aws-startup-advisor
Overview
AWS Startup Advisor puts an AWS Startups Solutions Architect in your coding assistant. It recommends the smallest architecture that fits your stage, credits, and team, then can scaffold it, migrate you from Azure, Google Cloud, Heroku, or Render, move your model calls to Amazon Bedrock, and investigate production incidents. It also answers your AWS Activate credits eligibility questions and finds startup offers you can use.
Ask it things like:
- "We're pre-seed with two engineers. What should our AWS architecture look like?"
- "Migrate us off Google Cloud and tell me what it costs."
- "Which runtime should this agent use?"
- "Our API started returning 502s after the last deploy."
The sixteen skills below are the catalog. Install them together.
Agent Skills
| # | Skill | Description | Documentation |
|---|---|---|---|
| 1 | architect-for-startups | Stage-aware AWS architecture guidance for non-agent workloads, tuned to team size, runway, and credits — advice rather than code changes. An AI agent as the workload goes to agent-advisor | SKILL.md |
| 2 | start-building-for-startups | Interactive discovery flow that gathers requirements, scans the codebase, then writes an AWS scaffold and implementation into the project | SKILL.md |
| 3 | agent-advisor | Runtime selection, migration planning, and an executable proof of concept for AI-agent workloads on AWS | SKILL.md |
| 4 | azure-to-aws | Six-phase Microsoft Azure to AWS migration. Discovery reads Terraform (azurerm_*), a live az CLI capture (read-only, consent-gated), a customer-supplied RDfA report (read-only, file-only), and application code — not Bicep, ARM templates, or Azure billing exports. Direct OpenAI, OpenRouter, and Anthropic usage-API discovery (read-only, consent-gated) supplement real AI spend. Generate is opt-in, and the what-if workshop is optional | SKILL.md |
| 5 | gcp-to-aws | Six-phase Google Cloud to AWS migration. Discovery reads Terraform, a live gcloud capture (read-only, consent-gated), application code, and billing exports. Direct OpenAI, OpenRouter, and Anthropic usage-API discovery (read-only, consent-gated) supplement real AI spend. Generate is opt-in. BigQuery is not mapped to an AWS analytics service | SKILL.md |
| 6 | heroku-to-aws | Six-phase Heroku to AWS migration with deterministic add-on mapping and an optional what-if repricing workshop | SKILL.md |
| 7 | render-to-aws | Render-to-AWS migration with consented read-only Blueprint, CLI, REST, application, and billing discovery; mandatory Clarify; deterministic design and estimates; optional workshop; and opt-in validated artifacts | SKILL.md |
| 8 | llm-to-bedrock | Rewrite OpenAI, Gemini, or Anthropic API call sites to Amazon Bedrock, evaluate quality, and deliver a git branch. The full migration requires gcp-to-aws or azure-to-aws alongside it (gcp-to-aws preferred when both are installed). Access-only mode (enable model access, no rewrite) does not | SKILL.md |
| 9 | tf-best-practices | AWS Terraform authoring posture, security-baseline spec, and a read-only policy gate over generated Terraform | SKILL.md |
| 10 | operate-on-aws | Incident investigation, root-cause analysis, and release-readiness review with AWS DevOps Agent, with cost, access, and code-egress disclosed up front | SKILL.md |
| 11 | knowledge-base-for-startups | AWS Activate FAQ, credits guide, programs, partner offers, sample architectures, and AWS-curated learn articles | SKILL.md |
| 12 | prompt-library-for-startups | AWS-curated copy-paste prompts for AI coding agents, plus downloadable installable agents | SKILL.md |
| 13 | contextual-offers-for-startups | Appends at most one relevant AWS Activate partner offer as optional context after another skill's output is final | SKILL.md |
| 14 | bedrock-migration-assessment | Scans your code for OpenAI, Anthropic, Gemini, and OpenRouter calls (or validates a provider-neutral ai-workload-profile.json), can measure OpenAI, Anthropic, and OpenRouter usage read-only with your consent, runs the AI-only migration questions, and designs a Bedrock target, endpoint, and client API per workload. Needs no infrastructure skill. This release stops after Design; Estimate and Generate land in later releases | SKILL.md |
| 15 | cost-regression-check | Prices an infrastructure-as-code change on every pull request before apply: the fixed monthly floor it adds and the change in cost per unit of usage, at list price before credits, against thresholds derived from your monthly ceiling | SKILL.md |
| 16 | self-hosted-llm-batch-inference | Recurring offline inference for a self-hosted 1B to 8B model on one GPU with AWS Batch on Spot, vLLM, deterministic shards, and verified completion and capacity release | SKILL.md |
contextual-offers-for-startups is consulted by the other skills rather than invoked directly: it runs only after a recommendation, plan, or build is already final, and it never influences the technical advice.
