SageMaker Marketplace Onboarding — an Agent Skill for Claude Code, Amazon Quick, Kiro, and Codex

An interactive Agent Skill that walks model providers through building an inference container that complies with the Amazon SageMaker Marketplace container contract — and, optionally, through publishing a Marketplace listing. Works with Claude Code, Amazon Quick, Kiro, Codex, and any other tool that supports the standard SKILL.md format.

You describe your model; your agent guides you phase by phase — inventorying an existing project or scaffolding a new one, enforcing the /ping and /invocations contract, packaging model weights, running a local validation gate, and pushing to ECR. It is 100% guidance and templates: no code runs on your behalf, and it never edits your existing project files.

Non-production disclaimer. The templates this skill scaffolds (app.py, Dockerfile, model_loader.py, inference.py, metering.py, websocket_handler.py, and the rest of templates/) are provided for demonstration and as a starting point only. They are not intended for production or Marketplace submission as-is and have not undergone security review. Before deploying or submitting a listing, you are responsible for your own security review and testing — including input validation, authentication/authorization where applicable, dependency pinning and vulnerability scanning, and any controls your use case and compliance obligations require.

---

What you get

• Works with existing projects or greenfield. For existing code, the skill produces a gap report and scaffolds a compliant version alongside your files — you merge by hand. For new projects, it scaffolds from templates.
• One /invocations implementation, four modes. The same container serves Real-Time, Streaming, Async Inference, and Batch Transform. SageMaker handles routing.
• A real /ping health check that exercises the inference path, instead of a static 200 that passes health checks while the model is broken (the most common Marketplace validation failure).
• Correct weights handling. Weights load from /opt/ml/model/ and ship separately as model.tar — no baked-in weights, no network calls at runtime.
• Opt-in modules for bidirectional WebSocket streaming and per-inference metering, scaffolded only if you need them.
• A local validation gate (test/test_local.sh) you run before you ever push to ECR.
• Optional listing hand-off that takes you through CreateModelPackage and the validation job if you want to publish.

Requirements

• One of: Claude Code, Amazon Quick, Kiro, Codex, or another agent that reads the standard SKILL.md format.
• For local container testing: Docker, and the AWS CLI configured if you push to ECR.
• Python 3.12+ if you use the WebSocket client SDK notes (see reference/websocket.md).

Install

Skill lives in sagemaker-marketplace-onboarding/. Repo root has a .claude-plugin/marketplace.json manifest so discovery tools find it without scanning.

Claude Code

/plugin marketplace add aws-samples/sample-sagemaker-marketplace-onboarding-skill
/plugin install sagemaker-marketplace-onboarding@sample-sagemaker-marketplace-onboarding-skill

Or via the skills CLI: npx skills add aws-samples/sample-sagemaker-marketplace-onboarding-skill -a claude-code

Or manually: download sagemaker-marketplace-onboarding.skill from Releases and unzip sagemaker-marketplace-onboarding.skill -d ~/.claude/skills/, or git clone + cp -r sagemaker-marketplace-onboarding ~/.claude/skills/.

Kiro CLI

npx skills add aws-samples/sample-sagemaker-marketplace-onboarding-skill -a kiro-cli

Kiro IDE

Manual flow, not CLI-based: Agent Steering & Skills sidebar > add skill > paste the folder URL: https://github.com/aws-samples/sample-sagemaker-marketplace-onboarding-skill/tree/main/sagemaker-marketplace-onboarding

Codex

npx skills add aws-samples/sample-sagemaker-marketplace-onboarding-skill -a codex

Amazon Quick

Desktop app, manual: Agents & skills > Skills tab > + Create > Import from file > select sagemaker-marketplace-onboarding/SKILL.md. See Amazon Quick skills docs.

Restart your agent (or start a fresh session) so it picks up the new skill.

Usage

Start a session with your agent and describe what you want. The skill activates on requests like:

• "Help me build a SageMaker Marketplace container"
• "Onboard my model to SageMaker Marketplace"
• "Package my model for SageMaker Marketplace"
• "Review my container against the SageMaker Marketplace spec"

Your agent asks two framing questions up front — container-only vs. also-list, and existing project vs. greenfield — then walks the rest one phase at a time, confirming before it moves on.

