Skip to main content

Changelog

What shipped, a week at a time. AcruxCore deploys continuously, so these are weekly summaries rather than numbered releases — the SDKs are the exception and follow semver.

Each major change gets its own heading and up to three bullets, one line each. Minor changes are a single line, no heading.

The two SDKs keep their own release notes, including migration steps for breaking changes: @acruxcoreai/sdk on npm and acruxcore on PyPI.

Beta

AcruxCore is in beta and pre-1.0. Breaking changes still happen; when one does, it is called out in the week it ships and in the SDK release notes.

Week of 7 September 2026​

Major​

The whole team's audit trail, in the dashboard​

  • Team → Audit trail lists every recorded action — keys, members, gateway, secrets, prompts, tools.
  • Filter by area, by a single event, and by who did it; the filters are in the URL, so a view is shareable.
  • Owner and admin only; a Bearer key cannot read it, because the trail records what people did. Reference →

One filter bar, on every screen that picks traffic​

  • Type tag:, prompt:, input:, meta.<key>: and the rest into a single box; each becomes a chip.
  • The same bar is on Traces, Feedback, and a dataset's Add example → From feedback.
  • Suggestions come from your team's own tags, metadata keys and prompt names. Reference →

Saved views​

  • Name a filter set and it is one click for the whole team, on Traces or Feedback.
  • Any member can rename or delete a view; the filters are stored exactly as you set them. Reference →

Add every row a filter selects, not just the page​

  • Add all N matching in From feedback builds from criteria instead of ticked boxes.
  • It reaches rows on pages you never opened, and says when more matched than one request takes. Reference →

Find a trace by what was actually said​

  • q now searches the captured request and response, not only names and attributes.
  • q_in narrows a search to input, output or name when a word appears on both sides.
  • Payload text is searchable wherever payload capture was on for the trace. Reference →

Filter by prompt, and filter the feedback feed​

  • prompt_id matches every version of a prompt, so you no longer need the version id.
  • The feedback feed takes the full trace filter set plus rating, source, label, has_comment.
  • One filter vocabulary now covers traces, feedback and dataset building. Reference →

Build a dataset from criteria instead of a hand-picked list​

  • Both from-feedback endpoints accept a filter in place of feedback_ids.
  • "Every thumbs-down with a comment on the checkout prompt" is one request.
  • The response reports matched, so a selection capped at 100 rows is visible. Reference →

Curate a dataset after you build it​

  • Add example now has a From feedback tab — pull real rows into a dataset that already exists.
  • A row brings its captured variables, its comment as criteria, and its session history.
  • A row already in the dataset is reported as skipped, never added twice. Reference →

Edit criteria, drop a row, delete a dataset​

  • The criteria cell edits in place, so a rubric inherited from a complaint can be reworded.
  • Every example row has a remove control, and a dataset can be deleted from either screen.
  • Both ask for confirmation first; past experiment runs keep their reports either way.

Choose the model and the instructions that optimize a prompt​

  • Optimizer model picks which model writes the rewrites, not just what they run on.
  • Optimizer prompt swaps the built-in instructions for your own, format still enforced.
  • An unregistered optimizer model is now rejected up front instead of failing mid-run. Reference →

Trace a LangChain or LangGraph agent, in Python or Node​

  • New tutorial builds a two-tool research agent in both languages, traced end to end over OTLP. Tutorial →
  • One instrumentor covers chain, tool and LLM spans; adding openai double-counts every call.
  • instrument: ['langchain'] ships in Node SDK 0.12.0; Python has had it since 0.11.0.

Evaluate a run your own app rendered​

  • Send variables beside prompt_version_id and they are stored on the span, not discarded.
  • A stored prompt with no placeholders can now produce dataset examples from real traffic.
  • Those messages are never re-rendered — you rendered them, so the values are lineage only. Reference →

Team API keys now appear in the audit trail​

  • Fixed — minting or revoking a team-scoped key wrote no audit event, so the trail missed it.
  • Both events now name the member who acted and mark the key as team-scoped. Reference →