Bundled Resources
Alongside the skills, the plugin ships supporting material the skills load on demand:
skills/shared/— Canonical reference material shared across skills rather than an invocable skill of its own: the phase-workflow interpreter contract, estimation schemas and complexity tiers, a snapshot of AWS infrastructure rates, phase-status schema, what-if workshop invariants, shared Clarify question fragments, and the AI-migration guardrails and Bedrock model references. Each consuming skill carries a byte-identical copy under its ownreferences/vendored/so the skill folder stays self-contained.agents/— Seven Claude Code subagent definitions: two generic, phase-agnostic migration phase workers (read/write and read/write/shell tiers) plus five specialists for the Bedrock rewrite flow (code analyzer, code rewriter, log ingestor, prompt evaluator, report generator).tools/— Containsmodel-staleness.py, which checks the Bedrock model-lifecycle registry's freshness contract (**Last updated**/**Staleness window**) and that every vendored copy underskills/*/references/vendored/ai/is byte-identical toskills/shared/ai/ai-model-lifecycle.md(--strictalso fails on a stale snapshot). It also containsbuild-model-catalogs.py, which renders the per-provider model-selection catalogs fromskills/shared/ai/models/bedrock-model-registry.json(--checkfails when they differ), andmodel-catalog-drift.py, which warns when a model the lifecycle registry marks Active is missing from the shared model-selection catalogs inskills/shared/ai/models/, or a catalog'sverified_atis older than the registry's staleness window (--strictfails on either).scripts/— Python helpers that validate generated artifacts, emit run summaries, and lint Bedrock model references: a migration report validator, a startup-program artifact check, a plan summary emitter (emit-plan-json.py),validate-artifacts.py, which checks every phase artifact in a run directory (or the fixture goldens, in CI) against the contract itsschema-*.md/ JSON Schema publishes — unknown keys, missing required keys, type and enum drift — andmodel-id-lint.py, which fails when a file outside the model catalog names a Bedrock model the lifecycle registry marks excluded or removed, or a fabricated dated Sonnet 4.6 / Opus 4.8 ID.scripts/artifact-contracts.jsonmaps artifacts to contracts;scripts/artifact-contracts-baseline.jsonis the explained burn-down list of known contract gaps. The Heroku flow's report validator ships insideskills/heroku-to-aws/scripts/.scripts/telemetry/collects optional usage data: nothing is sent until you accept, no skill depends on it, and/usage-datashows the setting or opts out.scripts/contracts/ai-workload-profile.schema.jsonadditionally defines the provider-neutralai-workload-profile.jsonobservation contract.render-to-awsproduces it from application discovery and recordsdeferred_shared_ai_engine;bedrock-migration-assessmentconsumes it without requiring an infrastructure skill.scripts/contracts/aws-design-ai.schema.json, canonicalized from Azure's prior single-scalarmetadata.ai_sourcecontract intoskills/shared/ai/schema-design-aws-ai.md, likewise defines a target-onlyaws-design-ai.jsondesign contract whosemetadata.ai_sources[]/metadata.gateways[]mirror the workload-profile contract exactly;design-ai.mdin both skills still writes against the prior per-skill shape, so this is a target, not a claim about current output.scripts/contracts/ai-capability-catalog.schema.jsondefines the target-only contract forskills/shared/ai/ai-capability-catalog.json, a static source-provider/gateway capability catalog; itsimplemented_adapter/routefields are null everywhere and it asserts no runtime support.scripts/contracts/model-selection-catalog.schema.jsondefines the contract for the shared model-selection catalogs inskills/shared/ai/models/that agent-advisor's selectors read; each model'sselection_status(selectableorreference_only) decides whether it can be picked automatically.scripts/contracts/estimation-ai.schema.jsonandscripts/contracts/generation-ai.schema.json, canonicalized intoskills/shared/ai/schema-estimate-ai.md/schema-generate-ai.md, likewise define target-onlyestimation-ai.json/generation-ai.jsoncontracts whosecost_by_source[]/artifacts_by_source[]carry a per-source/per-model traceability chain back throughaws-design-ai.json'smetadata.ai_sources[]/gateways[];estimate-ai.mdandgenerate-ai.md/generate-artifacts-ai.mdin both skills still write against their prior per-skill shapes, so these are targets, not claims about current output.fixtures/— Reference and regression fixtures for the migration report and estimation artifacts, used to keep generated output within the documented contract.fixtures/azure-iac-terraform/additionally carries deterministic asserters that pin the Azure Discover, Clarify, Design, and Estimate contracts against a committed synthetic estate.
