MLflow Release Notes

MLflow 3.16.0 includes several major features and improvements

Major New Features

🎨 Custom Trace Views: Design your own trace UI in plain English — describe the view you want and the MLflow Assistant builds it for you, laying out exactly the fields you care about with no config files or custom code. Save, name, and reuse views per experiment so your whole team shares the same lens on your traces.

🔭 Redesigned Trace Experience (now default): The trace explorer has been rebuilt from the ground up and is now the default — tighter row density, cleaner navigation, and a redesigned span-tree view. Reorder columns, add configurable custom columns from any trace tag or metadata field, and find multi-turn conversations inline via session grouping.

🔗 Span Links: Spans rarely stand alone. MLflow 3.16.0 adds first-class span links so relationships between spans — a retrieval step, a tool call, a downstream trace — are captured and navigable. Record links in your SDK and the trace explorer surfaces them in a dedicated Links tab to jump straight to the destination span.

Breaking changes

• [Server-infra / Tracking] basic-auth: enable fail-closed authorization by default (#25308, @PattaraS)
• [Build / Models] Drop cross version testing for pyspark < 3.4.4 (#25098, @harupy)
• [Gateway / Tracing] fix: Serve FastAPI-native routers under --static-prefix (#24511, @SeiichiroYoshioka)

Features

• [Evaluation / Tracing / Tracking] Add mlflow skills agent hints for GenAI anti-patterns (#25496, @james-fletcher-db)
• [Tracing] Add shell-based MLflow Tracing setup wizard (#25441, @B-Step62)
• [Gateway] Support per-user budget policies in the AI Gateway (USER scope + principal) (#24371, @joelrobin18)
• [Tracing / UI] Add trace-table column reordering to V4 traces tab (#25428, @joshuawong-db)
• [Tracing / UI] Add Expectation field to the traces table filter UI (#25460, @creater29)
• [Tracing / UI] Add saved views to MLflow demo (#25392, @joshuawong-db)
• [Tracing / UI] Add configurable custom trace columns (tags + metadata) to Traces V4 (#25427, @joshuawong-db)
• [Tracing / Tracking / UI] Refresh the Traces V4 table: row density, header, and navigation (#25354, @joshuawong-db)
• [Tracing / UI] Mount Traces V4 as the default traces experience (#25424, @joshuawong-db)
• [Tracing / UI] Port redesigned trace explorer (v2) to OSS (#25355, @joshuawong-db)
• [UI] Deprecate the experiment Sessions tab in favor of Traces session grouping (#25350, @joshuawong-db)
• [Evaluation / Tracing] feat(typescript): add logFeedback and typed trace assessments (#25169, @rajudandigam)
• [Tracing] Point coding agents at the MLflow tracing skill on import mlflow (#25247, @harupy)
• [Evaluation / Gateway] Support UC model services in built-in judges (#25246, @TomeHirata)
• [Tracing / UI] Enable shouldEnableModelTraceExplorerCustomTraceView in OSS (#25245, @aaronteo-db)
• [Tracking] jobs: add configurable Huey storage URL for Redis (#22653, @jeevan6996)
• [Tracking] [chart] Add optional External Secrets Operator support (#24523, @habibaikram1)
• [Evaluation] Enrich mlflow scorers list --builtin with the full scorer catalog (#25003, @joshuawong-db)
• [Tracing / UI] feat(ui): display scorer description in tooltip for custom judges (#22043, @adamgurary)
• [UI] Add Image Grid toggle for run parameters (#23705, @jdoughty04)
• [Tracing / Tracking / UI] [Custom View] Enforce saved view limit per experiment (#25064, @tanghaoji)
• [Tracking] [1/6] basic-auth: fail-closed authorization net + coverage guard (#25065, @PattaraS)
• [Build / Scoring] Bump model serving image to Ubuntu 24.04 and Python 3.11 (#25089, @harupy)
• [] Add SpanLink support to TypeScript SDK (#23480, @robinnarsinghranabhat)
• [Docs / Tracking] [Custom View] Increase experiment tag value limit to 20000 (#25062, @aaronteo-db)
• [Tracing / Tracking] feat(ts-sdk): add MLFLOW_TRACKING_AUTH=kubernetes/kubernetes-namespaced support (#24243, @robinnarsinghranabhat)
• [Tracing / UI] [Custom View] Add custom view deletion (#25057, @tanghaoji)
• [Tracing / UI] [Custom View] Support local Assistant providers with structured responses (#25034, @tanghaoji)
• [Tracing / UI] [Custom View] Submit custom view build prompts directly to Assistant (#25026, @aaronteo-db)
• [] Support optimizing chat-type prompts in optimize_prompts (#23488, @gaurav0107)
• [Docs / Tracking] Add --artifact-policy retarget to mlflow db move-resources for repointing experiment artifact roots (#24635, @gauthierpiarrette)
• [] [5 / N] Add custom view tab UI, save/rename, and experiment tag persistence (#24798, @aaronteo-db)
• [Evaluation] Expose generate_rationale_first through make_judge (#25002, @aaronteo-db)
• [Tracing / UI] [4/N] Add client-tool pause/resume for render_custom_view in Assistant (#24796, @aaronteo-db)
• [Tracing] Enable OpenCode tracing plugin to include reasoning/thinking content in the Span (#22997, @robinnarsinghranabhat)
• [UI] feat: enable parent-child run view in GenAI mode (#24790, @pasainatan)
• [Tracing] Add workspace support to Claude Code plugin (#23773, @Gkrumbach07)
• [Evaluation] Add Databricks scorer versioning support (#24940, @nihalmenon)
• [Tracking] Add artifact_path support for logged model artifacts (#23147, @Sanskar121543)
• [Tracing / UI] [3/N] Add custom view persistence model and prompt context builder (#24794, @aaronteo-db)
• [Tracing / UI] [2/N] Add data-source binding and A2UI message/template validator (#24792, @aaronteo-db)
• [Gateway] Add 13 new models to Databricks AI Gateway model catalog (#24893, @TomeHirata)
• [Evaluation / Tracking / UI] Add Assistant analyze action to evaluation runs (#24865, @joshuawong-db)
• [UI] [1/N] Updated A2UI catalog schema and added typed primitives for custom trace views (#24788, @aaronteo-db)