Minor​

  • The team audit endpoint takes event (a comma-separated list) and actorId; both narrow total.
  • A new /teams/:id/audit/actors lists everyone in the trail, removed members included.
  • A member role change or removal now reports the affected member's address, not just their id.
  • The Team page previews the five newest recorded actions, with a link to the full trail.
  • A sixth feature page, Audit, covers what the trail records, the filters, and the roles.
  • New guide walks the trail from the Team page to a filtered URL you can paste into a ticket. Guide →
  • Core concepts names the split: a trace is traffic your app made, an audit event is a change.
  • The home page's round-trip now opens on the trace and closes on the audit trail.
  • The compare page's audit-log and RBAC rows now link the guides behind them.
  • Fixed — the compare page and the comparison post said five competitors; there are six.
  • The nine-platform comparison gains an audit-trail row, and says which two we never checked.
  • The docs home and the project README now name Audit among the building blocks.
  • The best-open-source-LLMOps page marks best-for, limitations and licence with a coloured icon.
  • The comparison matrix marks each verdict with a glyph and shades the AcruxCore column.
  • The FAQ's comparison answer now names the audit trail as the row no paid plan gates.
  • Fixed — the FAQ said the comparison matrix weighs nine criteria; it weighs ten.
  • The security page covers roles and the trail: what it records, and what it keeps out.
  • The roles guide says role changes are recorded, and which roles can read the trail.
  • Dataset examples show a Prompt column naming the prompt version the row was captured from.
  • A dataset row with no variables now shows the prompt's own last message instead of a dash.
  • Long variable values in a dataset row clip to one line each, with Show all to read them.
  • Feedback rows now name the team member who posted them, not just "Developer".
  • Fixed — feedback skipped when building a dataset blamed payload capture even when it was on.
  • Fixed — the real reason is now named: a call that sent raw messages has no variables to replay.
  • Row delete and edit controls use proper icons and stay visible, instead of a hover-only ✕.
  • The dataset page's Optimize button and dialog now say they optimize a prompt, not the dataset.
  • Fixed — deleting a dataset logged a 404 in the browser console on the way out.
  • Fixed — optimize used a fixed model name a team may never have registered.
  • The model picker scrolls past seven models and shows a selected count, instead of growing.
  • The models field now says one set of rewrites is tested on every model picked, not one each.
  • Fixed — a dialog taller than the window put its own buttons out of reach; it scrolls now.
  • Fixed — two appends of the same feedback row at once could file it into a dataset twice.
  • Feature pages now explain fallbacks, OTel ingestion, bindings and the optimizer.
  • Fixed — the evaluation page's curl sample dropped every id from its URLs.
  • Fixed — that sample passed version_ids: ["v7","v8"]; the API takes version UUIDs.
  • Tracing is no longer gateway-only on the page; the OpenTelemetry path is named.
  • The docs site now publishes /llms.txt, a summary index of every page for AI crawlers.
  • A new FAQ page answers how AcruxCore compares, and where it does not fit.
  • A new best open-source LLMOps platforms page compares seven.
  • Each entry says what that platform is best for, where it falls short, and when it was checked.
  • The main site now publishes /llms.txt too, and both sites name the AI crawlers they allow.
  • The home page gained an at-a-glance block: category, licence, deployment, SDKs, price, audit trail.
  • Both /llms.txt files and the home page's structured data now state what the audit trail records.
  • The FAQ answers which platforms log who changed what, and what an audit log costs.
  • A new FAQ answer separates a trace from an audit event: traffic sent versus a change made.
  • The category page's seventh question asks who changed what, and whether it needs a paid plan.
  • The API reference now documents the team-wide audit trail, for owners and admins. Reference →
  • The gateway page names the native providers and the compatible connections.
  • Each feature page carries a real product screenshot and a first action of its own.
  • The homepage names the step it was missing: score a fix before promoting it.
  • The homepage lists evaluation and spend guardrails among the reasons to switch.
  • Fixed — sitemap lastmod dates trailed one commit behind the content they describe.
  • Sitemap dates now track every file a page renders, not only its own component.
  • Fixed — the "no eligible rows" error now names each real reason and how many rows hit it.
  • A skipped feedback row now says what is missing in a few words, and never points at a setting.
  • A tag or model shown on a span is now a link that opens the trace list filtered to it.
  • Filtering a list resets the page and clears any selection, instead of leaving both stale.
  • Fixed — a span's input and output showed as one long escaped line; nested JSON is decoded now.
  • Fixed — "Expand" on a payload removed the height cap and pushed the page away; it scrolls now.
  • Fixed — long attribute and metadata values were cut off with nothing to click; they expand now.
  • Fixed — a run cell's output was shown as quoted, escaped text instead of what the model wrote.
  • The dataset examples table says what a row is in a plain line above the table.
  • Both tutorial scripts and a runnable Python notebook ship with the output of a real run.
  • The notebook demonstrates a failing tool: the agent answers confidently, and gets it wrong.
  • Fixed — the traces.ingest() link in both SDK references jumped nowhere.
  • New landing-page overview video: one prompt from versioned template to promoted fix.
  • Fixed — a week in this changelog showed a duplicated entry and a stray heading.
  • The compare page, the README and the nine-platform post gain a prompt-optimizer row.
  • Fixed — the comparison post said no competitor automates a prompt rewrite; two do.
  • Each feature page now names its capability in the title, the heading and the page summary.
  • The six feature pages link each other by capability, not by a one-word label.
  • Fixed — search previews cut every marketing page's description off halfway through.