MCP Servers
| # | Server | Description |
|---|---|---|
| 1 | aws-mcp | AWS API access, documentation search, regional availability, and skill retrieval via AWS MCP Server |
aws-mcp is a single stdio server launched through uvx mcp-proxy-for-aws-cli==1.7.0. The skills use it for aws___search_documentation and aws___read_documentation (current AWS documentation), aws___get_regional_availability and aws___list_regions (service availability per region), aws___call_aws (authenticated AWS API calls), aws___run_script (sandboxed Python), and aws___retrieve_skill and aws___recommend (on-demand guidance). No other MCP server is required: in particular, render-to-aws neither configures nor invokes a Render MCP server and uses only its consented file, CLI, and GET-only REST discovery boundaries.
Installation
In Claude Code, install from this repository's marketplace:
/plugin marketplace add aws/agent-toolkit-for-aws
/plugin install aws-startup-advisor@agent-toolkit-for-aws
/reload-plugins
For Codex, Cursor, and Devin, see Quick start.
For standalone skill installs — Kiro, fx, and other hosts that consume skills directly — point the skills CLI at this plugin. The repository's top-level skills/ tree does not contain these skills, so the command in Quick start will not install them:
npx skills add aws/agent-toolkit-for-aws/plugins/aws-startup-advisor/skills --skill '*'
Install all 16 skills together rather than a subset: agent-advisor delegates to gcp-to-aws, and the migration skills share vendored fragments. A standalone install covers the skills only — it does not configure the aws-mcp server declared in .mcp.json, which the host needs separately.
Development checks
From the repository root, run mise -C plugins/aws-startup-advisor run ci.
The plugin's mise.toml inherits the root tool configuration. It runs the lint tasks, including model-reference lint, and every test suite: the contract tests in tests/, the agent-advisor, llm-to-bedrock, render-to-aws, and skill-script suites, telemetry, and the Heroku source-review tests.
It also declares Terraform under [tools]: the Render replay test runs terraform fmt, init -backend=false, and validate against each generated migration pack and fails closed if the binary is missing, so the workflow installs it through mise rather than relying on the runner image.
To add a check, define a task in this file and add it to tasks.ci.depends; the workflow keeps calling the same ci task. To add a tool a check needs, pin it under [tools] in this file; the workflow's mise-action step installs whatever this file declares.
The Startup advisor workflow runs on pull requests that change this plugin, the root mise.toml, or .github/workflows/startup-advisor.yml. It also supports manual runs.
Independently accept a migration run
From plugins/aws-startup-advisor in a repository checkout, use the local
acceptance checker after a GCP or Heroku infrastructure run. It
re-runs the contract, report, Terraform policy and generated-artifact checks
instead of trusting a job's Completed label or its saved PASS reports.
Python 3.10 or newer is required.
python3 scripts/accept-migration-run.py \
--run-dir /path/to/.migration/1030-1400 \
--skill gcp-to-aws \
--output /path/to/review/acceptance.json
The input directory stays unchanged; the output must be a new file outside it.
Exit 0 means PASS, 1 means FAIL, and 2 means UNVERIFIED. The JSON
includes the input and checker hashes, findings, and actual command results.
Checker hashes include the contract documents selected by the artifact manifest.
A decision-only pass covers the decision pack, including its rendered report;
it does not approve retained Terraform from an earlier execution run.
Execution packs remain UNVERIFIED until Terraform checks actually run.
After reviewing the generated configuration, add
--terraform /absolute/path/to/terraform to allow fmt, init -backend=false
and validate in a temporary copy. This downloads and executes only the
official HashiCorp AWS provider, with a separate Terraform CLI configuration
and no inherited cloud credentials. Modules, other providers, nested or JSON
configurations, and filesystem-reading functions need separate review and
remain UNVERIFIED. No plan, apply, migration script or telemetry runs.
For GCP execution packs, the Terraform floor includes main.tf, variables.tf,
outputs.tf, baseline.tf, its README.md, and a nonempty domain file.
Missing or empty prerequisite and validation scripts also fail acceptance.
Data, container, and secrets scripts are required when the current design
contains the corresponding resource categories.
For Heroku execution packs, missing Redis scripts, selected Elastic Beanstalk
deployment files, or required EKS files and resources fail acceptance. EKS packs
remain UNVERIFIED even when these outputs exist because Kubernetes manifests
and node-group wiring still need separate review.
