Model Registry

Register, manage, version, and deploy machine learning models. The model registry provides a centralized catalog for tracking models across your organization, regardless of their source or framework.

All endpoints require authentication via X-API-Key header and the appropriate scope.

---

RegisteredModel Object

{
  "_id": "model_abc123",
  "name": "Churn Prediction XGBoost",
  "description": "XGBoost model for customer churn prediction trained on Q4 2024 data",
  "framework": "xgboost",
  "algorithm": "XGBClassifier",
  "tags": ["production", "churn", "xgboost"],
  "source": "automl",
  "workspaceId": "ws_001",
  "organizationId": "org_xyz",
  "artifact": { "s3Key": "model-artifacts/org_xyz/upload_abc123.zip", "sizeMb": 12.5 },
  "features": { "featureNames": ["tenure", "monthly_spend", "support_tickets", "plan_type"], "targetName": "churned" },
  "training": { "problemType": "classification" },
  "metrics": { "classification": { "accuracy": 0.94 } },
  "deployment": {
    "status": "running",
    "resources": { "cpu": "2", "memory": "4Gi", "replicas": 1 },
    "deployedAt": "2025-02-01T10:00:00Z"
  },
  "access": { "owner": "user_456", "isPublic": false, "sharedWith": [] },
  "versions": [
    {
      "version": 1,
      "description": "Initial version",
      "artifact": { "s3Key": "model-artifacts/org_xyz/upload_abc123.zip", "sizeMb": 12.5 },
      "metrics": { "classification": { "accuracy": 0.94 } },
      "deployment": { "status": "registered" },
      "createdAt": "2025-01-15T10:30:00Z",
      "createdBy": "user_456"
    }
  ],
  "activeVersion": 1,
  "monitoring": { "recordInputs": true, "labelWindowDays": 14 },
  "latestDrift": { "resultId": "6ab95770...", "overallStatus": "ok", "driftScore": 0.04, "sampleSize": 1250, "calculatedAt": "2026-09-27T00:05:12Z" },
  "latestAnalysis": { "jobId": "6ab95740...", "status": "completed", "triggeredBy": "scheduler", "startedAt": "2026-09-27T00:00:03Z", "updatedAt": "2026-09-27T00:05:12Z", "finishedAt": "2026-09-27T00:05:12Z" },
  "createdAt": "2025-01-15T10:30:00Z",
  "updatedAt": "2025-02-01T10:00:00Z"
}

Field	Description
source	automl, upload, experiment or external
deployment	The serving state: status (registered, building, deploying, starting, running, idle, deployed, stopped, failed), resources, and deployedAt once deployed
access	owner (the owner's user ID), isPublic and sharedWith
versions	Every version: its artifact, metrics, deployment and who created it when
activeVersion	The version the model serves, and the one drift analyzes
monitoring	recordInputs (keep each prediction's inputs and whole output; default on) and labelWindowDays (how long an actual keyed by entity id may take to arrive; default 14)
latestDrift	The latest drift result's summary (driftScore is the mean PSI)
latestAnalysis	The latest drift analysis run: status (pending, deploying, running, completed, failed), triggeredBy (manual, scheduler, baseline), statusMessage, and error saying why it produced no result

---

GET /api/v1/mlops/models

List all registered models accessible to the authenticated user.

Scope: model-registry:read

Query Parameters

Parameter	Type	Required	Description
q	string	No	Search by model name or description
framework	string	No	Filter by framework: pytorch, tensorflow, sklearn, xgboost, lightgbm, onnx, custom, strongly-automl
source	string	No	Filter by source: automl, upload, experiment, external
status	string	No	Filter by deployment status (deployment.status): registered, building, deploying, starting, running, idle, deployed, stopped, failed
tag	string	No	Filter by tag
workspaceId	string	No	Filter by workspace ID
limit	integer	No	Number of results to return (default: 50, max: 200)
cursor	string	No	meta.nextCursor of the previous page; omit for the first page
sort	string	No	Sort field, prefixed with - for descending (default: -createdAt)