Bug fixes

• [Tracing] Fix host-only Databricks CLI commands in tracing setup (#25529, @B-Step62)
• [Model Registry] [BUGFIX] Preserve empty webhook fields in partial update_webhook calls (#25509, @CoralGarden52)
• [UI] Wrap the traces toolbar so custom-date controls don't overflow off-screen (#25512, @joshuawong-db)
• [Tracing / UI] Fix V4 trace drawer review action (#25510, @joshuawong-db)
• [Tracing / UI] Fix Custom Views in Traces V4 (#25488, @joshuawong-db)
• [Evaluation / Tracking] Fix proto conversion of DatasetRecordSourceType.UNSPECIFIED dataset record sources (#25230, @2sumtech)
• [Tracing / UI] Fix Traces V4 "1 of 0" footer count on exact trace-id search (#25483, @joshuawong-db)
• [Tracing] fix: prevent async trace queue deadlock during flush (#25264, @mukeshbhandarkar)
• [Evaluation] Fix IndexError in merge_records after to_dict/to_df on empty EvaluationDataset (#25498, @TomeHirata)
• [Tracking] Fix integer/varchar type mismatch in WorkspaceAwareSqlAlchemyStore experiment_id filters (#25232, @mayuriphad)
• [Model Registry] [BUGFIX] Preserve status in partial update_webhook calls (#24842, @abligail)
• [Tracking] Fix search_logged_models quoted-value parsing and invalid-operator message (#25398, @VenishPaneliya)
• [Tracing] Keep @mlflow/openai Chat Completion spans open during streaming (#25416, @copilot-swe-agent)
• [Build] Fix the Helm chart ServiceMonitor rendering a null metrics path (#25365, @aqeelat)
• [Scoring] Return 400 instead of 500 for a non-UTF-8 scoring request body (#25173, @tushardev-365)
• [UI] Fix: show assistant floating button on remote servers with MLFLOW_ENABLE_REMOTE_ASSISTANT=true (#25394, @TomeHirata)
• [UI] Include traceId query param in trace navigation links (#25375, @NJAHNAVI2907)
• [Models] Fix pyfunc load crash on ONNX (name, options) provider entries (#25310, @rrioh)
• [Tracking] fix: use chunked I/O in HdfsArtifactRepository to prevent OOM (#21941, @OneSizeFitsQuorum)
• [Build / Tracking] Fix Windows CI failures: drive-letter paths and OS error messages (#25382, @TomeHirata)
• [Models] Fix empty Responses Agent output streams (#25388, @copilot-swe-agent)
• [Prompts] Remove redundant text from GEPA reflective datasets (#25387, @copilot-swe-agent)
• [Evaluation] Fix stale expectations and record counts on dataset export (#25265, @mprahl)
• [Docs] fix: correct max_output_tokens for 10 Databricks-hosted models (#25360, @TomeHirata)
• [Tracing / UI] Normalize span links from OTLP endpoints in the trace explorer (#25319, @khaledsulayman)
• [Tracing] Flush the SpanBatcher from the trace flush paths so UC-table spans are exported (#25260, @alekseevpavel04)
• [UI] Fix autoscale layout state in metric charts (#25314, @PrithviBadiga)
• [UI] Preserve chart section when switching chart type in RunsChartsConfigureModal (#25283, @DaoDaoNoCode)
• [Scoring / Tracking] Fix scorer experiment_id VARCHAR binds on PostgreSQL + psycopg v3 (#25315, @rexminnis)
• [Gateway] Fix streaming passthrough skipping fallback (#25268, @TomeHirata)
• [Evaluation] Propagate eval_retry_context flags into the scorer thread pool (#25318, @joshuawong-db)
• [] feat: add rate limiting and retry support to automatic evaluation path (#24702, @rahul-oss-sap)
• [UI] fix: Ensure consistent ordering of runs for charts (#25019, @pasainatan)
• [Tracing] Fix basic auth overriding Kubernetes auth in Python SDK (#25131, @karimitanii)
• [Tracking] Ignore S3 directory markers in artifact listing (#23250, @Vinno97)
• [Tracking] [8/8] basic-auth: gate native FastAPI job routes on per-job ownership (#25102, @PattaraS)
• [Gateway / Tracking] [7/8] basic-auth: fail-closed enforcement for native FastAPI routes (#25101, @PattaraS)
• [Evaluation / Tracking] [6/6] basic-auth: gate presigned/invoke/demo/metric-twin; empty the debt list (#25071, @PattaraS)
• [Gateway] basic-auth: redact gateway secrets-config passphrase signal from non-admins (#25298, @PattaraS)
• [UI] Fix stale deps and example in the Optimize Prompt modal (#25143, @PattaraS)
• [Gateway / Tracking] [5/6] basic-auth: gate gateway guardrail/budget/config/discovery + filter list endpoints (#25070, @PattaraS)
• [Tracing / Tracking] Fix $0.00 cost panel for bare Databricks serving-endpoint model names (#24620, @adamgurary)
• [Tracing / Tracking] Read top-level cache_read/creation token counts in openai autolog (#24618, @adamgurary)
• [UI] [MLflow Assistant] Fix trace drawer closing when opening the Assistant provider picker (#25255, @aaronteo-db)
• [] [Security] Deny Assistant Read tool file access when no project directory is configured (#24513, @LiteshGhute)
• [Tracking] Skip entry_point.getLogger() when unavailable in dynamic token config provider (#25207, @serena-ruan)
• [UI] Fix visibility control bug for Evaluation Runs in Gen AI mode (#25094, @pasainatan)
• [Tracking] [4/6] basic-auth: gate job routes on per-job ownership (#25069, @PattaraS)
• [Tracing / Tracking] Offload log_spans() from the async event loop (#25185, @HumairAK)
• [Evaluation / Tracing / Tracking] Fix late-binding closure in configure_autologging_for_evaluation configuring the wrong flavor (#25135, @Koustav-github)
• [Evaluation / Tracking] [3/6] basic-auth: gate issue routes on experiment permission (#25067, @PattaraS)
• [Models] Parse schema datetimes as nanoseconds under pandas 3 (#25119, @harupy)
• [Models] Bound pandas in logged requirements for prophet <1.2.2 (#25118, @harupy)
• [Evaluation / Tracking] [2/6] basic-auth: gate evaluation dataset routes on experiment permission (#25066, @PattaraS)
• [Tracking] [1/6] basic-auth: fail-closed authorization net + coverage guard (#25065, @PattaraS)
• [Tracing / UI] Add copy button for markdown-rendered span values in trace explorer (#24619, @adamgurary)
• [UI] Link evaluation datasets from the run overview dataset drawer (#25061, @PattaraS)
• [Tracing] Map OpenInference prompt-cache token counts in OpenInferenceTranslator (#25039, @felix-xsenda)
• [Tracking] Prevent object-store artifact deletion from removing sibling paths (#25024, @yangbaechu)
• [Docs / Tracing] Honor explicit set_outputs() inside the @mlflow.trace decorator (#24624, @adamgurary)
• [Tracking] [tracking] Fix system metrics resume step without metric history mate… (#23401, @iis-MarkKuang)
• [] fix(gateway): handle missing finish_reason in openai_compatible streaming (#24111, @lavkeshdwivedi)
• [Tracking] fix: Stop SearchUtils._get_value from silently dropping dataset IN clause values (#25045, @kaikaila)
• [Tracing] Surface trace-export auth failures at ERROR instead of silently dropping traces (#24847, @adamgurary)
• [Models] Fix PyFunc artifact path collisions (#24681, @yangbaechu)
• [Build] Bump Docker base image from python:3.10-slim-bullseye to python:3.11-slim-trixie (#24961, @mille-printemps)
• [Models] Fix remaining torch_dtype deprecation warnings on load and save paths (#25017, @rrioh)
• [Tracking] Bound log_batch duplicate-metric recovery in SqlAlchemyStore by batch size (#24683, @jayceguo)
• [] Fix pydantic-ai-slim version detection in autologging (#24959, @jiahao6635)
• [Gateway] fix(gateway): preserve provider usage extras (#25010, @uuzzrm)
• [Tracking] fix(langchain): elevate secondary autolog exception handlers to logger.warning (#22040, @adamgurary)
• [UI] fix(ui): Batch run fetching in CompareRunPage to avoid one request per run (#24894, @daniellok-db)
• [Tracing / Tracking] perf(tracing): index trace archival candidate ordering (#24976, @medisean)
• [Gateway] Preserve provider usage extras in OpenAI-compatible gateway (#25000, @uuzzrm)
• [Models] Fix Windows separators in pyfunc artifact paths (#24432, @evanwrm)
• [Tracking] Fix cross-batch deadlock in concurrent log_batch latest_metrics writes (#24941, @Michael97-li)
• [Tracking] Avoid re-enabling externally-disabled loggers in _configure_mlflow_loggers (#24957, @aaronteo-db)
• [Evaluation] Preserve schema maps when enforcing strict JSON schemas (#24193, @sairavuri-sudo)
• [Tracing] Fix flaky test_search_traces_yields_expected_dataframe_contents by flushing before snapshotting (#24996, @harupy)
• [Gateway] Handle zstd-compressed request bodies on the AI Gateway (#24784, @midhunsomu)
• [Tracing / Tracking] Surface Anthropic prompt-cache token counts in mlflow.pydantic_ai.autolog (#23052, @JasonPan)
• [Tracing] Resolve UC path from active experiment when using get_trace with a plain ID (#24700, @aishdbx)
• [Tracing] Validate trace exists in link_prompts_to_trace (fixes #24067) (#24303, @zibranxo)
• [UI] Hide redundant metric names in chart cards (#23212, @xianjianlf2)
• [Gateway] Fix Bedrock gateway grouping of parallel tool results (#24309, @shivamAmrutia)
• [Tracing / Tracking] fix: use response model name for openai chat (#22325, @hansen-young)
• [Gateway] Reject unspecified linkage_type in gateway model configs instead of returning 500 (#24664, @KunalBharadwaj)
• [Docs / Evaluation / Tracing] Fix: Add trace completion buffer to prevent online evaluator from skipping long-running traces (#22006, @amitmodi)
• [Models] Fix torch_dtype deprecation warning on transformers model load (#24929, @serena-ruan)
• [Tracing] Fix basic-auth after-request handler registration (#24322, @sun-andsky)
• [Gateway] Support Vertex AI eu/us multi-region endpoints in gateway provider (#24932, @aaronteo-db)
• [UI] Allow valid Kubernetes workspace names with consecutive hyphens (#24229, @divya0709)
• [Models / Tracking] Fix model-metric-dataset association broken by stale cached ActiveRun outputs (#21855, @daniellok-db)
• [Tracing] Capture Bedrock prompt caching token fields and normalize input tokens to include them (#24899, @gauthierpiarrette)
• [Tracking] Pre-import databricks.sdk in Databricks to avoid telemetry deadlock (#24841, @aaronteo-db)
• [Tracing] Support haystack 3.x in haystack autologging (#24744, @harupy)
• [Tracing] Honor UC-backed trace storage in Databricks model serving (#24806, @james-fletcher-db)
• [Tracking] Do not hard-fail is_flavor_supported_for_associated_package_versions when flavor module is not installed (#24846, @adamgurary)
• [UI] fix(ui): Fix Chart view unresponsiveness via virtualization in RunsChartsDraggableCardsGridSection (#24012) (#24424, @Alwaysgaurav1)
• [] Fix claude_code provider exceeding Windows cmd.exe command-line limit (#24440, @joshuawong-db)
• [UI] Fix RestoreRunModal to show actual error instead of hardcoded message (#24731, @mihir2921)
• [UI] Fix DeleteRunModal to show actual error instead of hardcoded message (#24730, @mihir2921)
• [UI] [UI] Show error notification when permission-denied for USER role actions (#24722, @mihir2921)
• [Build / Models] Add support for pandas 3.x (#23719, @twanahc)
• [Model Registry] Fix model version string to int cast in registry stores (#24565, @qubeena07)

Documentation updates

• [Docs] Fix environment variable expansion in docker/README.md (#25474, @CoralGarden52)
• [Build] Use Chart.appVersion for the Helm example image (#25458, @CoralGarden52)
• [Docs / Tracing] docs: add TypeScript examples to Kubernetes auth page (#25083, @robinnarsinghranabhat)
• [Docs] Clarify trace-data scope in detect-issues.mdx (#25386, @copilot-swe-agent)
• [Docs] Add Red Hat OpenShift AI to hosting options (#25151, @B-Step62)
• [Docs / Tracing] Honor explicit set_outputs() inside the @mlflow.trace decorator (#24624, @adamgurary)
• [Docs / Tracing] docs: clarify Databricks profile, LangGraph autolog, runs-vs-traces, span attributes (#24954, @adamgurary)
• [Docs / Tracing] docs: warn that delete_traces() does not support Databricks Unity Catalog traces (#24647, @adamgurary)
• [Docs / Gateway] Clarify LiteLLM is an optional extension in AI Gateway docs (#24787, @TomeHirata)

Small bug fixes and documentation updates

#25558, #25553, #25392, #25421, #25407, #24828, #25295, #25249, #25055, #24827, @joshuawong-db; #25547, #25127, #24884, #25501, #24714, #24913, #24879, @B-Step62; #25544, #25542, #25540, #25538, #25537, #25534, #25531, #25530, #25506, #25466, #25436, #25408, #25405, #25390, #25320, #25256, #25023, #25014, #24993, #24992, #24988, #24986, #24955, #24872, #24855, #24856, @copilot-swe-agent; #25545, #25502, #25486, #25485, #25482, #25481, #25478, #25475, #25468, #25467, #25464, #25462, #25461, #25465, #25457, #25456, #25452, #25450, #25447, #25446, #25443, #25437, #25430, #25432, #25423, #25429, #25409, #25420, #25417, #25419, #25401, #25404, #25402, #25396, #25393, #25368, #25358, #25346, #25344, #25343, #25342, #25341, #25339, #25338, #25337, #25335, #25334, #25325, #25316, #25312, #25277, #25276, #25252, #25165, #25163, #25162, #25161, #25160, #25159, #25158, #25156, #25154, #25153, #25152, #25144, #25142, #25124, #25122, #25117, #25109, #25110, #25106, #25103, #25100, #25099, #25096, #25090, #25092, #25091, #25088, #25048, #25047, #25042, #25041, #25037, #25030, #25032, #25031, #25029, #25022, #24996, #24991, #24990, #24987, #24982, #24980, #24977, #24978, #24975, #24974, #24973, #24972, #24971, #24970, #24968, #24966, #24965, #24910, #24908, #24922, #24911, #24909, #24901, #24866, #24874, #24853, #24840, #24839, #24834, #24791, #24751, #24770, @harupy; #25533, #25528, #25305, #25304, #25302, #25296, #25222, @kriscon-db; #25412, #25309, @IgorShishkin12; #25520, #25519, #25015, #24945, #24943, #24944, #24902, #24904, #24905, #24903, @dependabot; #25081, #24880, #24802, @mprahl; #25505, #25500, #25480, #25379, #25360, #25351, #24754, @TomeHirata; #25477, #25459, @CoralGarden52; #25180, @miyakido; #25369, @d0w; #25400, @ArrayPD; #25391, #25329, #25104, @tanghaoji; #25374, @brandonhawi; #24276, #24222, @khaledsulayman; #25272, @kaikaila; #25300, #25097, #25046, #24838, @PattaraS; #24998, @juliafmorgado; #25251, #25212, #25211, #25186, #25060, #25059, #25028, #25025, @aaronteo-db; #25202, @aravgarg28; #22653, #jeevan6996; #25164, @jaceklaskowski; #24925, @darkdi; #25040, @rrioh; #24914, @smoorjani; #24892, @codechrl; #24927, @serena-ruan; #24323, #24678, @NeriCarcasci; #24516, @gauthierpiarrette; #24732, @mihir2921; #24779, @daniellok-db

MLflow 3.16.0 puts you in control of how you observe your GenAI apps. The headline: build your own trace UI just by describing it — the MLflow Assistant turns a plain-English prompt into a custom, reusable trace view. On top of that, this release ships a fully redesigned trace explorer as the default experience and first-class span links for connecting related spans across traces. Here's what's new.

