Nimbus Studio & Nimbus SDK: June 2026 Product Updates
June was a big month for Nimbus Studio and the Nimbus SDK.
Building on May’s update, we shipped in-session adaptive learning for Bayesian classifier pipelines, added zero-calibration deploy apps for focus, relaxation, and blinks, introduced a prompt-to-pipeline creation flow in the hub, and brought the desktop app to Windows for the first time.
The SDK release behind the headline feature is nimbus-bci==0.4.2.
Nimbus SDK 0.4.2: adaptive learning during live sessions
The biggest SDK-side update is Adaptive Deploy.
Bayesian classifier pipelines can now update themselves while a deploy session is running. Instead of treating a trained model as fixed for the entire session, Nimbus can use live feedback and game outcomes to improve the classifier as the session progresses.
In practice, this means the system can become more accurate as the user interacts with it.
What changed in the SDK
Every Nimbus SDK Bayesian head now supports in-session partial_fit:
NimbusLDANimbusQDANimbusSoftmaxNimbusSTS
Each head can update its posterior on labeled feedback during a live session.
The SDK's existing OnlinePerformanceTracker maintains a rolling accuracy window per sample. Adaptive updates are gated: if a candidate update would reduce accuracy, the batch is rejected and its buffer is drained instead of being silently retried. (If you’re new to entropy-based abstention, Confidence-Gated BCI: How Entropy and Rejection Policies Keep Decoders Reliable in the Real World covers the core idea.)
Signature parity has been verified across partial_fit, predict, predict_proba, and OnlinePerformanceTracker.update for all four heads in nimbus-bci==0.4.2.
What this looks like in Nimbus Studio
In Studio, Adaptive Deploy adds:
- A live accuracy readout
- An
Adaptive v{N}badge with the accuracy delta whenever a new adaptive revision is created - Explicit feedback controls, including correct / correction buttons
- Game outcomes as ground-truth sources, such as target hits in MI Target Hit
Adaptive Deploy also preserves the core Nimbus principle of what you build is what you deploy.
Your frozen, trained pipeline_state.pkl is never mutated. Adaptive changes are saved as session-scoped, versioned checkpoints, so the canonical Build artifact remains intact.
Adaptive Deploy is Pro-only and activates only when the pipeline head supports it.
Zero-calibration deploy apps: focus, relaxation, and blinks
June also added three ready-to-use interactive deploy apps that do not require a trained model:
- Focus Meter — turns live bandpower ratios into a real-time focus signal
- Relaxation Meter — uses the inverse signal for relaxation and meditation feedback
- Blink-to-Click — detects a frontal blink and turns it into a click event
These apps are powered by two new training-free nodes:
bandpower_indicatorfor focus and relaxationblink_detectorfor frontal blink events
They live in a new Indicators palette category, introduced alongside ADR-002 and the new INDICATOR and EVENT BCI data types.
Indicator-only sessions now show live signal and event traces on the Streaming Dashboard. The Focus Meter preset is also featured in the hub.
The important point: this is the same Deploy runtime used for classifier pipelines, but without requiring a classifier.
Hub AI Create: from prompt to pipeline
The projects hub now includes a Make it banner.
You can describe a pipeline in plain language — or choose one of six presets — and Hub AI Create will generate a new project, create a template-backed pipeline, and open it directly in the editor.
The flow includes:
- A typewriter-style prompt for free-text pipeline descriptions
- Curated presets for common starting points
- Optional assistant customization before the graph loads
- A terminal briefing after creation
That terminal briefing explains the template rationale, loaded graph facts such as dataset, paradigm, and preprocessing, and the evaluation or split plan, including CV method or catalog session holdout.
It is now the fastest path from an idea to a runnable BCI pipeline on the canvas.
Lab-grade calibration for researchers
June’s calibration work makes Nimbus Studio more useful for structured data collection, not just demos. For a full walkthrough of the structured flow, see BCI Calibration with Nimbus Studio: From Hardware to Trained Decoder.
The new calibration flow includes:
- Calibration protocols with paradigm and class presets
- Estimated session duration
- Per-trial rejection in the calibration modal
- Timing QC with device sync measured in CI
- A
hardware_smokecheck for post-run hardware timing validation - Session export bundles with
session_report, JSONL event logs, and artifact manifests
The behavioral sidecar now fans into preprocessing, feature, and model nodes, so behavioral events align with the correct downstream processing step.
HDF5 v1.1 event sidecars and the calibration_complete → custom_data.filePath handoff round out the researcher workflow.
Desktop 0.4.0: Windows support arrives
Nimbus Studio desktop reached 0.4.0 in June.
The biggest platform update: Windows x64 is now supported, alongside macOS.
The same pipeline builder, embedded backend, and cloud sync available in the Mac app now run natively on Windows.
Desktop 0.4.0 also includes:
- Support for adaptive learning and indicator apps
- More reliable live sessions
- Smoother reconnection
- Better behavior during long runs
- Fewer stalls during streaming
- More robust sign-in and token handling
- Tighter safeguards around data and project access
If you already have Nimbus Studio installed, the app checks for updates automatically. You can also use Help → Check for Updates, or download the latest installer from the releases page.
Getting started
- Web: studio.nimbusbci.com
- Docs: docs.nimbusbci.com
- Desktop: update in-app or download the latest installer from the releases page
- SDK:
pip install nimbus-bci