Response 200 OK

{
  "data": [
    {
      "_id": "model_abc123",
      "name": "Churn Prediction XGBoost",
      "description": "XGBoost model for customer churn prediction trained on Q4 2024 data",
      "framework": "xgboost",
      "tags": ["production", "churn", "xgboost"],
      "source": "automl",
      "workspaceId": "ws_001",
      "organizationId": "org_xyz",
      "training": { "problemType": "classification" },
      "deployment": { "status": "running" },
      "access": { "owner": "user_456", "isPublic": false, "sharedWith": [] },
      "activeVersion": 1,
      "createdAt": "2025-01-15T10:30:00Z",
      "updatedAt": "2025-02-01T10:00:00Z"
    }
  ],
  "meta": {
    "total": 34,
    "limit": 50,
    "nextCursor": null,
    "requestId": "req_abc123"
  }
}

List items leave out versions, artifact and the build output; get a model by ID for those.

---

POST /api/v1/mlops/models/uploads

A presigned link to upload a model bundle of up to 5 GB straight to storage. PUT the file to uploadUrl with Content-Length equal to fileSize, then register the model with the returned s3Key.

Scope: model-registry:write

Field	Type	Required	Description
fileName	string	Yes	The bundle's file name, e.g. my-model.zip
fileSize	number	Yes	Its exact size in bytes

Response 201 Created: data is { "uploadUrl": "...", "s3Key": "...", "bucket": "...", "expiresIn": seconds }.

---

POST /api/v1/mlops/models/uploads/validate

Upload a model artifact bundle to S3 and parse its manifest. Routes through the same upload path the UI uses, so manifest extraction, validation, and S3 upload behave identically. The returned s3Key + manifest are what POST /mlops/models (register) expects.

Scope: model-registry:write

Request (multipart/form-data)

Field	Type	Required	Description
file	file	Yes	Bundle binary -- either a .zip containing strongly.manifest.yaml at root (or one level nested), or a single .pkl / .joblib / .pt / .onnx artifact. Streamed up to 500 MB.

Response 201 Created

{
  "data": {
    "s3Key": "model-artifacts/org_xyz/upload_abc123.zip",
    "bucket": "strongly-models",
    "sizeMb": 12.5,
    "manifest": {
      "name": "Churn Prediction XGBoost",
      "framework": "xgboost",
      "version": "1.0.0",
      "entrypoint": "predict.py"
    },
    "manifestRaw": "name: Churn Prediction XGBoost\nframework: xgboost\n..."
  },
  "meta": { "requestId": "req_abc123" }
}

The manifest / manifestRaw fields are only populated when the bundle is a .zip containing a valid strongly.manifest.yaml.

---

POST /api/v1/mlops/models

Register a new model in the registry.

Scope: model-registry:write

Request Body

{
  "name": "Churn Prediction XGBoost",
  "framework": "xgboost",
  "description": "XGBoost model for customer churn prediction trained on Q4 2024 data",
  "source": "automl",
  "tags": ["churn", "xgboost"],
  "workspaceId": "ws_001",
  "problemType": "classification",
  "features": { "featureNames": ["tenure", "monthly_spend", "support_tickets", "plan_type"] }
}

Field	Type	Required	Description
name	string	Yes	Human-readable model name
framework	string	Yes	Model framework: pytorch, tensorflow, xgboost, sklearn, onnx, custom
description	string	No	Model description
source	string	No	Where the model originated: automl, upload, experiment, external (default: external)
tags	string[]	No	Array of tag strings for categorization
workspaceId	string	No	Workspace to register the model in
problemType	string	No	The model's task: classification, regression, multiclass, multilabel, timeseries or other (any other value is refused). Stored at training.problemType; drift monitoring needs classification or regression
features	object	Conditional	{ "featureNames": [...] } -- the training feature columns in input order. Required for tabular frameworks (sklearn, xgboost, lightgbm, strongly-automl) so predictions are labelled for drift analysis
artifact	object	Conditional	{ s3Key, s3Bucket?, sizeMb? } from /model-registry/upload. Required for a servable model
manifest	object	Conditional	The parsed strongly.manifest.yaml a zip bundle's upload returned
frameworkVersion	string	Conditional	The library version a single model file was saved with. Required for a single .joblib / .pkl / .pt file of sklearn, xgboost, lightgbm or pytorch: it is served with exactly that version
pythonVersion	string	Conditional	The Python a single .joblib / .pkl file was saved with. Required for a pickle
schema	object	Conditional	{ "input": { "fields": [{ "name", "type" }] } }: the model's inputs in the order it takes them (type float, int, string or bool). Required for a single model file: inputs are fed in this order
requirements	string[]	No	pip requirements a single model file's pickle imports beyond its framework (e.g. category-encoders==2.6.3), installed in its serving image
environmentVariables	object	No	The model's own environment variables { NAME: value }, set on its serving pod. Names the platform sets (MODEL_*, STRONGLY_*, AWS_*, FRAMEWORK, S3_BUCKET, INPUT_FIELDS, IMAGE_COLUMNS, PREDICT_PATH, HEALTH_PATH, PORT) are refused (400)

