API Reference

Complete reference for using the AICR API Server.

Overview

The AICR API Server provides HTTP REST access to recipe generation and bundle creation for GPU-accelerated infrastructure. Use the API for programmatic access to configuration recommendations and deployment artifacts.

Version numbers in the sample requests and responses below (server version, chart versions, driver versions) are illustrative. The authoritative, current versions are in the Component Catalog and the Container Images BOM.

┌──────────────┐      ┌──────────────┐
│ GET /recipe  │─────▶│   Recipe     │
└──────────────┘      └──────────────┘
        │
        ▼
┌──────────────┐      ┌──────────────┐
│ POST /bundle │─────▶│  bundles.zip │
└──────────────┘      └──────────────┘

API vs CLI

• Use the API for remote recipe generation and bundle creation
• Use the CLI for local operations, snapshot capture, and ConfigMap integration

Base URL

Local development (example):

http://localhost:8080

Start the local server:

docker pull ghcr.io/nvidia/aicrd:latest
docker run -p 8080:8080 ghcr.io/nvidia/aicrd:latest

Quick Start

Get a Recipe

Generate an optimized configuration recipe for your environment:

# GET: Basic recipe for H100 on EKS (query parameters)
curl "http://localhost:8080/v1/recipe?accelerator=h100&service=eks"
# GET: Training workload on Ubuntu
curl "http://localhost:8080/v1/recipe?accelerator=h100&service=eks&intent=training&os=ubuntu"
# POST: Recipe from a criteria envelope (YAML body)
curl -X POST "http://localhost:8080/v1/recipe" \
  -H "Content-Type: application/x-yaml" \
  -d 'criteria:
  service: eks
  accelerator: h100
  intent: training'
# Save recipe to file
curl -s "http://localhost:8080/v1/recipe?accelerator=h100&service=eks" -o recipe.json

Generate Bundles

Create deployment bundles from a recipe:

# Pipe recipe directly to bundle endpoint.
# The POST body must be a fully-hydrated RecipeResult; piping GET /v1/recipe
# output (as below) supplies one. Do not hand-author a partial body.
curl -s "http://localhost:8080/v1/recipe?accelerator=h100&service=eks" | \
  curl -X POST "http://localhost:8080/v1/bundle" \
    -H "Content-Type: application/json" -d @- -o bundles.zip
# Extract the bundles
unzip bundles.zip -d ./bundles
# Verify the complete extracted inventory before deployment
(cd ./bundles && aicr verify .)

Endpoints

GET /

Service information and available routes.

curl "http://localhost:8080/"

Response:

{
  "service": "aicrd",
  "version": "v0.14.0",
  "routes": [
    "/v1/recipe", "/v1/query", "/v1/bundle"
  ]
}

GET /v1/recipe

Generate an optimized configuration recipe based on environment parameters.

Query Parameters:

service=rke2 and accelerator=vr200 are Preview. They publish an early-adopter recipe path without the full production support and lifecycle qualification required for Supported status. See the published validation evidence at validation.aicr.run for current coverage.

service=k0s is Preview, covering the single k0s / h200 / ubuntu / training coordinate. See k0s H200 Setup for its prerequisites and known gaps.

Examples:

# Minimal request — at least one criteria dimension is required
curl "http://localhost:8080/v1/recipe?accelerator=h100"
# Specify accelerator with service
curl "http://localhost:8080/v1/recipe?service=eks&accelerator=h100"
# Full specification
curl "http://localhost:8080/v1/recipe?service=eks&accelerator=h100&intent=training&os=ubuntu&nodes=8"
# Using gpu alias
curl "http://localhost:8080/v1/recipe?gpu=gb200&service=eks&os=ubuntu&intent=training"
# Pretty print with jq
curl -s "http://localhost:8080/v1/recipe?accelerator=h100" | jq '.'

POST /v1/recipe

Generate an optimized configuration recipe from a criteria body. This endpoint provides an alternative to query parameters.

Content Types:

• application/json - JSON format
• application/x-yaml - YAML format

An explicit, supported Content-Type is required; a missing, aliased, or unsupported media type is rejected.

GET and POST return identical documents for equivalent criteria, including the criteria object echoed in the response: unspecified dimensions are reported as any on both. HEAD is accepted on /v1/recipe, /v1/query, and /metrics (/health, /ready, and / allow GET only) and returns the same headers without a body — it resolves the recipe to produce them, so it costs what GET costs rather than serving as a cheap probe. Use /health or /ready for liveness.

Request Body:

The request body is a strict envelope. Unknown fields are rejected, so a typo fails loudly instead of being silently ignored:

criteria:
  service: eks
  accelerator: gb200
  os: ubuntu
  intent: training
  platform: kubeflow
  nodes: 8
profile: gpuStack=operator-managed  # optional

profile is optional; omit it to take the composition’s declared default. If you also pass profile= as a query parameter, the two must agree.

The Kubernetes-style RecipeCriteria resource body that this endpoint accepted before v0.21 is no longer supported and returns 400.