1. Custom Trace Views

Every team inspects traces differently — so why settle for one fixed layout? In 3.16.0, you can design your own trace UI in plain English. Describe the view you want, and the MLflow Assistant builds it for you: the fields you care about, laid out the way you think about them. It's the power of an LLM and MLflow working together — no config files, no custom code, just a prompt.

Found a layout that clicks? Save it, name it, and reuse it. Views persist per experiment, so your whole team can share the same lens on your traces. Create as many as you need, rename them as your workflow evolves, and delete the ones you've outgrown. Whatever mental model you have for your agent, you can now make the UI match it.

2. A Redesigned Trace Experience

The trace explorer has been rebuilt from the ground up and is now the default traces experience. The new Traces table brings tighter row density, a cleaner header, and smoother navigation, alongside a redesigned span-tree explorer. Reorder columns to put what matters first, and add configurable custom columns backed by any trace tag or metadata field — so the details you care about are always one glance away.

Sessions are now folded directly into the traces view via session grouping, so multi-turn conversations live right alongside the rest of your traces.

3. Span Links

Spans rarely stand alone — a retrieval step, a tool call, or a downstream trace is often related to the span you're looking at. MLflow 3.16.0 adds first-class span links so those relationships are captured and navigable. Record links between spans in your SDK, and the trace explorer surfaces them in a dedicated Links tab — click through to jump straight to the destination span.

Full Changelog

For a comprehensive list of changes, see the release change log.

What's Next

Get Started

Upgrade to try these new features:

pip install mlflow==3.16.0

Share Your Feedback

We'd love to hear about your experience with these new features:

• GitHub Issues - Report bugs or request features
• MLflow Roadmap - See what's coming next and share your ideas
• ⭐ Star us on GitHub - Show your support for the project

Learn More

Check out the MLflow documentation for detailed guides

For a comprehensive list of changes, see the release change log, and check out the latest documentation on mlflow.org.

MLflow 3.15.2 is a patch release that includes several major features and improvements.

Features

• [Evaluation] Support immutable evaluation dataset versions (#24845, @danielseong1)
• [Evaluation] Add scorer_ensemble primitive for combining scorer results (#24749, @alkispoly-db)

Bug fixes

• [Evaluation] Preserve base judge invocation flow in MemAlign aligned judges (#24883, @veronicalyu320)
• [Tracking] Pre-import databricks.sdk in Databricks to avoid telemetry deadlock (#24841, @aaronteo-db)
• [Build / Tracking] Align runs.status constraint metadata (#24890, @joshuawong-db)

MLflow 3.15.2 is a patch release that includes several major features and improvements.

Features

• [Evaluation] Support immutable evaluation dataset versions (#24845, @danielseong1)
• [Evaluation] Add scorer_ensemble primitive for combining scorer results (#24749, @alkispoly-db)

Bug fixes

• [Evaluation] Preserve base judge invocation flow in MemAlign aligned judges (#24883, @veronicalyu320)
• [Tracking] Pre-import databricks.sdk in Databricks to avoid telemetry deadlock (#24841, @aaronteo-db)
• [Build / Tracking] Align runs.status constraint metadata (#24890, @joshuawong-db)

For a comprehensive list of changes, see the release change log, and check out the latest documentation on mlflow.org.

MLflow 3.15.1 is a patch release that includes bug fixes and documentation updates.

Bug fixes:

• [Model Registry] Skip env_pack on ARM client images (#24762, @qyc)
• [Scoring / Tracking] Harden version parsing against missing/non-PEP440 versions on Databricks Serverless (#24799, @PattaraS)

Documentation updates:

• [Docs] Clarify scorer versioning documentation (#24769, @nihalmenon)

For a comprehensive list of changes, see the release change log, and check out the latest documentation on mlflow.org.

MLflow 3.15.0 includes several major features and improvements

Major New Features

🧩 MCP Registry: A centralized catalog for registering, versioning, and sharing Model Context Protocol servers — with semantic-versioned configs, promotable aliases, tags, auto-discovered tools, and ready-made connection instructions for Claude Code and .mcp.json. Manage it from the UI, REST API, or Python!

🤖 MLflow Assistant enhancements: The in-app AI assistant now supports multiple LLM providers (Claude Code, Codex, and OpenAI-compatible/Gateway endpoints) chosen from a single settings page, displays live per-session token usage and estimated cost in the composer, and is easier to set up — mlflow agent setup can enable it in one prompt, with API keys stored securely in the Gateway's LLM Connections.

🗂️ Sharable table views: Save named views of the Runs table — capturing columns, order, widths, filters, and sort — and share them by URL.

⚡ Proxy-less artifact upload/download via presigned URLs: Large artifact transfers can now bypass the tracking server and talk directly to cloud storage (e.g. S3) through presigned URLs, cutting server load and timeouts on big files. We fall back to proxied transfer automatically for backward compatibility.

🖼️ Multi-modal attachments in LLM judges: {{ trace }} judges created with make_judge() can now see images captured in trace spans via a new get_span_image tool, which fetches attachments as base64 data URLs and passes them to multimodal models (Anthropic, Gemini, and OpenAI-compatible via litellm) — enabling evaluation of vision tasks, screenshots, and other visual content.

Other Assorted Features & Improvements:

• [Tracing / Tracking] Add Pydantic AI 2.x autologging support (#24721, @B-Step62)
• [Tracing / UI] [3/3] Add action recommendation for low-value issue detection result (#24580, @B-Step62)
• [Gateway] Support per-endpoint budget policies in the AI Gateway (#24370, @joelrobin18)
• [Evaluation / Gateway] Add sap-ai-core:/ provider for SAP AI Core Orchestration v2 with extra_headers support (#24356, @shivamxtech)
• [Model Registry] [Model Registry] Migrate UC model artifact download to native temp-creds (#24576, @AknA13)
• [Gateway / Tracing / UI] [2/3] Simplify issue detection start modal (#24579, @B-Step62)
• [Tracing / UI] [1/3] Run issue detection in the background (#24578, @B-Step62)
• [Tracing] [BUG] Fix otel parent context propagation (#24444, @iyashk)
• [Model Registry] [Model Registry] Gate UC model-registry client on native /api/2.1 endpoints (#24517, @AknA13)
• [Docs / Tracking] Enable artifacts-only mode with workspaces (#24452, @kramaranya)
• [Tracing / UI] Stop eagerly prefetching full neighboring traces in tracing UI (#24411, @DaoDaoNoCode)
• [Docs / Gateway] Authenticate OpenAI-protocol coding agents through the RBAC gateway (#24294, @joshuawong-db)
• [Tracing] feat: add search traces to typescript sdk (#23660, @WUMIKE233)
• [Prompts / UI] Add per-tool cards with a Monaco JSON editor and Format action to the Playground (#24129, @tanghaoji)
• [Prompts / UI] Redesign Playground Tools settings with an Add-tools entry flow (#24102, @tanghaoji)
• [] feat: add Cursor model catalog with Auto and Composer 2.5 pricing (#24078, @TomeHirata)
• [Tracing / Tracking] Support numeric assessment trace filters (#23948, @tanghaoji)

Bug fixes:

• [Tracing] Compute mlflow.llm.cost for Claude Code plugin traces on Databricks (#24416, @harupy)
• [UI] Fix artifact direct download capability checks in UI (#24748, @B-Step62)
• [Tracing / UI] Refresh traces saved-view on load across multiple windows (#24746, @joshuawong-db)
• [UI] Fix UnauthorizedError constructor message (#24729, @mihir2921)
• [UI] Fall back to nested spans for Chat Sessions turn content (#24703, @MannXo)
• [Tracing / Tracking] Fix experiment GC cascades for traces and logged models (#24351, @copilot-swe-agent)
• [Tracking] Fix workspace-enabled tracking store crash-loop when the default experiment is renamed (#24598, @daniellok-db)
• [Tracking] Fix Kubernetes auth token extraction for kubernetes-client 36+ (#24687, @HumairAK)
• [UI] Refresh Assistant config when leaving settings so the provider indicator isn't stale (#24704, @joshuawong-db)
• [Build] Fix imagePullSecrets for garbage collection CronJob (#24626, @nightcityblade)
• [UI] Lift Assistant floating button above bottom-pinned action bars (#24606, @joshuawong-db)
• [Evaluation / Tracing] Fix issue detection structured output schema (#24674, @B-Step62)
• [Models] Add MLFLOW_ALLOW_PICKLE_DESERIALIZATION guard to mlflow.statsmodels flavor (#24686, @B-Step62)
• [Tracking] Preserve static prefixes in notebook trace renderer URL (#24682, @mprahl)
• [UI] Fix blank and fields in run detail sidebar (#24634, @aishdbx)
• [Models / Scoring] Normalize reasoning content parts for ChatAgentMessage in parse_message (#24642, @iyashk)
• [UI] Fix Evaluation Runs Actions dropdown remaining clickable when no runs are selected (#24676, @daniellok-db)
• [UI] Surface truncated assistant streams as errors instead of a silent done (#24603, @joshuawong-db)
• [Tracking] Honor experiment grants on --artifacts-only servers when workspaces disabled (#24585, @PattaraS)
• [Tracking / UI] Bound get-history-bulk-interval memory by sampling metric history by row (#24305, @Mytolo)
• [Tracking] [BUG] Enforce artifact authorization on presigned download route in basic-auth (#24571, @iyashk)
• [Evaluation] Preserve explicit expectations in mlflow.genai.evaluate (#24561, @yangbaechu)
• [Gateway / UI] Support Portkey routing configuration in the AI Gateway (#24398, @rs-03)
• [UI] Show input modal submission errors inline instead of a transient toast in GenericInputModal (#23918, @Vedant-Agarwal)
• [Evaluation / Tracking] Enforce experiment permission on online scoring config endpoints (#24562, @aaronteo-db)
• [Tracing / Tracking] Propagate contextvars into trace read thread pool workers (#24456, @iyashk)
• [] Fix Windows path handling in artifact downloads (#20225, @acepace)
• [UI] Fix sidebar navigation and workflow type not updating on URL navigation or browser Back/Forward (#22116, @DaoDaoNoCode)
• [Gateway] Don't send stream_options from the Assistant to the gateway (#24552, @joshuawong-db)
• [Evaluation] Fix MemAlign guideline distillation against Databricks structured-output models (#24547, @veronicalyu320)
• [UI] fix: sort tags alphabetically in runs tag multiselect dropdown (#24134, @mysticalayushi)
• [Tracing] Fix LLM cost on autologged ChatDatabricks traces by preferring ls_provider (#24185, @james-fletcher-db)
• [Tracking] Fix pytorch pt2 export for batch-size-1 input_example (use Dim.AUTO) (#24494, @B-Step62)
• [Tracing / Tracking] Fix double-counted trace token usage for rollup parent spans in SqlAlchemyStore.log_spans (#24339, @gauthierpiarrette)
• [Tracking] Fix MlflowSparkStudy pruner and direction handling (#24425, @hellocybernetics)
• [Gateway / Tracing / UI] Fix gateway Try-in-Browser traces rendering as raw kvlist data in the UI (#24400, @joelrobin18)
• [UI] Fix role modal resource pickers failing with FEATURE_DISABLED when workspaces are disabled (#24441, @joelrobin18)
• [Model Registry] [Model Registry] Restore oneof on TemporaryCredentials.credentials (#24489, @AknA13)
• [UI] [fix] Fix runs-table column reordering before pinned anchors (#24465, @joshuawong-db)
• [UI] Fix admin permission grants failing with FEATURE_DISABLED when workspaces are disabled (#24383, @joelrobin18)
• [UI] Fix Columns dropdown anchoring on the runs table (#24419, @joshuawong-db)
• [UI] Fix artifact viewer crash on text files with many lines (#23899, @jayceguo)
• [Tracing] Fix async trace export dropping workspace context (#24093) (#24275, @vatsalyd)
• [Tracing] fix: store OpenAI agent Generation span attributes under correct SpanAttributeKey constants (#24290, @RudraDudhat2509)
• [Tracing / Tracking] Fix Postgres deadlock between start_trace() and log_spans() on trace metadata (#24338, @joshuawong-db)
• [Tracing / Tracking] Avoid eager span deserialization for TRACKING_STORE trace reads. (#24331, @shivamAmrutia)
• [Gateway] Resolve workspace context before running FastAPI auth validators (#24368, @joelrobin18)
• [Tracing] Fix RecursionError corrupting deeply nested traces (#24362, @joshuawong-db)
• [Evaluation / Tracing] Fix genai.evaluate crash when a cloned trace is unavailable (#24357, @joshuawong-db)
• [Scoring] Handle uncut {major}.x minor in DatabricksRuntimeVersion.parse (#24353, @PattaraS)
• [Evaluation / Tracing] Fix basic-auth before-request validators invoking view functions (#24354, @joshuawong-db)
• [UI] Fix promo modal focus (aria-hidden) and show it on the empty models page (#24278, @aminehd)
• [Scoring] Fix spark_udf InvalidVersion on Databricks Serverless for {major}.x runtime image strings (#24336, @PattaraS)
• [Tracing] Fix trace OTLP encoding for lone surrogates (#24318, @sanatb187)
• [Model Registry / Models] Forward tags to registered ModelVersion in Model.log() (#24271, @Ankita7033)
• [Tracing / UI] Fix duplicate full-trace fetch for TRACKING_STORE traces in the trace UI (#24319, @PattaraS)
• [Models / Scoring] Honor declared ONNX execution providers at load time (#24232, @qyc)
• [Evaluation] Fix pos_label not propagated to evaluate ROC/PR/calibration curves (#24049, @immu4989)
• [Tracking] Resolve experiment id from workspace-prefixed artifact paths in auth proxy (#24214, @rakshithbk-moeng)
• [] Fix DatabricksSDKModelsArtifactRepository auth on serverless executors (#24247, @harupy)
• [Evaluation] Send strict-compatible response_format for AI Gateway judges and scorers (#24048, @B-Step62)
• [Model Registry] Require read permission on the source run or model when creating a model version (#24293, @PattaraS)
• [] Fix Vertex AI Gemini multi-turn tool calling (#24266, @joshuawong-db)
• [Tracking] Enforce update-run authorization on the LogInputs endpoint (#24291, @PattaraS)
• [Tracking] Honor MLFLOW_S3_IGNORE_TLS in multipart artifact upload (#24113, @yashmahamulkar)
• [Gateway / Tracking] Reject non-HTTP scopes in WSGI mount to fix WebSocket crash (#24274, @joshuawong-db)
• [Model Registry / Tracing / Tracking] Fix tracing provider thread leak and set_experiment silently re-enabling tracing (#24267, @PattaraS)
• [Tracing] Fix @mlflow/opencode dropping session metadata against published core (#24252, @joshuawong-db)
• [UI] [BUG] Fix #24170: surface direct (per-resource) permissions in Admin UI (#24253, @nghessss)
• [Model Registry] Fix DNS-rebinding SSRF bypass in webhook delivery (#24258, @PattaraS)
• [Tracing / Tracking] Fix Summary ignoring valid=False Assessments (#24251, @joshuawong-db)
• [Evaluation] Route DSPy judge optimizer Databricks models to Mosaic AI Gateway for Unity Catalog names (#24177, @veronicalyu320)
• [UI] Allow granting of tool call permissions on MLflow assistant (#24084, @joshuawong-db)
• [Build] Fix Python 3.14 compatibility in skill_installer by using pathlib.Path (#24103, @krishtyagi0109-pixel)
• [UI] Render playground tool-call args and JSON responses as code blocks (#24086, @tanghaoji)
• [Gateway] Support Gemini thought signature in AI Gateway (#24051, @vatsalyd)
• [UI] fix(playground): render tool_calls so prompts run with tools no longer show empty result (#24073, @tanghaoji)

Documentation updates:

• [Docs] Add documentation for MCP Registry feature (#24713, @dkuc)
• [Docs / Tracing] docs: troubleshoot $0 trace cost for unknown model names (#24705, @daniellok-db)
• [Docs] docs: credential-expiry trace-drop cause in tracing FAQ, experiment_id in quickstart (#24692, @adamgurary)
• [Docs] Add Swagger UI API docs for MCP Server Registry (#24519, @dkuc)
• [Docs] docs: update auth REST API docs to reflect new RBAC endpoints (#23965, @Farzah11)
• [Docs] docs: fix broken auth REST API links causing Access Denied (#24280) (#24350, @Farzah11)
• [Docs] Add Artifact Hub repository metadata and OCI install docs for the Helm chart (#24313, @B-Step62)
• [Docs] docs: update Omnigent integration page for pure OTel setup (#24240, @TomeHirata)
• [Docs] docs: add Omnigent tracing integration page (#24072, @TomeHirata)

Small bug fixes and documentation updates:

#24746, #24747, #24671, #24637, #24655, #24459, #24510, #24629, #24426, #24554, #24555, #24556, #24359, #24460, #24445, #23676, #24446, #24268, #24417, #24358, #24288, #24250, #24206, #24165, #23939, #23936, #23881, @joshuawong-db; #24743, #24723, #24720, #24694, #24695, #24706, #24698, #24693, #24636, #24675, #24673, #24672, #24663, #24632, #24630, #24628, #24627, #24625, #24548, #24216, #24208, #24205, #24158, #24159, #24136, #24095, #24119, #24106, #24104, #24085, @harupy; #24736, @henishborad; #24717, #24583, #24582, #24584, #24685, #24621, #24604, #24521, #24482, #24495, #24484, #24480, #24481, #24490, #24374, #24317, #24299, @B-Step62; #24656, @rrioh; #24696, #24600, #24601, @daniellok-db; #24657, @DaoDaoNoCode; #24217, @nananosirova; #24645, #24377, #24492, #24380, #21417, #24346, #24297, #24108, @mprahl; #24477, @dkuc; #24572, #24520, #24479, @jonburdo; #24505, @PattaraS; #24242, @Killian-fal; #24558, #24540, #24090, #24139, #24089, #24061, #24032, #24030, @aaronteo-db; #24176, @zcin; #24412, @AknA13; #24437, @smoorjani; #24361, @sun-andsky; #23873, @WeichenXu123; #24240, #24203, #24178, @TomeHirata; #24225, #24204, #24166, #24054, @tanghaoji; #23794, #24171, @kramaranya; #23251, @aakankshaduggal; #24151, #24150, @xq-yin; #24075, @kevin-lyn

MLflow 3.15.0

MLflow 3.15.0 is all about making GenAI development faster and more collaborative. This release introduces a centralized MCP Registry for managing Model Context Protocol servers, a significantly upgraded MLflow Assistant with multi-provider support and a friction-free setup, shareable table views for the Runs table, proxy-less artifact transfers for big files, and multimodal LLM judges that can finally see the images in your traces. Here's what's new.

1. MCP Registry

Model Context Protocol servers are quickly becoming the connective tissue between agents and the tools they use — but until now, keeping track of them meant passing around config snippets and hoping everyone had the right version. The new MCP Registry gives you a single, centralized catalog to register, version, and share MCP servers across your team.

Every server gets semantic-versioned configs, promotable aliases (think @production and @staging), and tags for easy organization. MLflow auto-discovers each server's tools so you always know what a given version exposes, and it generates ready-made connection instructions for both Claude Code and .mcp.json — copy, paste, and you're connected. Manage the whole thing however you prefer: through the UI, the REST API, or Python.

2. A Smarter, Easier MLflow Assistant

The MLflow Assistant just got a major upgrade, and the theme is simple: less setup, more power.

• Bring your own model. The Assistant now supports multiple LLM providers — Claude Code, Codex, and API/Gateway endpoints — all selectable from a single settings page. Use the model that fits your workflow.
• Setup in seconds. Getting started is now as easy as pasting an API key directly into the pane. MLflow stores it securely in the Gateway's LLM Connections, so there's no fiddling with environment variables or config files.
• Full transparency as it works. Tool calls, approvals, and token costs are now displayed inline, so you can see exactly what the Assistant is doing — and what it's costing you — as it happens.

3. Shareable Table Views

Everyone builds their Runs table a little differently — the columns that matter, the sort order, the filters that cut through the noise. In 3.15.0, you can capture all of it. Save a named view that remembers your columns, their order and widths, your filters, and your sort — then share it with a teammate by simply sending them the URL. No more "here's how to set up your table" walkthroughs; just send the link.

4. Proxy-less Artifact Upload & Download

Large artifact transfers no longer have to funnel through the tracking server. Using presigned URLs, MLflow can now talk directly to your cloud storage (e.g. S3), cutting server load and eliminating the timeouts that plagued big-file uploads and downloads. It's faster for you and lighter on your infrastructure — and if presigned URLs aren't available, MLflow automatically falls back to the proxied transfer path, so existing setups keep working without any changes.