A single model file (.joblib, .pkl, .pickle, .pt, .pth, .onnx) is registered only with what serving it needs; a missing piece is refused (400) naming each. Its format decides its framework: a .joblib / .pkl is sklearn, xgboost, lightgbm or custom; a .pt is pytorch (TorchScript); a .onnx is onnx.

Response 201 Created

The new registered model.

---

GET /api/v1/mlops/models/:id

Get a single registered model by ID, including all versions.

Scope: model-registry:read

Path Parameters

Parameter	Type	Required	Description
id	string	Yes	Model ID

Response 200 OK

Returns the full RegisteredModel object.

---

PATCH /api/v1/mlops/models/:id

Update a registered model's metadata.

Scope: model-registry:write

Path Parameters

Parameter	Type	Required	Description
id	string	Yes	Model ID

Request Body

{
  "name": "Churn Prediction XGBoost v2",
  "description": "Updated model description with latest findings",
  "tags": ["production", "churn", "xgboost", "v2"],
  "problemType": "classification"
}

Only the fields sent change.

Field	Type	Required	Description
name	string	No	Model name
description	string	No	Model description
tags	string[]	No	Array of tag strings
problemType	string	No	The model's task: classification, regression, multiclass, multilabel, timeseries or other. Sets training.problemType and keeps the rest of the training record (and sets it on the active version). Drift monitoring needs classification or regression: set it on a model registered without one
training	object	No	The whole training record (replaces it)

Response 200 OK

The updated registered model.

---

DELETE /api/v1/mlops/models/:id

Delete a registered model. Deployed models must be undeployed before deletion. This action is irreversible.

Scope: model-registry:write

Path Parameters

Parameter	Type	Required	Description
id	string	Yes	Model ID

Response 204 No Content

---

POST /api/v1/mlops/models/:id/deploy

Deploy a registered model for inference. Provisions the required infrastructure and creates a prediction endpoint.

Scope: model-registry:write

Path Parameters

Parameter	Type	Required	Description
id	string	Yes	Model ID

Deploys the model's active version. To serve another version, deploy it from Versions.

Request Body

{
  "resources": { "cpu": "2", "memory": "4Gi", "replicas": 1 },
  "autoShutdownMinutes": 15
}

Field	Type	Required	Description
resources	object	No	The one size the model runs at: cpu, memory, gpu, gpuType, replicas (its container requests exactly what it is limited to)
autoShutdownMinutes	number	No	Idle minutes before the model scales to zero; > 0 deploys it on demand (it wakes on its first prediction), 0 or omitted keeps it always on
schedulingMode	string	No	always_on, on_demand or scheduled
scheduleRules	array	With scheduled	The start and stop times: [{ "id", "action": "start" | "stop", "dayOfWeek": [0-6] (0 = Sunday), "time": "HH:MM", "timezone": IANA name }], at least one start and one stop
autoscaling	object	No	The always-on HPA: enabled, minReplicas, maxReplicas, cpuThreshold, memoryThreshold. Only with always_on
useSpot	boolean	No	Run on spot capacity
spotFallback	boolean	No	With useSpot, fall back to on-demand capacity when spot is unavailable (default true); false requires spot
scalingMode	string	No	fixed (default: wakes to resources.replicas) or auto (scales between minReplicas and maxReplicas with live demand)
minReplicas	number	No	The awake minimum for scalingMode auto
maxReplicas	number	No	The awake maximum for scalingMode auto
targetConcurrency	number	No	In-flight predictions per replica that auto sizes for
instanceType	string	No	A node instance type to run on

Response 200 OK

The model, as GET /mlops/models/:id shows it: deployment.status reports the deploy.