Examples:

# POST with YAML body
curl -X POST "http://localhost:8080/v1/recipe" \
  -H "Content-Type: application/x-yaml" \
  -d 'criteria:
  service: eks
  accelerator: h100
  intent: training'
# POST with JSON body
curl -X POST "http://localhost:8080/v1/recipe" \
  -H "Content-Type: application/json" \
  -d '{
    "criteria": {
      "service": "eks",
      "accelerator": "h100",
      "intent": "training"
    }
  }'
# POST with criteria file
curl -X POST "http://localhost:8080/v1/recipe" \
  -H "Content-Type: application/x-yaml" \
  -d @criteria.yaml
# Pretty print response
curl -s -X POST "http://localhost:8080/v1/recipe" \
  -H "Content-Type: application/json" \
  -d '{"criteria":{"service":"eks","accelerator":"h100"}}' \
  | jq '.'

Error Responses:

• 400 Bad Request - No criteria provided: at least one of service, accelerator, intent, os, platform, or nodes must be non-zero. An empty request returns "no criteria provided: specify at least one of service, accelerator, intent, os, platform, nodes". This guard applies to GET /v1/recipe, POST /v1/recipe, GET /v1/query, and POST /v1/query.
• 400 Bad Request - Invalid criteria format, missing required fields, or invalid enum values
• 400 Bad Request - A stated criteria dimension is not honored by any applicable recipe overlay (uncovered dimension). This applies to GET /v1/recipe, POST /v1/recipe, GET /v1/query, and POST /v1/query: every dimension you state (service, accelerator, intent, os, platform) must be matched by at least one applied overlay, or the request fails instead of silently returning a recipe that ignores it. nodes is exempt — it is advisory and never required to be covered. The response’s details.uncovered array names the offending dimension(s), the requested value, and any validCompletions (additional criteria that would make the request coverable). Snapshot-driven resolution (CLI --snapshot / Go SDK) may additionally attach excludedOverlays and constraintWarnings to the error; the HTTP API resolves from criteria only and never emits those two fields.
• 405 Method Not Allowed - Only GET, HEAD, and POST are supported; the response carries Allow: GET, HEAD, POST

Uncovered-Dimension Error Example:

{
  "code": "INVALID_REQUEST",
  "message": "platform 'kubeflow' for criteria(service=eks, accelerator=h100, intent=training, platform=kubeflow) requires os (valid: ubuntu)",
  "details": {
    "uncovered": [
      {
        "dimension": "platform",
        "requestedValue": "kubeflow",
        "validCompletions": [{"os": "ubuntu"}]
      }
    ]
  },
  "requestId": "550e8400-e29b-41d4-a716-446655440000",
  "timestamp": "2025-01-15T10:30:00Z",
  "retryable": false
}

Response:

{
  "apiVersion": "aicr.run/v1",
  "kind": "RecipeResult",
  "metadata": {
    "version": "v0.22.0",
    "appliedOverlays": [
      "base",
      "eks",
      "eks-training",
      "gb200-eks-training"
    ],
    "excludedOverlays": [
      {
        "name": "h100-eks-ubuntu-training",
        "reason": "mixin-constraint-failed"
      }
    ],
    "constraintWarnings": [
      {
        "overlay": "h100-eks-ubuntu-training",
        "constraint": "OS.sysctl./proc/sys/kernel/osrelease",
        "expected": ">= 6.8",
        "actual": "5.15.0",
        "reason": "mixin-constraint-failed: expected >= 6.8, got 5.15.0"
      }
    ]
  },
  "criteria": {
    "service": "eks",
    "accelerator": "gb200",
    "intent": "training",
    "os": "any",
    "platform": "any"
  },
  "constraints": [
    {
      "name": "GPU.driver.version",
      "value": "580.82.07"
    },
    {
      "name": "GPU.driver.cudaVersion",
      "value": "13.1"
    }
  ],
  "componentRefs": [
    {
      "name": "gpu-operator",
      "type": "Helm",
      "chart": "gpu-operator",
      "source": "https://helm.ngc.nvidia.com/nvidia",
      "version": "v25.3.3"
    },
    {
      "name": "network-operator",
      "type": "Helm",
      "chart": "network-operator",
      "source": "https://helm.ngc.nvidia.com/nvidia",
      "version": "v25.4.0"
    }
  ],
  "deploymentOrder": [
    "gpu-operator",
    "network-operator"
  ]
}

metadata.excludedOverlays is optional. When present, each entry includes the overlay name and a machine-readable reason such as constraint-failed or mixin-constraint-failed.

metadata.gpuDriverState is optional and appears only for snapshot-driven recipes. It records the NVIDIA kernel driver state observed on the sampled GPU node — preinstalled or absent — and is omitted when no snapshot was provided or the snapshot carried no usable driver-loaded reading. The bundle-time CheckDriverOwnershipCoherence validation consumes it: a recipe whose snapshot observed no driver (absent) is blocked from bundling with the preinstalled-driver assumption, since that would leave GPU nodes driverless.

metadata.mariaDBOperatorState is optional and appears when a snapshot supplies MariaDB Operator conflict evidence during resolution of AICR-provided Slurm accounting. It records absent, api-detected, crs-detected, or unknown; query-generated recipes and older snapshots without the collector subtype omit the field. Recipe generation remains observational: api-detected, crs-detected, and unknown emit warnings but still produce a recipe. At bundle time, crs-detected and unknown block AICR-provided installation, while api-detected or omitted evidence warns but proceeds; absent proceeds silently.