Week of 31 August 2026​

Major​

Upstream rate limits answer 429 with the provider's own reason​

  • Provider 429s now answer 429 PROVIDER_RATE_LIMITED, not an opaque 502 PROVIDER_ERROR.
  • The provider's own reason and its Retry-After come through, so quota and pace differ. Reference →
  • Streaming requests answer the same way, as JSON, before any bytes are written.

Start a dataset from the dashboard, without waiting for feedback​

  • New dataset on Evaluations → Datasets creates an empty dataset — no traffic needed.
  • Add example writes one row by hand: named variable fields, or raw JSON for other types. Guide →

Evaluation runs and experiments can be deleted​

  • DELETE /api/v1/runs/:id removes a run and its cells; the Runs tab has a Delete action.
  • DELETE /api/v1/experiments/:id removes an experiment and every run under it. Reference →
  • Both answer 409 RUN_IN_FLIGHT while a run is still queued or running.

Report your own spans without waiting for the network​

  • traces.ingest(..., wait=false) buffers the trace and returns a usable trace id at once.
  • Measured at the call site: 6.5ms awaited, 0.017ms buffered, on a local API. Node · Python
  • New traces.flush() drains the buffer; both SDKs, both unchanged by default.

Minor​

  • Fixed — the feedback summary grouped by prompt version showed a raw id, not a name.
  • Feedback summary buckets now carry label and promptId alongside the unchanged key. Reference →
  • Sessions can be searched by id from the page itself, not only by editing the URL.
  • Fixed — the Evaluations → Runs tab was headed "Datasets" and described the wrong page.
  • Fixed — a custom provider base URL dropped every upstream response header.

Week of 17 August 2026​

Major​

Traces are named by what ran, not by a timestamp​

  • Traces sent over OTLP now take their root span's name instead of the time they started.
  • A trailing run id is trimmed, so two runs of the same crew share one searchable name.
  • A tool loop that joins an existing trace no longer renames it to runToolLoop. Guide →

Name a trace as a fallback, without overwriting a name it already has​

  • New x-trace-name-if-unset header (and body trace.nameIfUnset) on gateway completions.
  • It names a new trace, and is ignored on one that already has a real name.
  • x-trace-name is unchanged: still an instruction, still overwrites. Reference →

A chat() call with a trace option no longer counts itself twice​

  • Fixed — passing trace to gateway.chat() reported a second llm span per call.
  • Span counts, per-trace token totals and per-run call counts were all wrong, silently.
  • trace now also carries the name, trace id and session id through chat(), as on the loop. Reference →

A provider's 400 now names the rule you broke​

  • The gateway forwards the provider's own message for a 400, 404, 413 or 422.
  • A strict-mode response_format error names the exact property, not just "status 400".
  • 401, 403, 429 and every 5xx stay summarised — those describe the connection, not your request. Reference →

OpenAI's cached prompt tokens are now billed at the cached rate​

  • A repeated prompt prefix that OpenAI serves from cache is charged at half the input rate.
  • usage.cached_tokens is returned on the gateway response, as a subset of prompt_tokens.
  • Costs in traces, budgets and usage were overstated for any repeated system prompt.

Both SDKs released as 0.10.0​

  • npm i @acruxcoreai/sdk@0.10.0 and pip install -U acruxcore carry client_tools.
  • New prompts.list_aliases() / listAliases() reads which version an alias points at.
  • Python async with AcruxCore() now closes its HTTP connection pool on the way out. Guide →

Run a prompt's client tools without writing a dispatcher​

  • Both SDKs take client_tools / clientTools: tool name to the function that runs it.
  • Nothing is written to the catalog, so the tool's definition and its binding's pin stay put.
  • A missing implementation stops the run before the first model call, naming the tool. Guide →