---

PATCH /api/v1/mlops/models/:id/serving

Changes how the model is served (its Environment tab) without deploying it. Scheduling and demand scaling apply to a deployed model at once; resources and spot from its next deploy or start.

Scope: model-registry:write

Request Body: any of the deploy fields above, and environmentId (an environment to size the model by, or custom for resources). A strategy's settings are checked together: scheduled needs scheduleRules, on_demand an autoShutdownMinutes of at least 1, scalingMode auto needs on_demand with minReplicas, maxReplicas and targetConcurrency, and autoscaling is the always-on strategy's.

{
  "schedulingMode": "scheduled",
  "scheduleRules": [
    { "id": "weekday-start", "action": "start", "dayOfWeek": [1, 2, 3, 4, 5], "time": "09:00", "timezone": "America/New_York" },
    { "id": "weekday-stop", "action": "stop", "dayOfWeek": [1, 2, 3, 4, 5], "time": "17:00", "timezone": "America/New_York" }
  ],
  "useSpot": true
}

Response 200 OK: the model.

POST /api/v1/mlops/models/:id/stop

Stop a deployed model: its serving pod is removed and it moves to stopped (it reads stopping until then). A stopped model keeps its versions, monitoring and history; start it again, or deploy another version.

Scope: model-registry:write

Response 200 OK: the model, as GET /mlops/models/:id shows it, deployment.status stopping.

---

POST /api/v1/mlops/models/:id/start

Start a stopped model at the size it was deployed with. It is starting until its pod is ready, then running; a start that cannot complete (its image cannot be pulled, its container cannot start) leaves it failed with the reason in its deployment logs, scaled back to zero. Poll the model's deployment.status.

Scope: model-registry:write

Response 200 OK: the model, as GET /mlops/models/:id shows it, deployment.status starting.

---

GET /api/v1/mlops/models/:id/status

The model's lifecycle: status (its deployment's: registered when never deployed, building, deploying, starting, running, idle, waking, stopped or failed), canServe (whether predictions can be made now), activeVersion (the version that serves), versionCount and updatedAt.

Scope: model-registry:read

---

GET /api/v1/mlops/models/:id/logs

The model's image build logs, or with logType=runtime its serving container's logs. A model that is not running has no container: asked for runtime logs, it answers its build logs with type build, requested runtime and a warning saying why.

Scope: model-registry:read

Query	Type	Required	Description
logType	string	No	build (default) or runtime

Response 200 OK: data is { "logs": [...], "type": "build" | "runtime", "total": n }

---

POST /api/v1/mlops/models/:id/predict

Run inference on a deployed model (deployment status running). The request is proxied to the model's pod through the AI Gateway, since model pods are internal to the cluster. Every prediction is recorded with the model version that made it.

Scope: model-registry:write

Path Parameters

Parameter	Type	Required	Description
id	string	Yes	Model ID

Request Body

{
  "inputData": [[3.1, 4.2], [8.0, 9.1]],
  "entityId": "order-1042",
  "source": "production"
}

inputData is the feature payload forwarded verbatim to the model's serving endpoint. For the built-in tabular server pass the feature rows directly (a list of rows as above), a single { "feature": value, ... } row object, or { "instances": [ ...rows ] }. Do not wrap the rows in { "features": ... } -- the built-in server does not recognize that shape.