GET /v1/query

Query a specific value from a fully hydrated recipe. Resolves a recipe from criteria (same parameters as GET /v1/recipe), merges all base, overlay, and inline overrides, then returns the value at the given selector path.

Query Parameters:

All GET /v1/recipe parameters are supported, plus:

Response:

• Scalar values (string, number, bool) are returned as plain JSON values
• Complex values (maps, lists) are returned as JSON objects/arrays

Error Responses:

• 400 Bad Request - No criteria provided: at least one criteria dimension (service, accelerator, intent, os, platform, or nodes) must be non-zero. Returns "no criteria provided: specify at least one of service, accelerator, intent, os, platform, nodes". See the POST /v1/recipe error responses entry for full details.
• 400 Bad Request - Omitting selector returns INVALID_REQUEST. An explicitly empty selector= remains valid and returns the entire hydrated recipe.
• 400 Bad Request - A stated criteria dimension not honored by any applicable overlay (uncovered dimension) — same details.uncovered shape as described in the POST /v1/recipe error responses above.

Examples:

# Get a specific Helm value
curl -s "http://localhost:8080/v1/query?service=eks&accelerator=h100&intent=training&selector=components.gpu-operator.values.driver.version"
# Get deployment order
curl -s "http://localhost:8080/v1/query?service=eks&accelerator=h100&intent=training&selector=deploymentOrder" | jq '.'
# Get a component subtree
curl -s "http://localhost:8080/v1/query?service=eks&accelerator=h100&selector=components.gpu-operator.values.driver" | jq '.'

POST /v1/query

Alternative to GET /v1/query that accepts the criteria and selector in the request body. The body is a strict envelope with a criteria object (the same dimensions accepted as query parameters on GET /v1/query), a required selector string, and an optional profile. Unknown fields are rejected.

Content Types:

• application/json - JSON format
• application/x-yaml - YAML format

Request Body:

criteria:
  service: eks
  accelerator: h100
  intent: training
selector: "components.gpu-operator.values.driver.version"

Examples:

curl -X POST "http://localhost:8080/v1/query" \
  -H "Content-Type: application/json" \
  -d '{
    "criteria": {"service": "eks", "accelerator": "h100", "intent": "training"},
    "selector": "components.gpu-operator.values.driver.version"
  }'

The response format matches GET /v1/query: scalar values are returned as plain JSON values; maps and lists are returned as JSON objects/arrays.

Error Responses:

Same as GET /v1/query — see the GET /v1/query error responses section above. The no-criteria and uncovered-dimension 400 cases apply. selector must be present on both GET and POST; omitting it is a 400. Passing it explicitly empty (selector= on GET, "selector": "" in a POST envelope) is the documented way to ask for the entire hydrated recipe.

Profile and Slurm-accounting endpoints

The v1 in the route and the apiVersion in a recipe document are independent version axes. The route segment versions the HTTP contract; aicr.run/v1 and aicr.run/v1beta2 are the recipe schemas emitted from v0.22. /v1/bundle also still accepts the superseded aicr.run/v1alpha2 and aicr.run/v1alpha3, retired in v1.0.0: aicr.run/v1 for a default recipe and aicr.run/v1beta2 for a profile/configuration recipe, plus versionless legacy artifacts. Selecting a profile or resolving a Slurm accounting mode determines which schema track applies.

There is a single route family. /v1/recipe, /v1/query, and /v1/bundle serve every composition — profiled and unprofiled alike — with one contract. The AKS, GKE, and OKE families are the embedded profile adopters (gpuStack) and need no special routing.

GET /v1/recipe. Accepts the /v1/recipe criteria parameters plus optional profile=name=value, slurmAccountingMode, and gkeTcpxoInterfaces. Profile omission applies the resolved declaration’s required default. Slurm accounting accepts disabled, customer-managed, or aicr-provided; omission defaults a Slurm recipe to disabled. The setting is recorded at configuration.slurm.accounting.mode in an aicr.run/v1beta2 RecipeResult (the superseded aicr.run/v1alpha3 is still read). gkeTcpxoInterfaces carries the ordered eth1=<network>,...,eth8=<network> GPU-NIC Network mapping for the torch-distributed-tcpxo runtime; it is required when the resolved recipe ships that runtime (h100 GKE kubeflow training) and is recorded at configuration.gke.tcpxoInterfaces. The route rejects unknown query parameters and conflicting repeated values.