5. Multimodal Attachments in LLM Judges

LLM judges have been great for evaluating text — but a lot of real-world agents work with images, screenshots, and other visual content. Now your judges can see them too. {{ trace }} judges created with make_judge() gain a new get_span_image tool that fetches image attachments captured in trace spans, encodes them as base64 data URLs, and passes them straight to a multimodal model.

With support for Anthropic, Gemini, and OpenAI-compatible models (via litellm), you can now write judges that evaluate vision tasks, assess generated or captured screenshots, and reason about any visual content flowing through your traces — unlocking a whole new class of automated evaluation.

Full Changelog

For a comprehensive list of changes, see the release change log.

What's Next

Get Started

Upgrade to try these new features:

pip install mlflow==3.15.0

Share Your Feedback

We'd love to hear about your experience with these new features:

• GitHub Issues - Report bugs or request features
• MLflow Roadmap - See what's coming next and share your ideas
• ⭐ Star us on GitHub - Show your support for the project

Learn More

Check out the MLflow documentation for detailed guides

For a comprehensive list of changes, see the release change log, and check out the latest documentation on mlflow.org.

MLflow 3.14.0

MLflow 3.14.0 is a major release focused on closing the GenAI development loop, from getting an app instrumented in the first place to reviewing, testing, and iterating on it. Highlights include one-command observability onboarding with mlflow agent setup, durable low-latency tracing for Claude Code, Review Queues for collecting structured feedback on traces, a revamped evaluation dataset UI, a pytest integration that gates GenAI quality in CI, and an in-browser LLM Playground.

1. One-Command Onboarding with mlflow agent setup

Getting an app onto MLflow observability should not mean reading setup guides and wiring things together by hand. The new mlflow agent setup command does it all in a single line: it installs a curated set of MLflow skills and launches your favorite coding agent, which installs MLflow, sets up tracing, and instruments your app for you. Run it from inside your project, choose which coding agent to drive it (Claude Code, OpenAI Codex, or OpenCode) and where traces should land (a fresh local server, a Databricks workspace, or an existing one), and the agent adds tracing to your real entry point and verifies a trace end to end.

Because the actual instrumentation is delegated to your coding agent rather than a rigid script, it adapts to whatever package manager, framework, and entry point your project uses.

Run inside your project: your coding agent instruments it for MLflow Tracing

uvx mlflow@latest agent setup

Learn more in the Tracing Quickstart

2. Durable, Low-Latency Tracing for Claude Code

Rolling out Claude Code tracing to a whole team should be something you can switch on for everyone and stop thinking about. MLflow 3.14.0 makes it durable and low-overhead so organizations can adopt it with confidence: tracing never slows down the developer's Claude Code session, it will not overwhelm your central tracking server under load, and it does not lose traces when a network blip or a crash interrupts a session.

It achieves this with a write-ahead-log: each trace is written to local disk and uploaded by a background daemon that retries on failure, so the agent never blocks on the network and any traces already on disk are replayed even if the session exited unexpectedly. You can also send Claude Code traces to a Databricks Unity Catalog location with MLFLOW_TRACE_LOCATION.

Learn more about Claude Code tracing

3. Review Queues for Traces

Collecting feedback on GenAI traces used to be ad-hoc: spreadsheets, side channels, and no shared sense of what had already been reviewed. Review Queues turn that into a structured, in-UI workflow, and the quickest path needs no setup at all. From the trace table, select the traces you want a second opinion on and assign them directly to teammates, who find them waiting in their personal review queue. Each reviewer opens a focused review page and works through their traces one at a time.

When you want more structure, create a custom queue with a specific set of review questions (Pass/Fail, categorical, numeric, or free text) and assign it to a group of reviewers. Because a queue's status is shared across everyone assigned to it, the first reviewer to finish a trace clears it for the whole group. Either way, answers are never stored off to the side: each one is written straight back onto the trace, with feedback questions logging assessments and expectation questions logging ground truth. That makes review output immediately usable for evaluation, for aligning an LLM judge, or for building a dataset. The whole workflow is also scriptable through the mlflow.genai.review_queues SDK.

Learn more about Review Queues

4. Revamped Evaluation Dataset UI

Evaluation datasets are living collections of test cases that grow from production traces and hand-curated golden examples, but inspecting and maintaining them used to mean dropping into the Python SDK. MLflow 3.14.0 ships a fully revamped evaluation dataset experience in the UI. From an experiment's Datasets tab you get a searchable list of datasets, and clicking one opens a dedicated detail page with a sortable, paginated records table.

On the detail page, each record opens in a side panel where inputs and expectations are edited as validated JSON, with an unsaved-changes guard so you do not lose work. You can add records inline, multi-select and bulk-delete, choose which columns to show, and for any record that originated from a trace, open it in an in-page trace explorer to see where the test case came from. Combined with the existing "Add to evaluation dataset" action on the Traces tab, the whole build-and-maintain loop now lives in the UI.

Learn more about evaluation datasets

5. Pytest Integration for GenAI Regression Tests

If you already evaluate your agent with mlflow.genai.evaluate() and scorers, regression testing in MLflow 3.14.0 asks you to learn nothing new: add the @mlflow.test decorator to a plain pytest function, run your existing scorers, and assert on the result with assert result.passed, result.reason. When it fails, the failing scorers and their judge rationales show up right in the pytest output.

Because it is just a decorator on an ordinary pytest function, your regression suite plugs into the whole pytest ecosystem: parametrize a test across many cases, fan it out across workers with pytest-xdist, and run it in CI like any other test. Every run is captured to MLflow, and the Evaluation runs UI shows your test history and the detailed per-assertion judge results at a glance. The recommended loop is to grow the suite from real failures: each time the agent misbehaves, capture that case as a @mlflow.test so the same issue can never silently return.

import mlflow
from mlflow.genai.scorers import Guidelines

@mlflow.test
def test_answers_concisely(agent):
    result = mlflow.genai.evaluate(
        predict_fn=agent,
        data=[{"inputs": {"question": "What are your hours?"}}],
        scorers=[Guidelines(name="concise", guidelines="Answer in one sentence.")],
    )
    assert result.passed, result.reason

Learn more about regression testing

6. LLM Playground

Iterating on a prompt used to mean editing code and re-running it, or jumping out to a separate vendor playground that knew nothing about your registered prompts. MLflow 3.14.0 adds an in-browser LLM Playground, wired to the same MLflow AI Gateway endpoints and Prompt Registry versions you already govern. Open it from an experiment, pick a model endpoint, and compose a multi-turn conversation with system, user, and assistant messages.

A settings drawer lets you tune temperature, max tokens, top-p, penalties, and stop sequences; supply tool definitions; and constrain output to plain text or a strict JSON schema. Messages support {{ variable }} placeholders you fill in at submit time, and "Load prompt from registry" drops any registered version straight into the conversation, applying its stored model config automatically.

Learn more about the LLM Playground

Full Changelog

For a comprehensive list of changes, see the release change log.

What's Next

Get Started

Upgrade to try these new features:

pip install mlflow==3.14.0

Share Your Feedback

We'd love to hear about your experience with these new features:

• GitHub Issues - Report bugs or request features
• MLflow Roadmap - See what's coming next and share your ideas
• ⭐ Star us on GitHub - Show your support for the project

Learn More

Check out the MLflow documentation for detailed guides.

MLflow 3.14.0 includes several major features and improvements

Major New Features

• 🚀 One-command agent onboarding with mlflow agent setup: Install MLflow, set up tracing, and hand your favorite coding agent (Claude Code, OpenAI Codex, or OpenCode) the MLflow skills to instrument your app, all from a single command.
• ⚡ Durable, low-latency tracing for Claude Code: Roll out Claude Code tracing across a team with confidence: a write-ahead-log keeps it from slowing the agent, overwhelming the tracking server, or losing traces on a network blip or crash.
• 📝 Review Queues for traces: Assign traces to reviewers (or agents) and collect structured feedback and ground-truth annotations in the UI, written straight back onto the trace so they are immediately usable for evaluation.
• 🗂️ Revamped evaluation dataset UI: Browse, inspect, edit, and bulk-manage evaluation dataset records directly in the UI, with click-through to the source trace.
• 🧪 Pytest integration for regression testing: Write GenAI regression tests as plain pytest functions with the @mlflow.test marker, gate them in CI, and review test history and per-assertion judge results in the UI.
• 🎛️ LLM Playground: Iterate on prompts in the browser against your AI Gateway endpoints and Prompt Registry versions, with settings, tools, structured output, and template variables.

Breaking Changes

• [Models] Change mlflow.sklearn serialization_format default from cloudpickle to skops (#23987, @copilot-swe-agent)
• [Models] Change serialization_format default to "pt2" for mlflow.pytorch.log_model and mlflow.pytorch.save_model (#23988, @copilot-swe-agent)
• [Models] Change serialization_format default to "skops" in mlflow.lightgbm log_model/save_model (#23986, @copilot-swe-agent)

Other Assorted Features & Improvements

• [Evaluation / UI] [3/3] Show regression-test results in the existing eval-run UI (#23985, @B-Step62)
• [Prompts / UI] Add "Save prompt to registry" action to the Prompt Playground (#24021, @B-Step62)
• [Prompts] Prompt Playground (#23273, @TomeHirata)
• [Evaluation] [2/3] Add EvaluationResult.passed/.reason for @mlflow.test assertions (#23869, @B-Step62)
• [UI] Review queues: list the affected queues in the delete-question confirmation (#24002, @kriscon-db)
• [UI] Add shareable review queue URLs with a startReview deep link (#23941, @harupy)
• [UI] Allow editing a completed review in place in focus mode (#23967, @kriscon-db)
• [Tracing] Add x-mlflow-run-id support to OTLP trace ingestion (#23664, @sanatb187)
• [Evaluation / Tracing] [1/3] Add @mlflow.test pytest marker and assertion framework (#23864, @B-Step62)
• [UI] Improve review queue empty states with onboarding content (#23903, @B-Step62)
• [UI] Add mlflow skills view/list CLI (#23907, @joshuawong-db)
• [UI] Improve review queue list: flat layout, sortable columns, status filter (#23902, @B-Step62)
• [Tracing] Add MLFLOW_WORKSPACE support to OSS auth provider (#23927, @Nehanth)
• [Gateway] Add cached token pricing to Databricks model catalog (#23901, @TomeHirata)
• [Evaluation] Add MLFLOW_GENAI_JUDGE_DEFAULT_MODEL environment variable (#23860, @B-Step62)
• [Evaluation] Wire "Run judge(s)" submission in "Run Eval" in Evaluations Run page to POST /mlflow/genai/evaluate/invoke (#23781, @aaronteo-db)
• [Evaluation] Add rule-based built-in scorers: RegexMatch, PIIDetection, ResponseLength (#22571, @debu-sinha)
• [Tracing] Support Databricks backend in mlflow agent setup (#23783, @harupy)
• [Evaluation] Add POST /mlflow/genai/evaluate/invoke handler & job for UI-triggered eval runs (#23779, @aaronteo-db)
• [Tracing] [Claude Code] Support UC trace location via MLFLOW_TRACE_LOCATION (#23770, @B-Step62)
• [Tracing] [Codex] Support UC trace location via MLFLOW_TRACE_LOCATION (#23771, @B-Step62)
• [Evaluation / Tracking] [3/N] Label schemas: handlers + SDK + REST client (#23603, @kriscon-db)
• [Tracing / Tracking] Add run_id support for trace APIs (#23629, @sanatb187)
• [Evaluation / UI] Dataset v2 port (#23560, @B-Step62)
• [Evaluation / Tracking] Add OSS-native label schema entity, validation, and SQL store (#23597, @kriscon-db)
• [Tracing] Support mapping gen_ai.conversation.id to MLflow trace session (#23584, @SahilKumar75)
• [Tracing / UI] Added polling logic to live check and auto-refresh traces tab empty state with trace first ingestion (#23184, @vivian-xie-db)
• [Tracing] Add MlflowWalSpanExporter to hand traces off to the WAL daemon (#23641, @aaronteo-db)
• [Tracing] Support native UC trace ingestion from TypeScript SDK (#23562, @B-Step62)
• [Evaluation] Add Google ADK LLM judge scorers (Hallucination, Safety, ResponseEvaluation) (#22496, @debu-sinha)
• [Gateway] Add OpenAI /responses/compact passthrough route to AI Gateway (#23353, @id-jazzx)
• [Gateway] Add 21 new models to Databricks model catalog (#23520, @TomeHirata)