Field	Type	Required	Description
inputData	array | object	Yes	Feature rows forwarded to the model's serving endpoint (rows array, single row object, or {"instances": rows})
entityId	string	No	Your id for what is predicted; actuals can join on it
source	string	No	production (default) or test; test predictions are kept out of drift and A/B results
path	string	No	The route on the model to call when it serves more than one (for example parse_text); omit it for the model's default

Response 200 OK

{
  "data": {
    "predictions": [1, 0],
    "probabilities": [[0.18, 0.82], [0.74, 0.26]],
    "latencyMs": 12,
    "predictionId": "pZ8kQ2",
    "entityId": "order-1042"
  },
  "meta": { "requestId": "req_abc123" }
}

predictions holds one entry per input row. probabilities is present for classifiers that expose predict_proba (one probability vector per row); it is omitted for regressors. A single-row request also returns single-element arrays.

Keep predictionId (or your entityId) to record the prediction's actual later with POST /mlops/models/:id/actuals. Returns 409 if the model is not deployed.

Versions

A registered model accumulates immutable versions. Version 1 is created at registration; add a new version from a freshly uploaded artifact, then choose which one is served.

GET /api/v1/mlops/models/:id/versions

List the model's version history and which one is active.

{
  "data": {
    "versions": [
      { "version": 1, "description": "Initial version", "artifact": { "s3Key": "..." }, "createdAt": "..." },
      { "version": 2, "description": "retrained on more data", "artifact": { "s3Key": "..." }, "createdAt": "..." }
    ],
    "activeVersion": 2
  },
  "meta": { "requestId": "req_abc123" }
}

POST /api/v1/mlops/models/:id/versions

Add a new version. Upload the artifact first via POST /mlops/models/uploads/validate, then pass its returned reference.

Body:

{
  "artifact": {
    "s3Key": "uploads/model-v2.joblib", "s3Bucket": "...", "sizeMb": 4.2,
    "frameworkVersion": "1.6.1", "pythonVersion": "3.11.9"
  },
  "description": "retrained on more data"
}

Returns 201 Created with the new version, as the model's versions list it. artifact.s3Key is required. A single model file's version carries, on its artifact, the frameworkVersion and pythonVersion it was saved with and any requirements, as registering one does; schema gives its inputs when they differ from the model's.

POST /api/v1/mlops/models/:modelId/versions/:id/activate

Point the model's served artifact/metrics at :version without deploying. data is the model, as GET /mlops/models/:id shows it, its activeVersion the version.

POST /api/v1/mlops/models/:modelId/versions/:id/deploy

Set :version active and (re)deploy the serving pod with that version's artifact. Rolling back is just deploying an earlier version again.

Model Evaluation

Monitoring settings, actuals (the true outcomes of predictions) and drift baselines. Drift analyses, results and settings are in the Drift Detection API. Files of any size (actuals, reference rows) are uploaded straight to storage: request an upload link, PUT the file to it with the returned contentType, then start the job that reads it.

Response 200 OK: the model, as GET /mlops/models/:id shows it: deployment.status reports the deploy.

PATCH /api/v1/mlops/models/:id/monitoring

Change the model's monitoring settings. They apply from its next prediction.

Scope: model-registry:write

Field	Type	Required	Description
recordInputs	boolean	No	Keep each prediction's inputs and whole output. Turn it off for sensitive data: predictions are still recorded, but input drift cannot be measured.
labelWindowDays	number	No	How long an actual keyed by entity id may take to arrive and still label that entity's predictions: 1 to 90 days

Send at least one. Response 200 OK: the updated registered model.

POST /api/v1/mlops/models/:id/actuals

Record actuals for the model's predictions.

Scope: model-registry:write

{
  "actuals": [
    { "predictionId": "pZ8kQ2", "actual": "churn" },
    { "entityId": "order-1042", "actual": 1, "actualAt": "2026-09-20T12:00:00Z" }
  ]
}

Each actual is keyed by predictionId (the id a prediction returned) or entityId (your id sent with the prediction). An entity id actual labels that entity's predictions made within the model's labelWindowDays before actualAt (else its upload time); either may arrive first. A second actual for the same key replaces the first.

Response 200 OK

{
  "data": { "recorded": 1, "replaced": 1, "rejected": [], "batchId": "b_91c2" },
  "meta": { "requestId": "req_abc123" }
}

rejected lists each row that could not be recorded: { "row": 0, "reason": "..." }.

POST /api/v1/mlops/models/:id/actuals/uploads

A link to upload a CSV of actuals of any size: a predictionId or entityId column, an actual column, and optionally actualAt.