The walkthrough at a glance

Phase	What happens
0	Goal + project-state questions that branch the flow
1–2	Inventory an existing project, then a discovery interview
3–4	Gap report and/or scaffold from templates
5–10	Contract, Dockerfile, weights packaging, local testing, ECR push, pre-submission checklist
11	(also-list only) CreateModelPackage + validation job + AWS docs hand-off
12	(optional, ask-only) Operating a live listing — observability + IAM temporary delegation pointers
13	(optional, ask-only) Pipecat voice-agent ecosystem hand-off for STT/TTS WebSocket models

Modality- and mode-specific questions are gated: an LLM provider never sees text-to-speech questions, and a container-only user never sees listing steps. Phases 12 and 13 never trigger unprompted — they're pointers surfaced only if the user asks, or (for 13) if the modality/WebSocket/testing preconditions are already met.

Repository layout

sagemaker-marketplace-onboarding/        ← the installable skill
├── SKILL.md                             ← the walkthrough your agent follows, phase by phase
├── templates/                           ← code files scaffolded into your project
│   ├── Dockerfile
│   ├── requirements.txt
│   ├── app.py                           ← FastAPI server: /ping, /invocations, /execution-parameters
│   ├── model_loader.py
│   ├── inference.py                     ← smoke_check + predict + predict_stream + zip_outputs
│   ├── metering.py                      ← per-inference billing headers (opt-in)
│   ├── websocket_handler.py             ← bidirectional streaming (opt-in)
│   ├── supervisord.conf                 ← multi-process container config (opt-in)
│   ├── package_model.sh                 ← builds model.tar and uploads to S3
│   ├── PRE_SUBMISSION_CHECKLIST.md
│   └── test/                            ← test_input.json, test_local.sh, test_streaming.py, test_websocket.py
└── reference/                           ← docs your agent reads to cite constraints (not scaffolded)
    ├── contract.md                      ← endpoint/header/timing cheat sheet
    ├── timing.md                        ← hard timing limits
    ├── checklist.md                     ← full pre-submission checklist
    ├── gap-checks.md                    ← gap-analysis list for existing projects
    ├── websocket.md                     ← bidirectional streaming + client SDK notes
    ├── billing.md                       ← hourly vs. per-inference metering
    ├── logging.md                       ← CloudWatch logging patterns
    ├── marketplace-listing.md           ← CreateModelPackage skeleton for the also-list path
    ├── observability.md                 ← CloudWatch metrics catalog + EMF business-metric emission (Phase 12)
    ├── iam-temporary-delegation.md      ← buyer-approved temporary support access for a live listing (Phase 12)
    └── pipecat-integration.md           ← Pipecat voice-agent orchestration hand-off (Phase 13)

tests/ — repo smoke tests for the templates and docs (not part of the installable skill).

.claude-plugin/ — plugin.json + marketplace.json, so Claude Code's native /plugin marketplace add and the npx skills CLI both find the skill directly, without a directory scan.

sagemaker-marketplace-onboarding.skill — a prebuilt zip of the skill folder, attached to each release.

• templates/ — files copied into your project. Changing one changes what you ship.
• reference/ — docs your agent reads to answer questions and cite constraints, without touching your code.

Scope — what this skill does not do

By design, the skill:

• Never edits your existing project files. It produces a gap report and scaffolds a compliant version alongside; you merge by hand.
• Makes no AWS API calls on your behalf. No boto3, no live SageMaker actions — it guides, you act.
• Stops at CreateModelPackage. FDP enrollment, IAM roles, pricing, EULA, and publishing live with AWS account teams and the Marketplace Management Portal.
• Does not test your model's accuracy. It validates AWS Sagemaker contract compliance, not model quality.

Contributing

Contributions are welcome. Found a bug, an outdated constraint, or a gap? Open an issue or send a pull request. See CONTRIBUTING.md for details. The skill is entirely documentation and templates, so most changes are edits to SKILL.md, reference/*.md, or templates/.

Please test locally before opening a PR: install the skill into your own agent (Claude Code, Amazon Quick, Kiro, or Codex), run the walkthrough end to end, and for template changes do a full scaffold plus bash test/test_local.sh <image> <weights>. There is no CI — manual testing is the gate. Also run the stdlib-only repo smoke tests (template syntax, metering helpers, inference.count consistency) from the repo root: python3 -B -m unittest discover -s tests.

A few design principles are load-bearing and shouldn't be changed casually: /ping must exercise inference, ENTRYPOINT must be exec-form, no weights baked into the image and no runtime network calls, container code stays mode-agnostic, and every walkthrough question uses structured prompts rather than free text.

License

See LICENSE.