference/gateway/keys): Create scoped virtual keys that authenticate gateway calls without exposing your provider credentials.
- [Import](https://docs.acruxcore.com/api-reference/import): Create a new prompt and its first version from a previously exported JSON file.
- [Notifications](https://docs.acruxcore.com/api-reference/notifications): Read and change which email notifications you receive, per team — plus the unauthenticated one-click unsubscribe endpoints mail clients call.
- [Optimize Loop](https://docs.acruxcore.com/api-reference/optimize): Draft, score, and promote LLM-generated prompt rewrites against a dataset.
- [OTLP Trace Ingestion](https://docs.acruxcore.com/api-reference/traces/otlp): Point any OpenTelemetry exporter — CrewAI, LangChain, LlamaIndex, or a custom OTel SDK — straight at AcruxCore.
- [Prompt Aliases](https://docs.acruxcore.com/api-reference/prompts/aliases): List, render, and promote named pointers (production/staging) to a specific prompt version.
- [Prompt Diff](https://docs.acruxcore.com/api-reference/prompts/diff): Get a unified diff between two versions of a prompt's raw template.
- [Prompt Tool Bindings](https://docs.acruxcore.com/api-reference/prompts/tool-bindings): Connect a tool to a prompt by choosing the tool alias each prompt alias should call — one live write, no new prompt version.
- [Prompt Versions](https://docs.acruxcore.com/api-reference/prompts/versions): Commit immutable prompt versions and fetch them back by number.
- [Prompts](https://docs.acruxcore.com/api-reference/prompts): Read, search, and resolve prompts by API — create/edit/delete happen in the dashboard today.
- [Saved Views](https://docs.acruxcore.com/api-reference/traces/views): Name a filter set on the trace or feedback list so the whole team can apply it in one click.
- [Sessions](https://docs.acruxcore.com/api-reference/traces/sessions): Group related traces under a shared session id and read them back as one rolled-up view.
- [Tool Aliases](https://docs.acruxcore.com/api-reference/tools/aliases): List, render, and promote named pointers to a specific tool version.
- [Tool Execute](https://docs.acruxcore.com/api-reference/tools/execute): Invoke an HTTP-backed tool version server-side and record the call.
- [Tool Resolve](https://docs.acruxcore.com/api-reference/tools/resolve): Resolve tool names to model-ready function definitions, in one batched call.
- [Tool Sync](https://docs.acruxcore.com/api-reference/tools/sync): Create-or-commit-or-nothing: reconcile a tool defined in code with the catalog in one call.
- [Tool Versions](https://docs.acruxcore.com/api-reference/tools/versions): Commit immutable tool versions and fetch them back by number.
- [Tools](https://docs.acruxcore.com/api-reference/tools): Create, list, update, and delete tools in the Tool Catalog.
- [Tools Analytics](https://docs.acruxcore.com/api-reference/tools/analytics): Read usage analytics for tool executions.
- [Trace Analytics](https://docs.acruxcore.com/api-reference/traces/analytics): Read time-series analytics over every span ingested by the gateway or the SDK.
- [Trace Facets](https://docs.acruxcore.com/api-reference/traces/facets): Discover the distinct tags, metadata keys/values, resolved span models and declared error codes a team has actually used, for populating filter pickers.
- [Trace Settings](https://docs.acruxcore.com/api-reference/traces/settings): Configure team-wide trace capture behavior, such as whether payloads are stored.
- [Traces](https://docs.acruxcore.com/api-reference/traces): Ingest and query traces — the record of one request, its spans, tokens, and cost.
- [Traces Feedback](https://docs.acruxcore.com/api-reference/traces/feedback): Attach and read human feedback (rating, label, comment) on a trace or one of its spans.