Repository navigation
Releases
v2026.11.9
Docker images
docker pull acruxcore/api:2026.11.9
docker pull acruxcore/worker:2026.11.9
docker pull acruxcore/web:2026.11.9With the bundled docker-compose.yml, pin this release with ACRUXCORE_IMAGE_TAG=2026.11.9.
The same images are also tagged sha-8c23d90.
SDKs
| Package | Registry | Version | In this release |
|---|---|---|---|
@acruxcoreai/sdk |
npm | 0.22.0 |
Unchanged |
acruxcore |
PyPI | 0.22.0 |
Unchanged |
@acruxcoreai/mcp-server |
npm | 0.3.1 |
Unchanged |
What's changed
Fixes and smaller changes
- Fixed — OTLP ingest stored about 720 spans a second; it now stores about 1,700.
- Fixed — The trace list slowed as history grew. Its count stops 10,000 rows past the page.
- Fixed — Trace ingest queued a scoring job per LLM call even for teams with no eval rules.
- New post — Multi-agent systems: when splitting work across LLM agents helps, and when it fails.
Pull requests in this release (3)
- fix(traces): 2.4x faster OTLP ingest, flat trace list at 100k traces
- fix(online-eval): skip the job for a team with no eval rules
- fix(web): allow Cloudflare's Web Analytics beacon in the CSP
The weekly changelog describes every change, week by week.
To run AcruxCore yourself, see Self-hosting.
v2026.11.8
Docker images
docker pull acruxcore/api:2026.11.8
docker pull acruxcore/worker:2026.11.8
docker pull acruxcore/web:2026.11.8With the bundled docker-compose.yml, pin this release with ACRUXCORE_IMAGE_TAG=2026.11.8.
The same images are also tagged sha-4c5bfb0.
SDKs
| Package | Registry | Version | In this release |
|---|---|---|---|
@acruxcoreai/sdk |
npm | 0.22.0 |
Unchanged |
acruxcore |
PyPI | 0.22.0 |
Unchanged |
@acruxcoreai/mcp-server |
npm | 0.3.1 |
Unchanged |
Upgrade notes
- Database migrations (2). The
apicontainer applies them on start. Back up Postgres before you upgrade.20261111120000_trace_client_trace_id20261111120100_trace_client_trace_id_index
- New optional variable
EVAL_CALL_DEADLINE_MS. - New optional variable
EVAL_PER_TEAM_MAX_IN_FLIGHT. - New optional variable
ONLINE_EVAL_WORKER_CONCURRENCY.
What's changed
New
Security fixes from an internal audit
- Fixed — the Preview tab could run script from a teammate's template; it renders on the server now.
- The web app sends a Content-Security-Policy that blocks inline and eval'd scripts.
- No team could read another team's data; the fixes below close the remaining gaps.
Only owners and admins can bind team secrets into tools
- Sending
{{secret.NAME}}to a new URL, header or query parameter now needs owner or admin. - Editors can still run those tools and save versions that keep the URL and secret entries.
Budgets cap calls that send no output limit near the limit
- A call with no
max_tokensis capped only when the budget left can't pay for a long answer. - The cap is what the budget left buys at the dearest fallback's price, at most 4096 tokens.
- A cached answer still serves a nearly spent budget, and parallel calls can't overshoot it.
Fairer evaluation workers
- One team's large or slow eval runs no longer delay other teams' runs and online scoring.
- Eval workers run 4 jobs per queue, at most 2 per team per worker (
EVAL_PER_TEAM_MAX_IN_FLIGHT). - A timed-out eval call is not retried; a provider's quick 504 is. See
EVAL_CALL_DEADLINE_MS.
Sign-in rate limit keys on a verified client address
- Fixed — a forged
CF-Connecting-IPheader no longer gets around the sign-in rate limit. - Self-hosting behind Cloudflare with your own proxy? Set
AUTH_IP_HEADERS=cf-connecting-ip,x-forwarded-for.
Fixes and smaller changes
- Fixed — removing or demoting an admin now revokes the pending invite links they created.
- Fixed — local installs without
BETTER_AUTH_SECRETsigned tokens with a publicly known secret. - Fixed — a locally built Docker image could include the generated dev auth secret.
- Fixed — a call running when a budget period reset could leave extra headroom in the new period.
- Fixed — changing a budget's period during a call left that call's estimate counted as spent.