Connect a tool to a prompt by choosing its alias​

  • Bind a catalog tool to a prompt in one write — no version to commit, no alias to promote.
  • Give one prompt alias its own tools: a different build of one, or none at all.
  • Breaking: POST /prompts/:id/versions no longer accepts tools — bind the tool instead. Guide →

See what a render will really call​

  • render() returns toolResolutions: the tool alias followed, its version, and the binding.
  • source reads alias when the prompt alias has its own binding, default when inherited. Reference →

See when a tool's code changed​

  • New GET /api/v1/tools/:id/audit trail: version commits, code-sync pushes, and binding changes. Reference →

Evaluation rules: real models, real filters, custom judge prompts​

  • Judge model is now a required dropdown — no more silent fallback to an unregistered model.
  • Match filter's model, prompt+alias, and tags fields are now dropdowns, not free text.
  • Use one of your own Prompts as the judge's grading template instead of the built-in one. Guide →

Run a prompt's tools in two lines​

  • gateway.runPromptWithTools(rendered) takes a render result and fills in the rest.
  • Model, messages, bound tools and prompt lineage all come from the render, not the call site.
  • A binding pinned to an exact tool version now travels as a pin, not as its alias. Guide →

A tool-using agent can stream its answer​

  • stream: true on the tool loop yields typed events: content, tool_call, tool_result, done.
  • Same trace as an unstreamed run — one llm span per round, tool spans nested under it.
  • Both SDKs, and on runPromptWithTools too. Guide →

Client-side tool loops keep their prompt lineage​

  • Fixed — SDK loops produced llm spans with no link back to the prompt version.
  • POST /gateway/chat/completions now accepts prompt_version_id next to your own messages.
  • Both SDKs send it automatically; a version from another team is rejected, not stamped. Reference →

Both SDKs released as 0.9.0​

  • npm i @acruxcoreai/sdk@0.9.0 and pip install -U acruxcore carry everything above.
  • Breaking: committing a version no longer takes tools — bind the tool to the prompt.
  • Breaking: the tool-routes endpoints are gone; the binding methods replace them. Guide →

Minor​

  • tool_refs and POST /tools/resolve now take version to pin one exact tool build.
  • Fixed — an undecryptable provider credential now returns a clear 409, not an opaque 500.
  • Fixed — group_by=day trace analytics bucketed days in the server timezone, not UTC.
  • Fixed — the changelog showed one week as two sections, dated three days apart.
  • GET /traces/facets now also returns distinct resolved models, for filter pickers.
  • Corrected — comparison posts now reflect AcruxCore's rule-based online evaluation. Reference →
  • New tutorial: a travel planner that picks between three tools, or calls none at all. Tutorial →
  • Nine screenshots across the tutorials and guides now show the current alias-keyed Tools tab.
  • Fixed — Python tabs in five tutorials called Node method names and Node argument shapes.
  • Fixed — the Python/Node tabs for scripting gateway setup called two methods no SDK has.
  • Fixed — the API reference's datasets.createFromFeedback example named a method that does not exist.
  • Fixed — the trace-tagging guide still used the flat get_trace()/getTrace() removed in SDK 0.7.0.
  • Four tutorials now ship a runnable setup_prompt.py for the prompt-and-tool setup step.
  • Fixed — a malformed JSON body now returns 400 INVALID_JSON, not 500, so clients stop retrying it.
  • Fixed — an oversized body returns 413 and an unsupported Content-Encoding returns 415.
  • The travel-planner tutorial now shows every setup step twice: the dashboard click-through and the code.
  • Tool parameters: a checkbox now writes additionalProperties: false, no raw JSON needed.
  • Fixed — "Back to builder" on a tool version no longer looks dead; it greys out with a reason.
  • New guide: where a tool's definition should live, and what changes in traces and deploys. Guide →
  • That guide also ships as a runnable notebook, with a preflight check for a brand new account. Guide →
  • The travel-planner tutorial now ships as a runnable notebook too, written for a first-timer. Tutorial →
  • The Python SDK tool-calling tutorial ships as a runnable notebook, with both setup routes shown. Tutorial →
  • Fixed — the New tool dialog implied its description always reaches the model.
  • Resolving a tool now says it has no committed version, instead of answering a bare 404. Reference →
  • SDK errors now carry the API's own message, so the reason is in the exception you catch.
  • Fixed — three tutorials no longer install langchain-community, which is being sunset. Tutorial →
  • The Tavily tutorials now use Tavily's own tavily-python SDK instead of a LangChain wrapper. Tutorial →
  • The no-SDK REST tutorial ships as a runnable notebook, including what unthreaded traces look like. Tutorial →
  • The ReAct agent tutorial ships as a runnable notebook, covering who records a span on the BYO path. Tutorial →
  • The configurable-agent tutorial ships as a runnable notebook, swapping model and persona by alias. Tutorial →
  • The medical-information tutorial ships as a runnable notebook, with a citation check on every answer. Tutorial →
  • The supervisor multi-agent tutorial now ships as a runnable notebook, routing traps included. Tutorial →
  • The BYO-provider RAG tutorial now ships as a runnable notebook, with a live provider check. Tutorial →
  • The CrewAI tracing tutorial now ships as a runnable notebook, with four OTLP wiring traps. Tutorial →
  • Fixed — the CrewAI tutorial's install command was missing the tavily-python client.
  • The OpenAI Agents SDK tutorial now ships as a runnable notebook, handoff span included. Tutorial →
  • Fixed — a CrewAI trace is now named Crew.kickoff, not once per run id, so names group. Tutorial →
  • Fixed — an SDK error from your own provider now names the reason, not just the status.
  • Fixed — the SDK now says a model is required, rather than letting your provider guess one.
  • Fixed — screenshots in every runnable notebook now load in GitHub, nbviewer and Jupyter.
  • Runnable notebooks no longer repeat the API-key dialog screenshot; the menu path is enough.
  • The tutorials now run on five models across OpenAI, Anthropic, Gemini, Llama and Mistral. Tutorial →
  • Fixed — the ReAct agent tutorial said an OpenRouter key would not work; any provider does. Tutorial →
  • Fixed — the travel-planner notebook crashed on a model registered without prices.