Bug fixes

• [Evaluation] Fix ChrfScore RAGAS scorer instantiation due to class name mismatch (#24047, @B-Step62)
• [Tracing / Tracking] Map OpenAI Agents SDK guardrail spans to SpanType.GUARDRAIL (#24044, @B-Step62)
• [UI] Surface review-question modal failures as toasts (#24035, @kriscon-db)
• [Tracking] Prevent review queues from shadowing usernames (#24034, @kriscon-db)
• [Tracking] Make review-queue names unique case-insensitively (defined at table creation) (#24015, @kriscon-db)
• [Tracking / UI] Normalize review-queue add-items ids before the trace-existence check (#24029, @kriscon-db)
• [UI] Scope review-queue permission UX gate to the active workspace (#24031, @kriscon-db)
• [UI] Surface review-queue trace-removal failures and keep the selection on error (#24027, @kriscon-db)
• [UI] Surface assignable-users load error in review-queue pickers (#24020, @kriscon-db)
• [UI] Prefill review answers from the most recent assessment by timestamp (#24026, @kriscon-db)
• [UI] Surface review-queue self-assign failures with an error toast (#24018, @harupy)
• [Evaluation / Tracing] Fix genai.evaluate() dropping dataset expectations and tags with scorers=[] (#23957, @Incheonkirin)
• [UI] Require at least one question when saving review queue settings (#24007, @harupy)
• [UI] Send review-queue schema_ids only when the questions actually change (#24017, @kriscon-db)
• [UI] Block saving a review when a previously-answered question is cleared (#24008, @kriscon-db)
• [UI] Compare review-queue picker usernames case-insensitively (#24014, @kriscon-db)
• [Tracking] Bind review-queue completed_by to the authenticated caller (#24006, @kriscon-db)
• [UI] Fix non-functional JSON/Table toggle in the review queue full-trace explorer (#24005, @kriscon-db)
• [UI] Surface review-queue deletion failures instead of swallowing them (#24004, @harupy)
• [Tracing] Fix TS SDK traces storage when MLflow server uses a local FS artifact root without mlflow-artifacts:// uri schema (#23992, @aaronteo-db)
• [UI] Show minute fidelity in the review-queue "Date added" column (#23993, @kriscon-db)
• [UI] Review queues: show the optional rationale box in the question preview (#23995, @kriscon-db)
• [Tracing] Set model provider in Anthropic autolog so LLM cost is computed (#23972, @B-Step62)
• [Evaluation] Add missing ContextUtilization RAGAS scorer class (#23956, @B-Step62)
• [UI] Refresh per-trace queue membership after adding/removing review-queue items (#23940, @kriscon-db)
• [Gateway] Fix JSON response format for Gemini and Anthropic gateway providers (#23932, @tanghaoji)
• [Tracking] Fix metrics/get-history returning empty results when max_results is omitted (#23917, @Vedant-Agarwal)
• [UI] Auto-select default user queue on Review tab load (#23904, @B-Step62)
• [UI] Require at least one answer before completing a focused review (#23923, @kriscon-db)
• [Tracing / Tracking] Clean up review-queue items and assessment errors when a trace is deleted (#23913, @harupy)
• [Evaluation / Tracing] Support common RETRIEVER chunk content fields (#23867, @sanatb187)
• [Tracing / Tracking] Preserve OTel resource attributes during OTLP trace ingestion (#23829, @TomeHirata)
• [Gateway] Fix AI Gateway SSE large-frame read limit (#23880, @yashjiv15-jazzx)
• [Evaluation] Honor OPENAI_BASE_URL env var in OpenAI provider config (#23862, @B-Step62)
• [Build] @mlflow/XXXX package root points to missing dist/index.js (#23874, @WeichenXu123)
• [Build] Add auth extra for full docker image (#23892, @WeichenXu123)
• [Artifacts] Return 404 for missing Azure blob artifacts (#23832, @feynmanliang)
• [Tracking] Fix _stop_listen_for_spark_activity hanging indefinitely on CLOSE_WAIT socket (#23839, @kishor-rkrishnan)
• [UI] Install Codex/OpenCode skills at .agents/skills (#23847, @harupy)
• [Tracing / Tracking] Fix mlflow.openai.autolog span type resolution for ChatCompletions subclasses (#23759, @harupy)
• [Tracking] Fix mlflow db upgrade on a fresh database (#23752, @harupy)
• [Tracking] Expose workspace on experiment response (#23593, @joshuawong-db)
• [UI] Handle missing clipboard API in insecure HTTP contexts (#23598) (#23601, @srinjoy356)
• [Build / UI] Fix PDF artifact viewer import.meta SyntaxError (#23731, @harupy)
• [Tracking] Fix _parse_extra_conf for HDFS config values containing = (#23730, @copilot-swe-agent)
• [Prompts / UI] Hide experiment kebab on prompt details page (#23661, @harupy)
• [Tracking] Enforce upload artifact size for chunked requests (#23712, @dfgvaetyj3456356-hash)
• [Projects] Reject path traversal in project zip extraction (#23713, @dfgvaetyj3456356-hash)
• [Tracking] Prefer routed ASGI paths in FastAPI auth checks. (#23685, @HumairAK)
• [Tracing / Tracking] Restore mlflow.crewai autolog on crewai 1.14.5 (#23682, @harupy)
• [Tracing] Unwrap JSON-encoded session.id / user.id span attributes on ingest (#23642, @SahilKumar75)
• [Evaluation / Tracing / UI] Forward OpenAI custom base URL in Detect Issues flow (#23650, @harupy)
• [Tracking] Add ON DELETE CASCADE relationship for SqlTraceInfo to SqlExperiment (#23194, @Mytolo)
• [Tracing] Extend mlflow.sourceRun metrics filter to cover post-hoc linked OTLP traces (#23591, @RudraDudhat2509)
• [Tracing] UI does not show Judge costs (#23586, @WeichenXu123)
• [Tracking] [Security] Register auth validator for /ajax-api/3.0/mlflow/get-trace-artifact (#23317, @B-Step62)
• [Tracing] Fix pydantic-ai >= 1.78.0 ToolManager module rename (#23508) (#23528, @kishor-rkrishnan)
• [UI] Add .jsonl artifact previews (#23532, @bvolpato)
• [Tracking] Disable credentialed CORS when wildcard origins are configured (#23178, @B-Step62)
• [Evaluation] Fix judge fallback on event-based traces grading itself (#23445, @james-fletcher-db)

Documentation updates

• [Docs / Evaluation] Add docs page for @mlflow.test pytest regression testing (#24011, @B-Step62)
• [Docs] Fix make_judge doc: self-referential deprecation note and link typo (#24046, @B-Step62)
• [Docs] Add documentation for review queues and label schemas (#23975, @kriscon-db)
• [Docs] Surface mlflow agent setup in docs (#23859, @joshuawong-db)
• [Docs] Document MLFLOW_STATIC_PREFIX behavior change in migration guide (#23851, @Sanket2329)
• [Docs] Add Colab warning in Quickstart Step 4 (#23831, @Farzah11)
• [Docs] Fix undefined generate_response in tracing docs (#23814, @llljjjwww333)
• [Docs / Tracing] Use CLI for Claude Code plugin install in docs (#23679, @harupy)
• [Docs / Models] Deprecate validate_serving_input in favor of mlflow.models.predict (#23376, @B-Step62)
• [Docs] Fix incorrect output comment for best_run.info in tracking docs (#23571, @Aksh123100)

Small bug fixes and documentation updates

#24045, #24042, #24024, #24023, #23969, #23970, #23961, #23964, #23963, #23866, #23729, #23670, #23310, #23294, @B-Step62; #24022, #24019, #23937, #23758, #23737, #23735, #23605, #23579, #23545, #23511, #23526, @aaronteo-db; #24025, #24003, #23996, #23915, #23912, #23882, @kevin-lyn; #23910, #23994, #23990, #23974, #23984, #23934, #23935, #23946, #23938, #23931, #23925, #23921, #23924, #23846, #23926, #23844, #23886, #23887, #23885, #23884, #23878, #23879, #23876, #23875, #23807, #23804, #23801, #23799, #23795, #23604, #23599, #23613, @kriscon-db; #23997, #23834, #23853, #23823, #23780, #23630, #23614, @joshuawong-db; #23947, #23920, #23858, #23848, #23845, #23841, #23840, #23838, #23837, #23803, #23827, #23826, #23824, #23806, #23802, #23798, #23796, #23788, #23595, #23776, #23764, #23745, #23743, #23742, #23740, #23739, #23718, #23711, #23710, #23708, #23700, #23699, #23697, #23684, #23677, #23672, #23671, #23669, #23668, #23667, #23666, #23663, #23653, #23644, #23643, #23639, #23640, #23638, #23636, #23632, #23631, #23626, #23625, #23618, #23606, #23596, #23588, #23585, #23582, #23581, #23580, #23576, #23573, #23567, #23566, #23565, #23563, #23558, #23553, #23552, #23523, #23506, #23498, @harupy; #23893, @debu-sinha; #23722, @kishor-rkrishnan; #23833, #23741, #23732, #23727, @TomeHirata; #23769, @mprahl; #23589, @charlesverge; #23690, @pvelayudhan; #23658, @copilot-swe-agent; #23540, @jamesbraza

MLflow 3.13.0 includes several major features and improvements

Major New Features

🔐 Role-Based Access Control & Admin UI: A full RBAC system with reusable roles and workspace-scoped grants, plus a new web Admin UI for managing users, roles, and permissions on self-hosted MLflow.