# AKS, non-default value (omit profile= for the azure-managed default):
curl "http://localhost:8080/v1/recipe?service=aks&accelerator=h100&os=ubuntu&intent=training&profile=gpuStack=operator-managed"
# Slurm with AICR-provided accounting
curl "http://localhost:8080/v1/recipe?service=eks&accelerator=h100&intent=training&os=ubuntu&platform=slurm&slurmAccountingMode=aicr-provided"
# GKE h100 kubeflow training (required TCPXO mapping):
curl "http://localhost:8080/v1/recipe?service=gke&accelerator=h100&os=cos&intent=training&platform=kubeflow&gkeTcpxoInterfaces=eth1=gpu-nic-0,eth2=gpu-nic-1,eth3=gpu-nic-2,eth4=gpu-nic-3,eth5=gpu-nic-4,eth6=gpu-nic-5,eth7=gpu-nic-6,eth8=gpu-nic-7"

POST /v1/recipe. Accepts a strict JSON or YAML envelope. criteria is the plain criteria object, not a RecipeCriteria resource. Profile selection may be supplied in the envelope, as the profile query parameter, or in both places when the values agree. slurmAccountingMode is supplied as the same query parameter used by GET, as is gkeTcpxoInterfaces (required for the GKE h100 kubeflow training family — e.g. gkeTcpxoInterfaces=eth1=gpu-nic-0,...,eth8=gpu-nic-7). Conflicting selections are rejected:

criteria:
  service: aks
  accelerator: h100
  intent: training
profile: gpuStack=azure-managed

Unknown fields, duplicate or trailing documents, malformed selections, and selections against an unprofiled composition fail with 400 INVALID_REQUEST. POST envelopes require Content-Type: application/json or Content-Type: application/x-yaml; missing, aliased, or unsupported media types are rejected.

GET and POST /v1/query. GET accepts the recipe parameters, including slurmAccountingMode and gkeTcpxoInterfaces, plus selector. POST accepts the same strict envelope with a required selector. POST profile selection follows the same query/envelope agreement rule as /v1/recipe:

criteria:
  service: aks
  accelerator: h100
profile: gpuStack=azure-managed
selector: metadata.selectedProfile

POST /v1/bundle. Uses the query parameters and ZIP response documented under POST /v1/bundle below. It carries no profile-selection field because its body is an already-selected RecipeResult. It accepts legacy aicr.run/v1alpha2 and aicr.run/v1 default recipes, including older artifacts that omit apiVersion, and strictly decodes profiled or accounting-configured aicr.run/v1alpha3 and aicr.run/v1beta2 recipes. The request requires Content-Type: application/json or Content-Type: application/x-yaml; missing, aliased, or unsupported media types are rejected.

# The AKS, GKE, and OKE families are the embedded adopters (gpuStack profiles).
# -f stops on an HTTP error so a 4xx/5xx recipe body is never staged and
# an error response is never written to bundles.zip. POSIX sh suffices:
# the commands are sequential (no pipeline), so pipefail is not needed.
set -eu
curl -fsS -o recipe.json \
  "http://localhost:8080/v1/recipe?service=aks&accelerator=h100&os=ubuntu&intent=training&profile=gpuStack=operator-managed"
curl -fsS -X POST "http://localhost:8080/v1/bundle" \
  -H "Content-Type: application/json" -d @recipe.json -o bundles.zip

Profile-bearing responses record metadata.selectedProfile; accounting-aware responses record configuration.slurm.accounting. Both use recipe apiVersion aicr.run/v1beta2 (the superseded aicr.run/v1alpha3 is still read). Their owned paths are immutable across AICR’s supported override surfaces: divergent static values, intersecting dynamic paths, owned-component removal, and argocd-helm install-time values fail closed before output.

A recipe resolved without an explicit profile or slurmAccountingMode uses the default-track response shape. That track is aicr.run/v1 from v0.22, and the schema also still admits the superseded aicr.run/v1alpha2 so a client generated from this spec reads artifacts captured earlier. Profile and Slurm-accounting selection are available on every endpoint; no composition needs special routing.

The AKS, GKE, and OKE families are the embedded profile adopters (gpuStack). Their compositions resolve like any other: omit profile= to take the declared default, or select one explicitly.

# GKE: omit profile= for the gke-default default, or select
# gpuStack=bundle-installer explicitly, then POST to /v1/bundle.
# -f stops on an HTTP error so a 4xx/5xx recipe body is never staged and
# an error response is never written to bundles.zip.
set -euo pipefail
curl -fsS -o recipe.json \
  "http://localhost:8080/v1/recipe?service=gke&accelerator=h100&os=cos&intent=training&profile=gpuStack=bundle-installer"
curl -fsS -X POST "http://localhost:8080/v1/bundle" \
  -H "Content-Type: application/json" -d @recipe.json -o bundles.zip

POST /v1/bundle

Generate deployment bundles from a recipe.

Query Parameters:

Request Body:

The request body is the recipe (RecipeResult) directly. No wrapper object is needed. This release emits apiVersion: aicr.run/v1 or aicr.run/v1beta2 and kind: RecipeResult; its bundle readers additionally accept the superseded aicr.run/v1alpha2 and aicr.run/v1alpha3, respectively. The profile track identifies recipes carrying metadata.selectedProfile, typed configuration.slurm.accounting, or both; profile-bearing artifacts must use /v1/bundle. New clients should preserve the version emitted by recipe resolution.

For backward compatibility, the endpoint also accepts:

• Legacy artifacts that omit apiVersion or kind, or carry them as empty strings after a decode/remarshal round trip.
• The kind: Recipe value this contract published through v0.18.0.