- Fixed — re-sent trace spans could trigger evaluation-rule judge calls past the daily limit.
- Fixed — a trace id already used by another team made your trace ingest fail or split.
- Fixed — heavy trace searches by one team could make other teams' requests fail.
- Trace search terms are capped at 256 characters, and a search that runs too long returns 422.
- Fixed — diffing two very different prompt versions slowed the API; it now says "too large".
- Fixed — a streamed answer that never sent a line break could tie up the API.
- Fixed — a short template or tool transform could produce an enormous result; it now errors.
- A team renders at most two Preview tab drafts at once; the tab retries when it is busy.
- An invalid
EVAL_CALL_DEADLINE_MSorEVAL_PER_TEAM_MAX_IN_FLIGHTis now logged, then ignored.
Pull requests in this release (1)
- fix: security audit findings (14 confirmed)
The weekly changelog describes every change, week by week.
To run AcruxCore yourself, see Self-hosting.
v2026.11.7
Docker images
docker pull acruxcore/api:2026.11.7
docker pull acruxcore/worker:2026.11.7
docker pull acruxcore/web:2026.11.7With the bundled docker-compose.yml, pin this release with ACRUXCORE_IMAGE_TAG=2026.11.7.
The same images are also tagged sha-ab65793.
SDKs
| Package | Registry | Version | In this release |
|---|---|---|---|
@acruxcoreai/sdk |
npm | 0.22.0 |
Unchanged |
acruxcore |
PyPI | 0.22.0 |
Unchanged |
@acruxcoreai/mcp-server |
npm | 0.3.1 |
Unchanged |
Upgrade notes
- Database migrations (1). The
apicontainer applies them on start. Back up Postgres before you upgrade.20261113090000_provider_kind_runware
What's changed
New
Runware is a gateway provider
- Add a Runware key as a credential, with no base URL to set. Setup →
- The model picker lists every Runware chat model, 64 today. Streaming works.
- Tool calls work on every model family except Gemini, whose second turn fails at Runware.
Fixes and smaller changes
- New model fills in a model's prices from the Runware or OpenRouter model list.
- New guide — Supported providers: what each gateway provider supports.
Pull requests in this release (2)
- docs(seo): internal links to the alternatives posts
- feat(gateway): Runware provider (BYOK, all chat models, listed prices)
The weekly changelog describes every change, week by week.
To run AcruxCore yourself, see Self-hosting.
v2026.11.6
Docker images
docker pull acruxcore/api:2026.11.6
docker pull acruxcore/worker:2026.11.6
docker pull acruxcore/web:2026.11.6With the bundled docker-compose.yml, pin this release with ACRUXCORE_IMAGE_TAG=2026.11.6.
The same images are also tagged sha-751f0f4.
SDKs
| Package | Registry | Version | In this release |
|---|---|---|---|
@acruxcoreai/sdk |
npm | 0.22.0 |
Updated from 0.21.0 — release notes |
acruxcore |
PyPI | 0.22.0 |
Updated from 0.21.0 — release notes |
@acruxcoreai/mcp-server |
npm | 0.3.1 |
Unchanged |
Upgrade notes
docker-compose.ymlchanged. Copy the new file before you pull the images.
What's changed
New
Direct SDK calls to Anthropic and Gemini use their own APIs
- Provider
kindnow takesanthropicandgemini. The base URL defaults to their own host. - Tool calls, streaming, structured output and token counts work as they do on OpenAI.
- Ships in SDK 0.22.0. Call your provider directly →
Get started has a path for calling your provider directly
- Call your provider directly takes two steps: an API key, then your first trace.
- The snippet covers OpenAI, Anthropic, Gemini and OpenRouter, in Python and Node.
- Pages that need a stored key, such as the Playground, say so and show a lock in the sidebar.
New accounts land on a getting-started checklist
- After sign-in, a team with setup left opens Get started, not the empty prompt list.
- Four steps, each marked done from your team's data: provider, model, API key, first request.
- A key created there is filled into a curl, Python or Node snippet you can paste and run.
A model with blank prices costs nothing on traces too
- Fixed — SDK and OTLP spans for a model saved with blank prices took a built-in price.
- The same model through the gateway already cost nothing, so cost charts mixed the two.
- Your model's prices decide on both paths. Built-in prices cover only models you never added.