Week of 10 August 2026​

Major​

Score live traffic automatically with evaluation rules​

  • Create a standing rule that judges matching production calls with no run and no click.
  • Filter by prompt, model, or tag; sample and cap spend; get alerted below a threshold. Guide →

OTLP trace ingestion​

  • New POST /api/v1/traces/otlp accepts real OpenTelemetry exports over OTLP/HTTP
  • Works with CrewAI, LangChain, LlamaIndex via openinference-instrumentation-* — no code change
  • Reference →

SDK 0.8.0 (Python) — acruxcore.otel helper​

  • acruxcore.otel.register() wires the OTLP pipeline in one call, with instrument=[...] for CrewAI, LangChain, LlamaIndex, OpenAI, and the OpenAI Agents SDK
  • New optional extra: pip install 'acruxcore[otel]'
  • Guide →

SDK 0.8.0 (Node) — @acruxcoreai/sdk/otel helper​

  • New subpath export wires the OTLP pipeline in one call, with instrument: [...] for OpenAI and the OpenAI Agents SDK
  • @opentelemetry/* packages are new optional peer dependencies
  • Guide →

Minor​

  • Fixed — OTLP exports with input/output data could fail on retry instead of succeeding.
  • Fixed — LLM cost showed blank for provider-dated model ids (e.g. gpt-4o-mini-2024-07-18).
  • New tutorials: trace a CrewAI crew and an OpenAI Agents SDK app over OTLP.
  • Blog tag and author pages (/blog/tags, /blog/authors, and their archives) are now marked noindex so they stop competing with the posts they link to in search results.
  • Corrected — comparison posts now state the gateway's spend caps and rate limits correctly. Reference →
  • Compare — a row where a competitor lands in the same place as AcruxCore is now marked "Tie". Reference →
  • Compare — the Tool catalog row now records our edge over MLflow's MCP server registry.
  • Fixed — the comparison table cut off its last column on wide screens, at any zoom level.
  • Repo README now opens with a 35-second demo: prompt, version, tool, a real model call, the trace.

Week of 3 August 2026​

Major​

SDK trace analytics and sessions bindings​

  • New hub.traces namespace: analytics, facet discovery, and payload-capture settings.
  • New hub.sessions namespace lists sessions and reads one session's full trace history.
  • Feedback summary and the team-wide feedback feed are now reachable via hub.traces too.

SDK 0.6.7 — trace tags and metadata​

  • chat() and the tool loop now accept tags/metadata in trace options, sent as gateway headers.
  • Both SDK packages now link back to the public GitHub repo.

AcruxCore is now open source​

Past evaluation runs are now listed in one place​

  • A new Runs tab on Evaluations lists every run, newest first, with its score and best variant.
  • A run's report is reachable long after the fact — closing the tab no longer loses it.
  • GET /api/v1/runs returns the same history, filterable by status, dataset or prompt. Reference →

Evaluations and optimize now use full conversation context​

  • Feedback on a session now carries its prior turns into the dataset example.
  • A run replays that history before the new turn, so candidates see the real context.
  • The judge and the optimizer read it too, so scores and rewrites match the conversation.