🗄️ Trace Retention & Auto Archival: Automatically move aged trace span data out of your SQL backend into object storage (e.g. S3) while keeping every trace fully readable in the UI and APIs.

🤖 One-click observability & governance for coding agents: Onboard Claude Code, OpenAI Codex, or Gemini CLI to the AI Gateway in one click for tracing, usage tracking, budgets, and guardrails.

✨ New engines for MLflow Assistant: Run MLflow Assistant on a local Ollama model, the OpenAI Codex CLI, or any MLflow AI Gateway endpoint, in addition to Claude Code.

☸️ Helm chart for Kubernetes: An official, production-ready Helm chart for deploying the MLflow tracking server to any Kubernetes cluster.

🌐 Hermes Agent support: Route the Hermes Agent runtime through the AI Gateway and capture its end-to-end traces in MLflow over OpenTelemetry.

🪵 Span log levels: Python-logging-style severity levels on spans, with a "Minimum log level" filter in the trace UI to hide low-level noise.

Breaking Changes

The permission system has been overhauled into a unified Role-Based Access Control model. The legacy per-resource permission tables, REST endpoints, and client methods are removed and replaced by roles backed by role_permissions, default_permission now acts as a floor rather than an override, and a workspace USE grant is sufficient to create experiments and registered models. Code that relied on the old per-resource permission APIs must migrate to the new role-based APIs. (#22855, #22859, #22941, #23337, #23379, @PattaraS)

MLServer is no longer available as a pyfunc serving backend. The previously deprecated enable_mlserver option has been removed, so mlflow models serve always uses the built-in scoring server. (#23356, @harupy)

mlflow autolog claude no longer installs the old Python autolog hook; Claude Code tracing is now provided by the official Claude plugin, which must be installed separately. (#23339, @B-Step62)

The default optimizer used by judge.align() is now MemAlign, so existing alignment workflows may produce different judges than before unless an optimizer is passed explicitly. (#23254, @veronicalyu320)

Pointing the tracking or model registry store at a local file-system path now raises an error by default; set MLFLOW_ALLOW_FILE_STORE=true to keep using a file-based store. (#22773, @harupy)

Other Assorted Features & Improvements

• [Gateway] Support AI Gateway as a backend of MLflow Assistant (#23559, @B-Step62)
• [UI] Make admin pickers target the workspace they're granting into (#23543, @PattaraS)
• [UI] Bring direct-grant picker to parity with role picker (#23420, @PattaraS)
• [UI] Cherry-pick: Add OpenAI Codex CLI as assistant provider (#22566) (#23517, @B-Step62)
• [Tracing] Support settings.local.json for Claude Code tracing config (#23285, @Gkrumbach07)
• [UI] Add coding-agent endpoint creation flow in AI Gateway UI (#23430, @TomeHirata)
• [Tracking / UI] Unified per-user permission APIs: grant / revoke / get / list under /mlflow/users/permissions/* (#23247, @PattaraS)
• [Evaluation / Tracing] Add mlflow.genai.test_agent for automated agent stress-testing (#22990, @serena-ruan)
• [Gateway] Add /gateway/proxy/{endpoint_name}/{path} raw proxy endpoint (#23330, @TomeHirata)
• [UI] Add Ollama as assistant provider (#22098, @SuperSonnix71)
• [Model Registry / Tracking] feat(tracking): Add reader/writer instance routing for database replicas (#22910, @ravidarbha)
• [Tracing / Tracking] Add workspace trace archival configuration plumbing (#22164, @HumairAK)
• [Prompts / Tracking] Promote prompt to a first-class RBAC resource_type (#23248, @PattaraS)
• [Tracing] Publish claude code integration as a plugin and add a setup wizard. (#23218, @B-Step62)
• [UI] Forward Usage tab MetricsFilter metric filters to chart View traces links by translating to Traces page URL DSL (#23239, @aaronteo-db)
• [Tracing] Claude Code Agent Typescript (#20414, @joelrobin18)
• [UI] Adding metric filter component for Experiment Usage tab (#23120, @aaronteo-db)
• [Tracing] Add Link entity and LiveSpan.add_link() for OpenTelemetry Span Links (#22797, @khaledsulayman)
• [Docs / Evaluation] Add Google ADK and third-party scorers (#22299, @debu-sinha)
• [UI] Open /admin to workspace managers (scoped per their workspace) (#23086, @PattaraS)
• [Tracing] Trace Runner.run_streamed() in OpenAI Agents SDK autolog (#22962, @ktrk115)
• [Tracking] Feature/sagemaker build network option (#22996, @pdifranc)
• [Docs / Tracing / UI] Add log levels for Trace Spans with UI switch to filter (#23017, @rrtheonlyone)
• [Tracing / UI] Improve trace page empty state onboarding with setup and code blocks (#22533, @vivian-xie-db)
• [Tracing / Tracking] Add UC traces upsell message for set_experiment calls on Databricks (#23038, @xsh310)
• [Server-infra / UI] [Admin-UI-3/4] Add Platform Admin pages (#22929, @PattaraS)
• [UI] [Admin-UI-2/4] Add /account page and bottom-left account widget (#22973, @PattaraS)
• [Build] Add Helm charts for deploying mlflow to kubernetes cluster (#21973, @WeichenXu123)
• [Tracking] Fix Databricks unified auth support when MLFLOW_ENABLE_DB_SDK=true (#20599, @vb-dbrks)
• [Tracking] [Admin-UI-1/4] Add backend auth endpoints (#22928, @PattaraS)
• [Evaluation] Support multiple assessments per trace in MemAlign optimizer (#22846, @veronicalyu320)
• [Docs / Model Registry / Prompts] Include workspace in webhook delivery envelopes when workspaces are enabled (#22873, @copilot-swe-agent)
• [Server-infra] Seed default RBAC roles and grant creator on workspace creation (#22857, @PattaraS)

Bug fixes

• [UI] Warn on submit with an unsaved direct-grant draft (#23612, @PattaraS)
• [Tracing] Clear archive-now requests for non-archivable leftovers (#23655, @HumairAK)
• [Tracking] Forward MLflow client telemetry from inside Databricks workloads (#23483, @smoorjani)
• [UI] Drop the vestigial directPermissions parallel pass; hide synthetic user<id>_ roles on Account/UserDetail (#23578, @PattaraS)
• [Gateway] Tighten response format JSON schema type (#23290, @fenil210)
• [Tracking] fix(tracking): return <console> for mlflow.source.name when sys.argv[0] is empty (#23352, @xodn348)
• [Scoring] Fix UnicodeEncodeError on artifact download with non-ASCII filename (#23241, @1fanwang)
• [Artifacts / UI] Preserve pdfjs-dist bundles in webpack build (craco.config.js) (#23349, @B-Step62)
• [Tracking] Improve misleading DB SDK auth error (#23374, @B-Step62)
• [Tracing] Make mlflow.get_trace V4 retry policy configurable (#23443, @artjen)
• [Tracking] Release _post_import_hooks_lock before firing hooks (#23466, @harupy)
• [Model Registry / Prompts] RBAC: extend prompt resource_type to after-request handlers (#23426, @PattaraS)
• [Evaluation] Surface mlflow version mismatch when deserializing scorers (#23215, @smoorjani)
• [Tracing] Ship compiled dist/ in @mlflow/mlflow-openclaw so openclaw plugins install works (#23220, @B-Step62)
• [Tracing] Fix ended LiveSpan state mutation (#23152, @SahilKumar75)
• [Gateway] Fix AmazonBedrockProvider._build_converse_kwargs tool-call history and validation for Bedrock Converse (#23223, @copilot-swe-agent)
• [UI] fenil-fix: experiment name error (#23199, @fenil210)
• [Tracking] Add workspace isolation on scorers when creating a guardrail (#23115, @mprahl)
• [Evaluation] [Security] Add MLFLOW_ALLOW_PICKLE_DESERIALIZATION guard to PickleEvaluationArtifact (#23183, @TomeHirata)
• [UI] Fix getExperimentNameValidator showing incorrect "deleted state" error for active experiments (#23169, @copilot-swe-agent)
• [Tracking] Fix runs:/<run_id>/<model_name> loading by resolving logged-model artifacts via models:/<model_id> (#23130, @copilot-swe-agent)
• [Gateway] Fix Vertex AI gateway to use Anthropic API format for Claude models (#23175, @TomeHirata)
• [Tracing] Fix invalid stop-hook command when using pixi environment manager (#23030, @copilot-swe-agent)
• [Gateway] Fix MySQL-incompatible NULLS LAST syntax in list_endpoint_guardrail_configs (#23168, @copilot-swe-agent)
• [Evaluation] gateway: honor Anthropic api_base from secret auth_config (#23167, @sairavuri-sudo)
• [Gateway] Fix nested array items being stripped from function tool schemas (#23053, @shyamspr)
• [Tracing / Tracking] Fix OTLP trace ingestion: double-encoded request ID and missing trace tags (#23067, @sairavuri-sudo)
• [Gateway] Add per-image/video/audio pricing to amazon.nova-2-multimodal-embeddings-v1:0 in Bedrock catalog (#23117, @copilot-swe-agent)
• [Evaluation] Skip re-alignment of unchanged traces in MemAlignOptimizer (#23008, @veronicalyu320)
• [Tracing] Fix trace API authorization vulnerability (#23014, @TomeHirata)
• [Models] Fix sentence_transformers pyfunc predict for v5.4+ (#23108, @harupy)
• [Tracing / UI] Fix grouped trace session counts in the UI (#23012, @lavaFreak)
• [Tracing] Fix Azure OpenAI streaming usage tracing (#23036, @Genmin)
• [Tracing] Add session_count trace metric for grouped traces (#23011, @lavaFreak)
• [Evaluation] Fix gateway_adapter not forwarding workspace header to judge endpoints (#23047, @sairavuri-sudo)
• [Scoring] Use TaskContext.artifactDir to get the correct unpacked artifacts directory (#22969, @WeichenXu123)
• [Models / Tracking] Add MLFLOW_SKIP_PIP_REQUIREMENTS_CHECK env var to bypass pip validation in air-gapped environments (#22920, @copilot-swe-agent)
• [Tracking] Aggregate role-based grants in workspace-level permission checks (#22954, @PattaraS)
• [UI] Fix unclickable "View logs for this period" link in ScrollableTooltip when many data series are shown (#22917, @copilot-swe-agent)
• [Tracking] Fix delete_user FK constraint failure when user has dependent rows (#22922, @PattaraS)
• [Tracing] Preserve cache_read tokens in @mlflow/claude-code TypeScript plugin for cache observability (#22906, @dgokeeffe)
• [Tracing] Add OpenClaw tracing plugin (#22717, @B-Step62)
• [UI] Fix uncaught rejection in CreateBudgetPolicyModal submit (#22903, @PattaraS)