SDKs 0.22.0 on npm and PyPI
- Both SDKs: direct Anthropic and Gemini calls, described above. Both have breaking changes.
- Tool loops take
extra_body/extraBody, so a direct Anthropic or Gemini loop can think. - Migration steps: TypeScript, Python.
Fixes and smaller changes
- New guide — Call your provider directly with your own key, with Python and Node scripts.
- Get started has an optional last step that gives the model a tool, on both paths.
- SDK tool loops take
extra_body/extraBody, so a direct Anthropic or Gemini loop can think. - Fixed — a Node SDK tool loop with
responseFormathid that it stopped atmaxIterations. - Fixed — "Go to app" on the home page skipped Get started for a team still setting up.
- Fixed — a Gemini 3 tool call through the gateway failed on the next turn with a 400.
- Fixed — parallel Gemini tool calls streamed through the gateway merged into one broken call.
- Fixed — Gemini through the gateway rejected tool and output schemas with
additionalProperties. - The New model dialog lists the models your key can call, and you can search the list.
- The dashboard opens in light mode. The theme toggle still switches to dark and remembers it.
- Get started ends with links to guides for tools, feedback, the optimizer, judges and datasets.
- A new model's blank prices now stay blank, not filled in from a built-in list.
- Self-hosting — Redis in
docker-compose.ymlkeeps its queue in a named volume across restarts. - New guide — Find the slow step in your app with
traces.step. - The SDK reference no longer says shipped options such as
reasoningEffortcome in a later release. - Fixed — after a streamed OpenAI call, the gateway opened a new connection for the next call.
Pull requests in this release (9)
- feat(web): land new users on a getting-started checklist
- fix(deploy): prune old images on deploy and give Redis a volume
- BYO as a first-class path: native Anthropic and Gemini in both SDKs, Get started choice, lock panels, guide
- docs: "Find the slow step in your app" guide + drop stale SDK version notes
- feat(sdk): extra_body / extraBody on the tool loops
- docs: trim and correct CLAUDE.md and AGENTS.md
- fix: blank-price models stay unpriced on traces; streamed calls keep their connection
- ci: run the old-blog-host check in the release-PR job
- chore(sdk): release both SDKs as 0.22.0
The weekly changelog describes every change, week by week.
To run AcruxCore yourself, see Self-hosting.
v2026.11.5
Docker images
docker pull acruxcore/api:2026.11.5
docker pull acruxcore/worker:2026.11.5
docker pull acruxcore/web:2026.11.5With the bundled docker-compose.yml, pin this release with ACRUXCORE_IMAGE_TAG=2026.11.5.
The same images are also tagged sha-cef100f.
SDKs
| Package | Registry | Version | In this release |
|---|---|---|---|
@acruxcoreai/sdk |
npm | 0.21.0 |
Unchanged |
acruxcore |
PyPI | 0.21.0 |
Unchanged |
@acruxcoreai/mcp-server |
npm | 0.3.1 |
Unchanged |
What's changed
Fixes and smaller changes
- Comparison posts and pages now cover Langfuse's AI Gateway, in public beta since 4 November.
- New post: Langfuse AI Gateway vs LiteLLM vs AcruxCore, all self-hosted.
Pull requests in this release (1)
- docs(blog): Langfuse AI Gateway vs LiteLLM vs AcruxCore, self-hosted benchmark
The weekly changelog describes every change, week by week.
To run AcruxCore yourself, see Self-hosting.
v2026.11.4
Docker images
docker pull acruxcore/api:2026.11.4
docker pull acruxcore/worker:2026.11.4
docker pull acruxcore/web:2026.11.4With the bundled docker-compose.yml, pin this release with ACRUXCORE_IMAGE_TAG=2026.11.4.
The same images are also tagged sha-789e010.
SDKs
| Package | Registry | Version | In this release |
|---|---|---|---|
@acruxcoreai/sdk |
npm | 0.21.0 |
Updated from 0.20.0 — release notes |
acruxcore |
PyPI | 0.21.0 |
Updated from 0.20.0 — release notes |
@acruxcoreai/mcp-server |
npm | 0.3.1 |
Unchanged |
Upgrade notes
docker-compose.ymlchanged. Copy the new file before you pull the images.
What's changed
New
SDK direct calls to OpenAI send the token limit gpt-5 expects
- Fixed — BYO calls to OpenAI with a token limit failed on gpt-5 and the o-series.