Optimize and experiments can now pick their baseline alias​

  • New alias field — target staging, dev, or any alias instead of production.
  • Baseline still defaults to production, falling back to the latest version if none.
  • Feedback-built datasets now warn (never block) if examples came from a different prompt.

SDK prompt version lifecycle​

  • Both SDKs now manage prompts end to end: create, commit, list, diff, and promote versions.
  • Export and import move a version between teams or environments as one JSON document.
  • Look up every trace a specific prompt version produced, from either SDK.

SDK tool catalog lifecycle​

  • hub.tools/client.tools gained: create, list, get, update, delete, versions, promote, analytics.
  • Available in both TypeScript and Python SDKs — see the Tool Catalog guide.

Apache License 2.0​

  • Permissive, OSI-approved, with nothing gated.
  • Fork it, self-host it, or sell what you build; the AcruxCore name and logo stay trademarked.
  • @acruxcoreai/sdk and acruxcore ship under the MIT license.

SDK 0.7.0 — Resource-based namespace pattern​

  • Breaking: All flat client methods removed. Use hub.gateway.chat(), hub.prompts.render(), hub.traces.ingest() etc. instead of hub.chat(), hub.renderPrompt(), hub.trace().
  • hub.gateway.stream() is now a standalone method (previously hub.chat({stream: true})).
  • hub.gateway.flush() / hub.gateway.close() replace hub.flush() / hub.close().

Evaluations are now scriptable from both SDKs​

  • hub.datasets, hub.experiments, hub.runs, and hub.optimize expose 19 methods for the full evaluations domain.
  • Create datasets, run experiments, poll results, read reports, and promote optimizer candidates without leaving your code.

One-command local self-host​

  • New docker-compose.local.yml bundles Postgres, Redis, API, worker and web in one file.
  • docker compose -f docker-compose.local.yml up --build — no .env to fill in first.

Minor​

  • Fixed — the trace settings API reference showed the wrong payload-capture default.
  • Fixed — six API reference pages showed a stale 401 error message.
  • POST /datasets/from-feedback now accepts at most 100 feedback ids per request.
  • Fixed — the Python SDK tool-calling tutorial's decorator example had a broken import.
  • Tutorial and guide pages now show a short snippet plus a link to the full runnable script.
  • Tutorial script links now point to scripts that were actually run and verified.
  • New guide — evaluate a prompt with conversation history.
  • New guide — view trace analytics.
  • New guide — configure trace payload capture.
  • New guide — look up the traces a prompt version produced.
  • A LICENSE, TRADEMARK.md, and CLA.md now ship at the repo root.
  • Fixed — the /sdk page's Python tutorial links 404'd (wrong docs path).
  • The /sdk page now links five capability guides per language, and its code samples show a decorated tool call and session tracing.
  • Three SDK guides — chat, tracing, and gateway routing — now show Python code alongside Node's.
  • Added — an "Optimize" button on a dataset's page starts a run without rebuilding it.
  • Added — GET /api/v1/health reports database and Redis reachability for load balancers and uptime monitors.
  • Terms, Privacy and the site footer now name AcruxCore without a corporate suffix.
  • Fixed — routing requests to OpenAI reasoning models (o1, o3, o4-mini, gpt-5) no longer 400s with Unsupported parameter: 'max_tokens'; the gateway now sends max_completion_tokens to OpenAI, while openai_compatible providers keep max_tokens.
  • Fixed — the model-page "Test" button works for reasoning models (o1, o3, o4-mini, gpt-5); the connectivity ping no longer sends a 1-token cap those models can't meet.
  • Privacy names AcruxCore as the controller for the hosted service, with a contact address.
  • New guide — Product tour: tools, streaming, traces, feedback, and evaluation in one walkthrough.
  • Fixed — rendering a prompt with a {% for %} loop no longer wrongly demands the loop variable as an input.
  • /compare pages now show AcruxCore's real one-command Docker self-host.
  • /compare's Self-hosting row now reads "docker compose up" for every column, matching the identical command.
  • The product name is now one word — "AcruxCore" across the site, docs, and both SDKs.
  • SDK 0.7.1 — the rename reaches both packages' metadata; no API or behaviour change.
  • The repo README now carries the competitor comparison table, losses and ties included.
  • The SDKs page now links the full TypeScript and Python API reference.
  • Fixed — marketing pages answered on two addresses; /pricing/ now redirects to /pricing.
  • Fixed — marketing pages showed the flat client methods SDK 0.7.0 removed.
  • 11 blog post titles and 12 descriptions were shortened so search engines stop truncating them.