Scope: model-registry:write | Body: { "filename": "actuals.csv" }

Response 200 OK: data is { "uploadId": "...", "uploadUrl": "https://...", "contentType": "text/csv" }

POST /api/v1/mlops/models/:id/actuals/imports

Import a CSV of actuals: an upload, or a file on a shared volume you can read. A job reads it.

Scope: model-registry:write

{ "source": { "type": "upload", "uploadId": "...", "filename": "actuals.csv" } }

or { "source": { "type": "volume", "volumeId": "...", "path": "outcomes/actuals.csv" } }.

Response 201 Created: the import, with status pending.

GET /api/v1/mlops/models/:id/actuals/imports

The model's actuals imports, newest first: status (pending, running, completed, failed), rowsRead, recorded, replaced, rejected, and errorMessage when one failed.

Scope: model-registry:read | Sort: -createdAt (default), status, rowsRead, recorded, rejected

GET /api/v1/mlops/models/:modelId/actuals/imports/:id/rejections

The rows an import could not record, each with its row (line in the file) and reason.

Scope: model-registry:read

POST /api/v1/mlops/models/:id/baselines/uploads

A link to upload a CSV of a version's reference rows (the data it was trained or validated on) of any size.

Scope: model-registry:write | Body: { "filename": "reference.csv" }

POST /api/v1/mlops/models/:id/baselines

Build a version's drift baseline from its reference rows: a column per input feature the model declares, and optionally actual, prediction and confidence (to record the version's accuracy, and estimate accuracy before actuals arrive).

Scope: model-registry:write

{
  "source": { "type": "upload", "uploadId": "...", "filename": "reference.csv" },
  "version": 2,
  "analyzeDaily": true
}

Field	Type	Required	Description
source	object	Yes	An upload (type: "upload", uploadId, filename) or a volume file (type: "volume", volumeId, path)
version	number	No	The version the rows describe (default: the active version). A model registered without versions has one baseline for the model: leave version out
analyzeDaily	boolean	No	Turn on the model's daily drift analysis once the baseline is kept (default false); the baseline's first analysis counts as that day's run

Response 201 Created: the job, with status pending. Once built, the baseline becomes the version's active one (the one it replaces is kept, inactive) and the version's first drift analysis starts, when the version has production predictions from the last 7 days (else the job's driftWaiting says so).

GET /api/v1/mlops/models/:id/baselines/jobs

The model's baseline jobs, newest first: status (pending, running, completed, failed), modelVersion, sampleSize, baselineId, analyzeDaily, driftJobId (the first analysis), driftWaiting, driftError, driftScheduleError, and errorMessage when one failed.

Scope: model-registry:read | Sort: -createdAt (default), status, modelVersion

POST /api/v1/mlops/models/:modelId/baselines/:id/activate

Make an earlier baseline its version's active one again; drift compares the version with it from the next analysis on.

Scope: model-registry:write

Response 200 OK: data is { "baselineId": "...", "modelVersion": 2, "isActive": true }

Usage analytics

GET /api/v1/mlops/models/:id/analytics

The model's usage over 7, 30 or 90 days, as its Analytics tab shows it: summary (predictions, errors, errorRate, callers, latencyAvgMs, latencyP50Ms, latencyP95Ms, avgConfidence, lastPredictionAt), daily (predictions, errors and latency per day), sources, versions and A/B variants (predictions and errors of each, with the test's abTestName and abTestDeleted when it has been deleted since), and weekdayHour (predictions by weekday, Monday first, and UTC hour).

Scope: ml-workbench:read

Query	Type	Required	Description
range	string	No	7d, 30d (default) or 90d

GET /api/v1/mlops/models/:id/analytics/callers

Who called the model over the range, one row per caller and what the call came through (via: app, workflow, workspace or direct, with viaName): name, email, predictions, errors, lastSeen. Paged.

Scope: ml-workbench:read | Sort: -predictions (default), errors, lastSeen, name

Query	Type	Required	Description
range	string	No	7d, 30d (default) or 90d
q	string	No	By name, email, app or workflow
limit, cursor	number	No	Paging