- New provider
kindfield for Azure OpenAI and proxies: Node, Python. - A BYO
extraBodycan no longer add a second token limit. All three ship in 0.21.0.
SDKs 0.21.0 on npm and PyPI
- Both SDKs: BYO token-limit fix and provider
kind, described above. Both have breaking changes. - TypeScript: after
register(), OTel-instrumented calls in a step join the step's trace. - Migration steps: TypeScript, Python.
Fixes and smaller changes
- GitHub releases now list each release's changelog entries, not PR titles.
- New guide — Migrate your prompts from Langfuse with
acruxcore-migrate. - New guide — Self-host AcruxCore with Docker, from the Docker Hub images.
- The local Compose file reads
WEB_PORTandAPI_PORTfrom.env, for when 8080 or 3001 is taken. - Fixed — the gateway ignored
max_completion_tokens, so the call ran with no output limit. - Fixed — a non-streamed call to OpenAI, Anthropic or Gemini after a 4 s pause reconnected.
Pull requests in this release (6)
- docs: migrate your prompts from Langfuse (guide + acruxcore-migrate), README category line
- fix(sdk): send max_completion_tokens on BYO calls to api.openai.com
- fix(gateway): accept max_completion_tokens as an alias of max_tokens
- fix(gateway): keep idle provider connections open for 60 s
- docs: self-host AcruxCore from the Docker Hub images
- chore(sdk): release both SDKs as 0.21.0; release PR lists only merged PRs
The weekly changelog describes every change, week by week.
To run AcruxCore yourself, see Self-hosting.
v2026.11.3
Docker images
docker pull acruxcore/api:2026.11.3
docker pull acruxcore/worker:2026.11.3
docker pull acruxcore/web:2026.11.3With the bundled docker-compose.yml, pin this release with ACRUXCORE_IMAGE_TAG=2026.11.3.
The same images are also tagged sha-8132794.
SDKs
| Package | Registry | Version | In this release |
|---|---|---|---|
@acruxcoreai/sdk |
npm | 0.20.0 |
Unchanged |
acruxcore |
PyPI | 0.20.0 |
Unchanged |
@acruxcoreai/mcp-server |
npm | 0.3.1 |
Unchanged |
What's changed
- docs: cover Langfuse's AI Gateway in comparisons
The weekly changelog describes each change for users.
To run AcruxCore yourself, see Self-hosting.
v2026.11.2
Docker images
docker pull acruxcore/api:2026.11.2
docker pull acruxcore/worker:2026.11.2
docker pull acruxcore/web:2026.11.2With the bundled docker-compose.yml, pin this release with ACRUXCORE_IMAGE_TAG=2026.11.2.
The same images are also tagged sha-bf767e0.
SDKs
| Package | Registry | Version | In this release |
|---|---|---|---|
@acruxcoreai/sdk |
npm | 0.20.0 |
Unchanged |
acruxcore |
PyPI | 0.20.0 |
Unchanged |
@acruxcoreai/mcp-server |
npm | 0.3.1 |
Unchanged |
Upgrade notes
- Database migrations (3). The
apicontainer applies them on start. Back up Postgres before you upgrade.20261107090000_span_payload_tools20261107091000_trace_name_is_derived20261107100000_trace_tool_sets
What's changed
- docs(blog): LLM-as-a-judge — how reliable LLM judges are and how to evaluate an LLM app
- fix: gateway trace issues #789–#793 (tool definitions, trace names, SDK OTel steps, provider headers)
The weekly changelog describes each change for users.
To run AcruxCore yourself, see Self-hosting.
v2026.11.1
Docker images
docker pull acruxcore/api:2026.11.1
docker pull acruxcore/worker:2026.11.1
docker pull acruxcore/web:2026.11.1With the bundled docker-compose.yml, pin this release with ACRUXCORE_IMAGE_TAG=2026.11.1.
The same images are also tagged sha-48bc45d.
SDKs
| Package | Registry | Version | In this release |
|---|---|---|---|
@acruxcoreai/sdk |
npm | 0.20.0 |
Unchanged |
acruxcore |
PyPI | 0.20.0 |
Unchanged |
@acruxcoreai/mcp-server |
npm | 0.3.1 |
Unchanged |
What's changed
- First tagged release. Earlier history is in the commit log.
The weekly changelog describes each change for users.
To run AcruxCore yourself, see Self-hosting.