Week of 27 July 2026​

Major​

Tracing no longer slows your model calls down​

  • Spans queue in the background — the model's answer no longer waits on a trace write.
  • About 570 ms off every traced call, ten times what the gateway's own routing costs.
  • Reading traces right after a call needs await hub.flush() first — both SDKs at 0.6.5.

Tools are now defined once, in code​

  • A decorated function is the tool — no create → commit → promote for a code-owned tool.
  • POST /tools/sync commits only when the spec really changed, and moves the alias. Reference →
  • Both SDKs at 0.5.0 — breaking: raw tool definitions move to toolDefs / tool_defs=.

Streamed gateway completions are traced​

  • A streamed call was billed but wrote no llm span, so it was invisible in traces.
  • Tool spans under a streamed call were orphaned, and now nest where they belong.
  • Streaming records the same span a non-streamed call does.

Queued evaluation runs and outbound email could stall forever​

  • The API and the worker could each reach a different Redis, so neither read the other's work.
  • Runs sat at queued, and invites, verification mail and digests were never sent.
  • Fixed — restart any run of yours still sitting at queued.

The judge scored correct answers as failures on feedback-derived criteria​

  • A feedback comment describes the reply that provoked it, not the answer you want.
  • The judge read it as describing the output, so a correct rewrite could still score 0.
  • Re-run any optimize or experiment run you judged against feedback criteria.

Optimize runs could fail while the rewrites were fine​

  • The optimizer's example showed a lone system message, so candidates dropped the variables.
  • One bad escape in the model's JSON threw the response away; near-valid JSON is now repaired.
  • A rejected candidate is now named with its reason instead of a bare failure.

A team member now holds exactly one role — a breaking API change​

  • Invites and role updates take {"role": "editor"}; the array form now returns 400.
  • GET /auth/me, /auth/teams, /teams/:id/members and /invites return a role string.
  • Nobody lost access — anyone who held more than one role keeps their highest. Invite a teammate →

You can now use the SDK without our AI gateway​

  • Bring your own OpenAI-compatible key and base URL; the key never touches our servers.
  • Tracing still works — the SDK reports its own spans, streaming and non-streaming. SDK guide →
  • Both SDKs at 0.6.0, no breaking changes; an HTTP 429 now retries like a 5xx.

The SDK render cache could send the wrong prompt to the model​

  • The 60-second cache key left out your variables, so a new question got the first render.
  • Variables are now part of the key, and their order does not split an entry.
  • cacheTtl / cache_ttl of 0 really disables the cache instead of always serving stale.

response_format support on the gateway​

  • Pass structured-output requests straight through for OpenAI and Gemini models.
  • Anthropic models get the same contract via an internal forced-tool-call translation — no caller-visible difference.
  • Both SDKs at 0.6.6 accept responseFormat/response_format directly, like tools/toolChoice.
  • A cookie banner now shows once, site-wide; analytics cookies are set only if you accept.
  • One consent choice covers both acruxcore.com and docs.acruxcore.com — no re-prompt.
  • Change your choice any time from "Cookie preferences" in the footer.

SDKs now forward trace tags and metadata to the gateway​

  • chat() and runToolLoop() / run_tool_loop() pass tags and metadata as gateway headers automatically.
  • runToolLoop also forwards them as x-span-tags / x-span-metadata, tagging each gateway LLM span. New guide →