Documentation updates

• [Tracing] Clarify trace archival max-traces behavior. (#23656, @HumairAK)
• [Docs] docs: fix admonition rendering broken by Docusaurus 3.10 (#23635, @B-Step62)
• [Docs / Tracing] Refresh Claude Code tracing docs and split Claude Agent SDK page (#23633, @B-Step62)
• [Tracing] Document trace archival setup and behavior. (#23371, @HumairAK)
• [Docs] docs: add Role-Based Access Control page; refresh permissions docs (#23133, @PattaraS)
• [Docs] Update codex.mdx to use Codex openai_base_url config (CLI + ~/.codex/config.toml) (#23272, @copilot-swe-agent)
• [Docs] Fix ANTHROPIC_BASE_URL example in Claude Code gateway docs (#23269, @copilot-swe-agent)
• [Docs] Add trace integrations docs for n8n and Openweb UI (#23249, @WeichenXu123)
• [Docs / Tracing] Docs: add Hermes Agent AI Gateway and tracing guides (#23216, @B-Step62)
• [Docs] Fix webhook auth endpoint placeholders in basic-http-auth.mdx (#23198, @copilot-swe-agent)

Small bug fixes and documentation updates

#23637, #23405, #23181, #23673, #23659, #23293, #23440, #23441, #23127, #23182, #23179, #23180, #22992, @B-Step62; #23557, @AayushShah-904; #23594, #23592, #23583, #23496, #23417, #23415, #23414, #23413, #23412, #23410, #23409, #23408, #23407, #23406, #23399, #23398, #23452, #22861, #22933, #22888, @PattaraS; #23531, #23282, #23281, #23280, #23279, #23278, #23277, #23276, #23275, #23274, #23173, #23189, #23174, #23171, @TomeHirata; #23495, #23447, #23448, #23422, #23360, #23322, #23316, #23313, #23085, #23304, #23268, #23084, @kriscon-db; #23494, #23493, #23489, #23473, #23470, #23468, #23467, #23463, #23461, #23459, #23462, #23457, #23454, #23450, #23449, #23427, #23428, #23424, #23411, #23395, #23391, #23390, #23388, #23387, #23386, #23385, #23383, #23380, #23381, #23373, #23372, #23361, #23357, #23354, #23351, #23348, #23343, #23342, #23331, #23321, #23320, #23318, #23315, #23307, #23297, #23288, #23286, #23283, #23262, #23264, #23256, #23260, #23255, #23240, #23228, #23234, #23233, #23230, #23227, #23207, #23203, #23206, #23197, #23187, #23185, #23177, #23166, #23157, #23156, #23154, #23155, #23153, #23149, #23148, #23143, #23142, #23141, #23140, #23135, #23134, #23132, #23131, #23129, #23126, #23123, #23122, #23118, #23112, #23110, #23107, #23105, #23093, #23090, #23010, #23088, #22999, #22998, #22988, #22989, #22987, #22986, #22981, #22975, #22960, #22958, #22602, #22937, #22923, #22912, #22907, #22908, #22898, #22894, #22893, #22892, #22889, #22887, #22886, @harupy; #23423, #23464, #23287, @aaronteo-db; #23368, @ynachiket; #23366, @mprahl; #23363, #23333, #23136, #23005, #22777, #22934, #22605, #22497, #22256, @HumairAK; #22832, @james-fletcher-db; #23214, @SomtochiUmeh; #23224, #23103, #23101, #22959, #22924, @copilot-swe-agent; #22977, @4binas; #22961, #22955, @serena-ruan; #22865, @iis-MarkKuang; #22799, @artjen

MLflow 3.13.0 is a major release for running AI observability at scale, focused on access control, the lifecycle of your trace data, and richer support for agents. Highlights include a full Role-Based Access Control system with a new Admin UI, automatic trace archival, one-click onboarding for coding agents, new engines for MLflow Assistant, span log levels, an official Helm chart for Kubernetes, and Hermes Agent support.

1. Role-Based Access Control and Admin UI

Sharing a self-hosted MLflow server across a team used to mean granting permissions one resource at a time, with no central place to manage them. A new Role-Based Access Control (RBAC) system replaces that: define roles as reusable bundles of permissions, assign them to users, and let workspace-level grants express both membership and admin authority. A user's effective access is the union of their roles, and experiments, models, prompts, scorers, and AI Gateway endpoints are all covered.

A new web Admin UI makes this manageable without touching REST endpoints, with a self-service /account page for viewing your roles and changing your password, and Platform Admin pages for managing users, roles, and grants. Just start mlflow server with authentication enabled.

2. Trace Retention and Auto Archival

On a long-running tracking server, trace span data piles up in your SQL backend and eventually slows it down. Trace archival keeps it in check: a background pass automatically moves traces older than your retention window out of SQL and into cheap object storage such as S3, while keeping every trace fully readable in the UI and through the APIs. Retention is policy-driven, resolving from server to workspace to experiment.

Enable it by pointing the server at a YAML config:

trace_archival:
  enabled: true
  location: s3://my-bucket/trace-archive
  retention: 7d
  interval_seconds: 60

Workspace managers and experiment owners can then tighten retention from the UI or CLI, and the effective cutoff shows up as an "Archive after" badge on each trace.

3. One-Click Observability and Governance for Coding Agents

Putting a coding agent like Claude Code, OpenAI Codex, or Gemini CLI under observability and governance used to require wiring up a gateway endpoint by hand. The AI Gateway QuickStart now does it in one click: pick your agent and MLflow provisions a pre-configured endpoint (no API key needed, since the agent brings its own credentials) and hands you a ready-to-paste starter snippet. From then on, every request the agent makes is captured as a trace and subject to usage tracking, budgets, and guardrails.

4. New Engines for MLflow Assistant

MLflow Assistant launched in 3.9.0 tied to Claude Code. You can now choose the engine that powers it: run the Assistant on a local Ollama model, the OpenAI Codex CLI, or any MLflow AI Gateway endpoint, all selectable from the setup wizard. That means you can keep everything on your own machine with Ollama, or reuse a provider you already route through the Gateway.

5. Helm Chart for Kubernetes Deployment

Deploying MLflow on Kubernetes used to mean writing and maintaining your own manifests. An official, production-ready Helm chart now does it for you, with TLS, persistent storage, Ingress, Prometheus metrics, a restrictive NetworkPolicy, RBAC, and optional mlflow gc garbage collection built in. Download the chart and install it (requires Kubernetes 1.23+ and Helm 3.8+):

helm install mlflow ./charts \
--namespace mlflow \
--create-namespace \
--set storage.enabled=true \
--set mlflow.backendStoreUri="sqlite:////mlflow/mlflow.db" \
--set mlflow.artifactsDestination="/mlflow/artifacts"

6. Hermes Agent Support

Agent observability now reaches beyond coding agents to long-running autonomous runtimes. Hermes Agent from Nous Research integrates on two surfaces: route its model calls through the AI Gateway for centralized usage tracking, budgets, and guardrails, and capture full end-to-end traces, including LLM calls, tool invocations, and long-running sessions, through MLflow Tracing over OpenTelemetry.

To route Hermes through the Gateway, create an endpoint and run hermes setup model; for tracing, install the community hermes-otel plugin to export OTLP traces to your MLflow server.

7. Log Levels for Trace Spans

A busy trace can bury the spans you care about under chain plumbing and parser calls. Spans now carry Python-logging-style severity levels (DEBUG through CRITICAL), assigned automatically from the span type, so LLM, tool, and retriever calls surface as INFO while internal steps stay DEBUG, and any span that raises is promoted to ERROR. In the trace explorer, a new Minimum log level slider hides everything below the threshold, with no code changes required. You can also set a level explicitly:

import mlflow
with mlflow.start_span("plumbing", log_level="DEBUG") as span:
    ...

Full Changelog

For a comprehensive list of changes, see the release change log.

What's Next

Get Started

Upgrade to try these new features:

pip install mlflow==3.13.0

Share Your Feedback

We'd love to hear about your experience with these new features:

• GitHub Issues - Report bugs or request features
• MLflow Roadmap - See what's coming next and share your ideas
• ⭐ Star us on GitHub - Show your support for the project

Learn More

Check out the MLflow documentation for detailed guides

We're excited to announce MLflow 3.13.0rc0, which deepens agent observability, tightens permissions, and broadens deployment options:

Major New Features

RBAC + Admin UI: Major overhaul of MLflow's Role-Based Access Control — legacy per-resource permission tables collapsed into role_permissions, unified per-user permission APIs under /mlflow/users/permissions/*, workspace USE permission lets users create experiments and registered models, default roles are seeded on workspace creation, prompt is promoted to a first-class RBAC resource_type, and a new 4-page Admin UI (account widget, /account page, Platform Admin pages, backend auth endpoints) opens to workspace managers scoped per their workspace. (#22855, #22857, #22859, #22928, #22929, #22941, #22973, #23086, #23247, #23248, #23337, #23379, @PattaraS)

Coding-Agent Tracing as Plugins: Claude Code, OpenClaw, Ollama, and OpenAI Codex are now wired into the AI Gateway as first-class assistant providers, plus a Claude Code TypeScript plugin with a setup wizard and settings.local.json support. The legacy Python autolog hook for mlflow autolog claude is replaced by the new official plugin, and a coding-agent endpoint creation flow is now available directly in the AI Gateway UI. (#20414, #22098, #22566, #22717, #23218, #23285, #23339, #23430, #23517, @B-Step62, @joelrobin18, @Gkrumbach07, @SuperSonnix71, @TomeHirata)

Trace Archival: End-to-end trace archival across the tracking stack. Includes archival configuration models, OTLP and artifact helpers, SQLAlchemy archival passes, archive-aware retrieval fallback, plus workspace/experiment/server-level archival settings in the UI. Read archived traces back seamlessly. (#23359, @mprahl)

Helm Charts for Kubernetes Deployment: First-class Helm chart for deploying MLflow to Kubernetes clusters — production-ready configuration, ingress, persistence, and appVersion wired to the released MLflow image. Get from helm install to a running tracking server without writing your own manifests. (#21973, @WeichenXu123)

mlflow.genai.test_agent for Automated Agent Stress-Testing: New API for stress-testing GenAI agents — generate adversarial inputs, replay them through your agent, and review the resulting traces in MLflow. Wires into the existing evaluation flow and assessment APIs. (#22990, @serena-ruan)

OpenTelemetry Span Links: Tracing now supports the OpenTelemetry Link entity via LiveSpan.add_link(), letting you connect causally related spans across traces. (#22797, @khaledsulayman)

Database Replica Routing: The SQL tracking store now supports reader/writer instance routing for database replicas, so read-heavy MLflow deployments can scale horizontally without overloading the primary. (#22910, @ravidarbha)

Stay tuned for the full release, which will include even more features and bug fixes.

To try out this release candidate, please run:

pip install mlflow==3.13.0rc0