All three shapes reach the bundler identically: the endpoint normalizes kind on ingest, stamping kind: RecipeResult when the request carries an absent, empty, or legacy Recipe kind. The generated bundle’s recipe.yaml therefore always carries the canonical kind and reloads through aicr bundle -r, aicr validate -r, and the tooling that reads a bundle’s recipe.yaml (TestGrid publication, evidence synthesis). Only kind is rewritten — a request that omits apiVersion still produces an artifact with an empty apiVersion, which every reader accepts as the legacy shape.

Any other kind is rejected with a 400, so the endpoint never emits an artifact it would refuse to read back. This matches the /v1/bundle decode path, and the CLI file loader for the same values — aicr bundle -r accepts a RecipeMetadata file as an overlay to hydrate, but as a hydrated RecipeResult artifact it too accepts only RecipeResult or an absent kind. apiVersion is validated separately, as described next.

The shared artifact gate rejects any apiVersion outside aicr.run/v1alpha2, aicr.run/v1, aicr.run/v1alpha3, and aicr.run/v1beta2 with a 400, on this endpoint as well as on the CLI file-load path. An absent or empty apiVersion is still admitted as the legacy shape on RecipeResult inputs through v0.22, and v1.0.0 stops admitting it along with the alpha values. The tolerance is scoped to RecipeResult, which predates the field: a RecipeMetadata overlay is a catalog document however it arrives, so aicr bundle -r and aicr validate -r reject a headerless one exactly as a --data catalog scan does. The reader and emitter clocks are separate: v0.21 and v0.22 both read the alpha values, the target values, and the empty header, while generated recipes carried alpha headers through v0.21 and carry the target values from v0.22 onward. On the CLI file-load path, reading an alpha or headerless artifact logs a deprecation warning naming the file; these endpoints take the artifact as a request body, so there is no file to name and no equivalent signal. See Catalog and binary compatibility for the release-by-release table.

Components

These are the recipe components in recipes/registry.yaml — the names the bundlers query parameter accepts (a request may only name components the recipe declares). The registry is the authoritative source — see the component catalog for detailed descriptions, pinned versions, and per-component caveats.

Examples:

Note: The POST body must be a fully-hydrated RecipeResult — the server adopts the body as-is and does not hydrate registry defaults, so a hand-authored partial body (missing namespace, valuesFile, overrides, dependencyRefs) yields empty values and namespaces in the generated bundle. Obtain a complete body from aicr recipe ... --format json --output - (the CLI defaults to YAML, but POST /v1/bundle JSON-decodes its body) or GET /v1/recipe and pass it unchanged. The inline bodies below are elided for brevity (only a few component fields shown) — use a generated RecipeResult, not these literals.

To bundle a subset of the recipe’s components, use the bundlers query parameter (e.g. ?bundlers=gpu-operator,network-operator) rather than hand-trimming componentRefs — trimming the body silently drops required dependencies and breaks deployers like Helmfile on dangling dependencyRefs. The filter prunes those edges safely (a filtered-out dependency is assumed satisfied externally) and rejects unknown or disabled component names with HTTP 400. Slurm accounting adds required-component checks: customer-managed mode requires slinky-slurm, while AICR-provided mode also requires mariadb-operator-crds, mariadb-operator, and slurm-accounting-mariadb. A bundlers filter that omits any required component is rejected with HTTP 400; required components are not automatically added to the selection.

Enabled Helm refs must reference a deployable primary: an external chart (a source repository plus an effective version — empty, whitespace-only, or a bare v is rejected; the chart name falls back to the component name when chart is unset, but a chart without a source is rejected) or local primary manifestFiles. chart, source, and version values carrying surrounding whitespace are rejected — deployers consume them verbatim. Incoherent refs are rejected with HTTP 400 naming the component. Component ref names must also be unique within a recipe (enabled or disabled refs) when non-empty; a duplicate non-empty name is rejected with HTTP 400 naming the conflicting positions. Refs with an empty name are exempt from the uniqueness check.