Minor​

  • A "Beta" badge in the landing hero, matching the one the signed-in app already showed.
  • A product demo plays on the home page.
  • The home page's code panel switches between TypeScript and Python, with a tab to pin one.
  • New writing — a hands-on comparison of LangSmith, Langfuse, PromptLayer and AcruxCore.
  • Re-measured — how much overhead an LLM gateway adds now benchmarks five paths including bring-your-own-key, on one fresh run.
  • A real logo — a crescent and compass rose, across the site, docs, browser tab and previews.
  • Tool release notes are separate from the model-facing description.
  • Code-owned tool warning in the dashboard before an edit the next deploy will supersede.
  • Tool loops route by executor type — http runs on the platform, client runs locally.
  • Rendering a prompt now returns its version id and number, linking a trace to that version.
  • chat() can thread manual calls into one trace via trace: { traceId, sessionId }.
  • A bring-your-own provider URL is checked for HTTPS — warned once if it is plain http://.
  • Hardening — the gateway's token estimate is bounded, so one odd prompt cannot slow others.
  • New guide — build a RAG agent without the gateway.
  • New guide — improve a prompt from feedback.
  • The streaming-trace post now carries the production run that verified the fix.
  • The Quickstart's Python tab now uses the acruxcore SDK instead of raw requests.
  • Core concepts was rewritten around the new tool path.
  • Fixed — a signed-in visitor clicking the public site's logo or nav was bounced into the app.
  • Fixed — the sign-in page and the signed-in app's sidebar still showed the old placeholder mark.
  • Fixed — list bullets stopped rendering on the site's written pages.
  • Fixed — a run report's delta badge printed +66.66666666666667 beside a score reading 66.7.
  • Fixed — errors from local development were reported to the monitor alongside production ones.
  • Fixed — several site code samples still showed the pre-0.5.0 way of passing a prompt's tools.
  • Fixed — the TypeScript SDK's npm package page linked a private source repo that 404s for visitors.
  • Fixed — simultaneous tool syncs could create two tools with one name, or fail with a 500.
  • Fixed — a first sync targeting a custom alias reported success without creating it.
  • Fixed — a duplicate tool name now returns 409 TOOL_NAME_TAKEN, not a hidden second tool.
  • Fixed — accepting the cookie banner on the site did not actually start analytics.
  • New Tutorials section — end-to-end agent builds split out from single-feature guides.
  • New guide — manage team roles and permissions.
  • New guide — scope access with virtual keys.
  • New guide — alias and track usage of tools in the catalog.
  • New guide — automatic model fallbacks.
  • New guide — set spend limits with gateway budgets and rate limits.
  • New guide — diff, export, and import your prompt library.
  • Evaluate a prompt now frames the baseline comparison as a pre-ship gate.
  • New writing — how much latency and spend exact-match gateway caching actually saves.
  • New writing — comparing Anthropic, OpenAI, and Gemini request shapes behind one gateway API.
  • Fixed — a malformed prompt id in the URL returned a raw 500 instead of a clear 400.
  • Fixed — deleting a tool could leave a secret it referenced permanently undeletable.
  • New guide — tag and filter traces.
  • New tutorial — build a ReAct agent with a real Yahoo Finance news tool.
  • New tutorial — build a configurable ReAct agent with real Tavily search.
  • New tutorial — build a supervisor multi-agent system with real finance/research/writing subagents.
  • New tutorial — build a Medical-Information QA agent demonstrating response_format.

Week of 20 July 2026​

Major​

Self-hostable authentication​

  • Auth moved from Supabase to Better Auth, so no hosted identity provider is needed.
  • Existing sessions and sign-in flows are unchanged.

API keys are hashed at rest and shown once​

  • A key is displayed a single time when created, then stored only as a SHA-256 hash.
  • Keys created before this change were removed — they could not be migrated.

Transactional email, event notifications and a weekly usage digest​

  • Invites, email verification and password resets now send real mail.
  • Each team can opt in to event notifications and a weekly usage summary.
  • Every message carries a one-click unsubscribe.

Python SDK​

  • acruxcore on PyPI — async, and at parity with the TypeScript SDK.
  • Prompt render, gateway chat and streaming, tool loops, traces and feedback.
  • Shipped with a text-to-SQL agent guide.

Minor​

  • Prompt default model — a render no longer repeats the model the prompt was written for.
  • Error monitoring across the API, the worker and the dashboard.
  • A security hardening pass across the gateway, prompt rendering, membership and traces.
  • The public site and docs site were rebuilt for launch, including SEO and footer pages.
  • Fixed — a worker start-up race could silently drop the email, eval-run and digest workers.
  • Fixed — clicking the in-app logo now opens the landing page instead of doing nothing.

Week of 13 July 2026​

Major​

One trace per agent run​

  • A client-side tool loop threads a trace id, so a multi-step run is a single trace.
  • The gateway's llm spans and your tool spans appear in one tree.
  • Previously every model call produced a trace of its own.

Tool calls run in parallel​

  • When a model asks for several tools at once, the loop dispatches them concurrently.
  • Previously they ran one after another.

The SDK gained the core LLM methods​

  • Chat, streaming and feedback, alongside prompt rendering.
  • An agent no longer needs a provider client of its own.

Minor​

This changelog starts on 13 July 2026. Anything before that predates the public beta.