# Basic: pipe recipe to bundle
curl -s "http://localhost:8080/v1/recipe?accelerator=h100&service=eks" | \
  curl -X POST "http://localhost:8080/v1/bundle" \
    -H "Content-Type: application/json" -d @- -o bundles.zip
# Advanced: with value overrides and Argo CD deployer
curl -s "http://localhost:8080/v1/recipe?accelerator=h100&service=eks" | \
  curl -X POST "http://localhost:8080/v1/bundle?deployer=argocd&repo=https://github.com/my-org/my-gitops-repo.git&set=gpuoperator:gds.enabled=true" \
    -H "Content-Type: application/json" -d @- -o bundles.zip
# With node scheduling for system and GPU nodes
# (recipe.json must be a fully-hydrated RecipeResult, e.g. from GET /v1/recipe)
curl -X POST "http://localhost:8080/v1/bundle?system-node-selector=nodeGroup=system&system-node-toleration=dedicated=system:NoSchedule&accelerated-node-selector=nvidia.com/gpu.present=true&accelerated-node-toleration=nvidia.com/gpu=present:NoSchedule" \
  -H "Content-Type: application/json" \
  -d @recipe.json \
  -o bundles.zip
# Override the shared DRA eviction label when bundling DRA with GPU Operator
curl -X POST "http://localhost:8080/v1/bundle?dra-eviction-node-label=example.com%2Fdra-ready%3Denabled" \
  -H "Content-Type: application/json" \
  -d @recipe.json \
  -o bundles.zip
# Generate bundles from a saved (fully-hydrated) recipe
curl -X POST "http://localhost:8080/v1/bundle" \
  -H "Content-Type: application/json" \
  -d @recipe.json \
  -o bundles.zip
# Elided literal body (NOT complete — use a generated RecipeResult instead)
curl -X POST "http://localhost:8080/v1/bundle" \
  -H "Content-Type: application/json" \
  -d '{
    "apiVersion": "aicr.run/v1",
    "kind": "RecipeResult",
    "componentRefs": [
      {"name": "gpu-operator", "type": "Helm", "chart": "gpu-operator", "source": "https://helm.ngc.nvidia.com/nvidia", "version": "v26.7.0", "namespace": "gpu-operator", "valuesFile": "components/gpu-operator/values.yaml"},
      {"name": "network-operator", "type": "Helm", "chart": "network-operator", "source": "https://helm.ngc.nvidia.com/nvidia", "version": "26.1.1", "namespace": "nvidia-network-operator", "valuesFile": "components/network-operator/values.yaml"}
    ],
    "deploymentOrder": ["gpu-operator", "network-operator"]
  }' \
  -o bundles.zip

Response Headers:

Before writing the response, the server stages a private, revalidated closed-world inventory. The ZIP contains only the inventory-derived directories and regular files, including recipe.yaml when present; unverified entries are rejected rather than archived. X-Bundle-Files and X-Bundle-Size are derived from that same frozen inventory.

Bundle Structure

bundles.zip
├── deploy.sh                    # root automation script (executable)
├── README.md                    # root deployment guide
├── checksums.txt                # SHA256 for every regular payload file in the archive
├── recipe.yaml                  # canonical post-resolution recipe (helm deployer)
├── 001-<component>/             # per-component folder (NNN-prefixed)
│   ├── install.sh               # component install script
│   ├── values.yaml              # static Helm values
│   ├── cluster-values.yaml      # per-cluster dynamic values
│   └── upstream.env             # CHART/REPO/VERSION (upstream-helm only)
└── 002-<component>/
    ├── install.sh
    ├── values.yaml
    └── cluster-values.yaml

Checksums are root-level only; component folders carry install.sh at their root (no scripts/ subdirectory), and no uninstall.sh/undeploy.sh is generated. After extraction, aicr verify . performs full closed-world verification: every manifest digest must match and every additional file or directory, symlink, or other non-regular object is rejected, except the exact allowed inventory metadata paths.

Server-Side Signing

Pass ?attest=true to POST /v1/bundle to receive a cryptographically signed bundle. The server signs the bundle as itself using an operator-configured signing identity. This is the trust boundary: no signing key, token, or identity is ever taken from the request, so any client that can reach the endpoint gets bundles signed under the server’s identity, never its own.

A signed bundle additionally carries, inside the returned zip:

attestation/bundle-attestation.sigstore.json   # signature over the bundle
attestation/aicr-attestation.sigstore.json     # aicrd tool-provenance attestation

attest=true requires a configured signing identity. If none is configured the request is rejected with HTTP 400 (Server is not configured for attestation). An unparseable attest value is also HTTP 400. Absent or false returns an unsigned bundle, as before.

The server supports two mutually exclusive signing modes, selected by environment variables at startup. The configuration is validated fail-fast: a malformed or ambiguous setting stops the server from starting.

Mode A: KMS key. Set AICR_SIGNING_KEY to a cosign KMS URI. The bundle is signed with a long-lived key held in the KMS; no OIDC identity is involved.

Mode B: keyless against a private Sigstore. Set AICR_FULCIO_URL to a private Fulcio CA plus a token source. The server obtains a short-lived signing certificate from Fulcio using its own OIDC identity. Operator setup for Mode B:

• Run a private Fulcio that trusts the cluster’s ServiceAccount token issuer.
• Mount a projected ServiceAccount token with audience sigstore into the aicrd pod, and point AICR_IDENTITY_TOKEN_FILE at it. The token is read fresh for every signed request because ServiceAccount tokens rotate.
• Alternatively, when aicrd itself runs inside GitHub Actions, its ambient OIDC environment is used as the token source.

Setting both AICR_SIGNING_KEY and the keyless variables is ambiguous and the server refuses to start.

Server signing also requires the aicrd binary attestation (aicrd-attestation.sigstore.json, issued under the NVIDIA-CI identity and bound to the aicrd binary digest) to be shipped inside the container image. The server verifies it once at startup and embeds it as tool provenance (attestation/aicr-attestation.sigstore.json) in every signed bundle. If signing is enabled but that attestation is missing or invalid, the server fails to start. Producing that attestation in the CI/release pipeline is a separate dependency, tracked outside this feature.

By default the server discovers that attestation next to its own executable. Set AICR_BINARY_ATTESTATION_FILE to point at an explicit path when the image stages it elsewhere: a ko-built image places assets under KO_DATA_PATH (/var/run/ko/aicrd-attestation.sigstore.json), not next to the binary. Only the attestation file path changes; it is still verified against the aicrd binary’s own digest.

GET /health

Service health check (liveness probe).

curl "http://localhost:8080/health"

Response:

{
  "status": "healthy",
  "timestamp": "2026-01-11T10:30:00Z"
}

GET /ready

Service readiness check (readiness probe).

curl "http://localhost:8080/ready"

Response:

{
  "status": "ready",
  "timestamp": "2026-01-11T10:30:00Z"
}

GET and HEAD /metrics

Prometheus metrics endpoint. HEAD returns the same headers with no body; every other method is rejected with 405 and an Allow: GET, HEAD header.

curl "http://localhost:8080/metrics"

Key Metrics:

Complete Workflow Example

Fetch a recipe and generate bundles in one workflow:

#!/bin/bash
# Step 1: Get recipe for H100 on EKS for training
echo "Fetching recipe..."
curl -s "http://localhost:8080/v1/recipe?accelerator=h100&service=eks&intent=training" \
  -o recipe.json
# Display recipe summary
echo "Recipe components:"
jq -r '.componentRefs[] | "  - \(.name): \(.version)"' recipe.json
# Step 2: Generate bundles from recipe (pipe directly)
# recipe.json is the fully-hydrated RecipeResult fetched in Step 1.
echo "Generating bundles..."
curl -s -X POST "http://localhost:8080/v1/bundle" \
  -H "Content-Type: application/json" \
  -d @recipe.json \
  -o bundles.zip
# Alternative: one-liner without intermediate file
# curl -s "http://localhost:8080/v1/recipe?accelerator=h100&service=eks" | \
#   curl -X POST "http://localhost:8080/v1/bundle" \
#     -H "Content-Type: application/json" -d @- -o bundles.zip
# Step 3: Extract and verify
echo "Extracting bundles..."
unzip -q bundles.zip -d ./deployment
# Verify the complete inventory (checksums.txt is at the bundle root)
echo "Verifying bundle inventory..."
cd deployment
aicr verify .
# Step 4: Deploy (example)
echo "Bundle ready for deployment:"
ls -la

Error Handling

Error Response Format

{
  "code": "ERROR_CODE",
  "message": "Human-readable error description",
  "details": { ... },
  "requestId": "550e8400-e29b-41d4-a716-446655440000",
  "timestamp": "2026-01-11T10:30:00Z",
  "retryable": true
}

Error Codes

INVALID_REQUEST is not always 400: POST /v1/query and POST /v1/recipe return it with HTTP 413 Request Entity Too Large when the request body exceeds the server’s body-size limit (MaxRecipePOSTBytes).

Handling Rate Limits

# Check rate limit headers. HEAD resolves the recipe exactly as GET does and
# returns the same headers without the body -- it is not a cheap probe.
curl -I "http://localhost:8080/v1/recipe?accelerator=h100"
# Response headers:
# X-RateLimit-Limit: 100
# X-RateLimit-Remaining: 95
# X-RateLimit-Reset: 1736589000

When rate limited (HTTP 429), use the Retry-After header:

# Retry with backoff
response=$(curl -s -w "%{http_code}" "http://localhost:8080/v1/recipe?accelerator=h100")
if [ "${response: -3}" = "429" ]; then
  # HTTP headers end in CRLF, so strip the carriage return before sleep reads
  # the value.
  retry_after=$(curl -sI "http://localhost:8080/v1/recipe?accelerator=h100" | grep -i "Retry-After" | awk '{print $2}' | tr -d '\r')
  echo "Rate limited. Retrying after ${retry_after}s..."
  sleep "$retry_after"
fi

Rate Limiting

• Limit: 100 requests per second (a single process-global token bucket shared across all clients, not per-IP)
• Burst: 200 requests
• Headers: X-RateLimit-Limit, X-RateLimit-Remaining, X-RateLimit-Reset
• 429 Response: Includes Retry-After header

Criteria Allowlists

The API server can be configured to restrict which criteria values are allowed. This enables operators to limit the API to specific accelerators, services, intents, or OS types.

Configuration

Allowlists are configured via environment variables when starting the server:

Behavior:

• If an environment variable is not set, all values for that criteria are allowed
• If an environment variable is set, only the specified values are permitted
• The any value is always allowed regardless of allowlist configuration
• Allowlists apply to /v1/recipe, /v1/query, and /v1/bundle

Example Configuration

# Start server allowing only H100 and L40 GPUs on EKS
docker run -p 8080:8080 \
  -e AICR_ALLOWED_ACCELERATORS=h100,l40 \
  -e AICR_ALLOWED_SERVICES=eks \
  ghcr.io/nvidia/aicrd:latest

Error Response

When a disallowed criteria value is requested:

curl "http://localhost:8080/v1/recipe?accelerator=gb200&service=eks"

Response (HTTP 400):

{
  "code": "INVALID_REQUEST",
  "message": "accelerator type not allowed",
  "details": {
    "requested": "gb200",
    "allowed": ["h100", "l40"]
  },
  "requestId": "550e8400-e29b-41d4-a716-446655440000",
  "timestamp": "2026-01-27T10:30:00Z",
  "retryable": false
}

CLI Behavior

The CLI (aicr) is not affected by allowlists. Allowlists only apply to the API server, allowing operators to restrict API access while maintaining full CLI functionality for administrative tasks.

Programming Language Examples

Python

import requests
import zipfile
import io
BASE_URL = "http://localhost:8080"
# Get recipe
params = {
    "accelerator": "h100",
    "service": "eks",
    "intent": "training",
    "os": "ubuntu"
}
resp = requests.get(f"{BASE_URL}/v1/recipe", params=params)
resp.raise_for_status()
recipe = resp.json()
print(f"Recipe has {len(recipe['componentRefs'])} components")
# Generate bundles — the (fully-hydrated) recipe is the request body.
resp = requests.post(
    f"{BASE_URL}/v1/bundle",
    json=recipe,
)
resp.raise_for_status()
# Extract zip
with zipfile.ZipFile(io.BytesIO(resp.content)) as zf:
    zf.extractall("./deployment")
    print(f"Extracted {len(zf.namelist())} files")

Go

package main
import (
    "encoding/json"
    "fmt"
    "io"
    "net/http"
    "net/url"
    "os"
)
func main() {
    baseURL := "http://localhost:8080"
    // Get recipe
    params := url.Values{}
    params.Add("accelerator", "h100")
    params.Add("service", "eks")
    
    resp, err := http.Get(baseURL + "/v1/recipe?" + params.Encode())
    if err != nil {
        panic(err)
    }
    defer resp.Body.Close()
    var recipe map[string]interface{}
    json.NewDecoder(resp.Body).Decode(&recipe)
    
    fmt.Printf("Got recipe with %d components\n", 
        len(recipe["componentRefs"].([]interface{})))
}

JavaScript/Node.js

const BASE_URL = "http://localhost:8080";
async function main() {
    // Get recipe
    const params = new URLSearchParams({
        accelerator: "h100",
        service: "eks",
        intent: "training"
    });
    
    const recipeResp = await fetch(`${BASE_URL}/v1/recipe?${params}`);
    const recipe = await recipeResp.json();
    
    console.log(`Recipe has ${recipe.componentRefs.length} components`);
    
    // Generate bundles — the (fully-hydrated) recipe is the request body.
    const bundleResp = await fetch(`${BASE_URL}/v1/bundle`, {
        method: "POST",
        headers: { "Content-Type": "application/json" },
        body: JSON.stringify(recipe),
    });
    
    // Save zip
    const buffer = await bundleResp.arrayBuffer();
    require("fs").writeFileSync("bundles.zip", Buffer.from(buffer));
    console.log("Bundles saved to bundles.zip");
}
main();

Shell Script (Batch Processing)

#!/bin/bash
# Generate recipes for multiple environments
# /v1/recipe accepts every family, profiled and unprofiled alike.
environments=(
  "os=ubuntu&accelerator=h100&service=eks"
  "os=cos&accelerator=h100&service=gke"
  "os=ubuntu&accelerator=h100&service=aks"
)
for env in "${environments[@]}"; do
  echo "Fetching recipe for: $env"
  curl -s "http://localhost:8080/v1/recipe?${env}" \
    | jq -r '.componentRefs[] | "\(.name): \(.version)"'
  echo ""
done

OpenAPI Specification

The full OpenAPI 3.1 specification is available at: api/aicr/v1/server.yaml

Generate client SDKs:

# Download spec
curl https://raw.githubusercontent.com/NVIDIA/aicr/main/api/aicr/v1/server.yaml \
  -o openapi.yaml
# Generate Python client
openapi-generator-cli generate -i openapi.yaml -g python -o ./python-client
# Generate Go client
openapi-generator-cli generate -i openapi.yaml -g go -o ./go-client
# Generate TypeScript client
openapi-generator-cli generate -i openapi.yaml -g typescript-fetch -o ./ts-client

Troubleshooting

Common Issues

“Invalid accelerator type” error:

# Use valid values: h100, h200, gb200, gb300, b200, a100, l40, l40s, rtx-pro-6000, vr200, any
curl "http://localhost:8080/v1/recipe?accelerator=h100"

“Recipe is required” error:

# The body IS the RecipeResult itself — not wrapped in a {"recipe": ...} field.
# Pass a fully-hydrated RecipeResult (e.g. from GET /v1/recipe) directly:
curl -s "http://localhost:8080/v1/recipe?accelerator=h100&service=eks" | \
  curl -X POST "http://localhost:8080/v1/bundle" \
    -H "Content-Type: application/json" -d @- -o bundles.zip

Empty zip file:

# Check recipe has componentRefs
curl -s "http://localhost:8080/v1/recipe?accelerator=h100" | jq '.componentRefs'

Connection refused (local):

# Start local server first
make server

See Also

• CLI Reference - Command-line interface
• Agent Deployment - Kubernetes agent for snapshot capture
• Installation Guide - Setup instructions
• Data Flow - Understanding recipe data architecture
• Automation Guide - CI/CD integration patterns
• Kubernetes Deployment - Self-hosted API server deployment