This is a local, bounded acceptance check. It does not change ATX job status, support AI/billing-only routes, or prove deployment and data-migration success.
For GCP workshop runs, acceptance also checks scenario identities, snapshot paths, and saved inventory/design/preference fingerprints. Generate and acceptance share the v2 Terraform-report contract. The generated-artifact checker catches unsupported RDS instance password arguments and resolvable local ECS secret-selection mismatches.
The PostgreSQL script examples support a full offline copy with writers frozen and a tested data verifier. Online replication and final delta synchronization require a separate implementation. A missing verifier or unsupported transfer must remain on HOLD in the generated guide and report.
Startup Architecture Advice
The architect-for-startups skill answers "what should we build on AWS?" for a workload that is not an AI agent, the way a Startup Solutions Architect would: it establishes the company's stage, team size, runway, and credit position first, then recommends the smallest architecture that clears the bar, and names what to revisit at the next stage.
How It Works
- Stage-aware defaults — Recommendations differ for a pre-revenue prototype and a Series B workload with paying customers; the skill asks before it assumes.
- Cost and credit awareness — Options are framed in terms of run-rate and credit burn, so the architecture stretches AWS Activate credits rather than consuming them in a month.
- Reviews and diligence — Existing architectures can be reviewed for a fundraise or technical diligence conversation, with gaps ranked by what an investor or acquirer will ask about.
- Multi-tenant isolation — For a multi-tenant SaaS, picks a silo, pool, or bridge model per layer and pushes isolation into IAM and the database, so a missing tenant filter fails closed rather than leaking another tenant's data.
- Advice, not edits — This skill recommends and reviews. When you want the architecture written into the repository, it hands off to
start-building-for-startups. When the workload is an AI agent — deploying one, picking its runtime, or moving agents to AWS — it hands off toagent-advisor.
Examples
- "We're pre-seed with two engineers — what should our AWS architecture look like?"
- "Review our architecture before our Series A technical diligence"
- "How do we make our Activate credits last another six months?"
- "Is Aurora Serverless the right call at our stage?"
- "How should we isolate tenants in our pooled SaaS before our first enterprise customer?"
Guided Building
The start-building-for-startups skill runs a structured discovery flow and then writes code. It gathers intent, scope, constraints, and preferences through picker-based questions, infers what it can from the existing codebase, and produces an AWS architectural scaffold and implementation in the project.
Examples
- "Help me build the backend for our new app on AWS"
- "Scaffold an AWS project for a multi-tenant SaaS MVP"
- "Expand our existing service to add async processing"
- "Refactor this app onto managed AWS services"
AI Agent Runtimes
The agent-advisor skill is the entry point for agentic work on AWS. It covers runtime selection, migration planning for agents already running elsewhere, and building an executable proof of concept, as one phased flow.
How It Works
- Runtime selection — Compares Amazon Bedrock AgentCore, Amazon ECS, Amazon EKS, AWS Lambda, and Lambda MicroVMs against the workload's latency, session, tool-calling, and operational requirements.
- Migration planning — Plans a move for existing agent workloads, including adding AgentCore capabilities (memory, gateway, identity, policy, observability) to an agent that already runs on AWS.
- Durable execution — Covers running Temporal workers on AWS and the Temporal Cloud versus self-hosted decision. Temporal workflow code is not rewritten into another orchestrator.
- Proof of concept — Produces a runnable POC rather than a slide-level recommendation.
The skill requires at least one agentic component. Non-agent compute or data migrations route to azure-to-aws, gcp-to-aws, heroku-to-aws, or render-to-aws, and a pure LLM SDK rewrite routes to llm-to-bedrock.
Examples
- "Which runtime should I use for my agent — AgentCore, ECS, EKS, or Lambda?"
- "Move our LangGraph agents to AWS"
- "Add memory and a gateway to the agent we already run on ECS"
- "We orchestrate with Temporal — how does that run on AWS?"
Cloud Migration
The azure-to-aws, gcp-to-aws, heroku-to-aws, and render-to-aws skills run the same six-phase flow: discover, clarify, design, estimate, generate, and feedback. Clarify must finish before design, estimate, or generate, so the plan is never built on unstated assumptions. Generate runs only after you choose to produce the artifacts at the post-estimate decision gate. The what-if workshop is optional and sits beside estimate; it is not an extra phase.
How It Works
- Discover —
azure-to-awsreads Terraform (azurerm_*) and application code, can capture a live subscription through theazCLI (read-only, consent-gated), and can parse a customer-supplied Resource Discovery for Azure (RDfA) report (read-only, file-only). It does not read Bicep or ARM templates or Azure billing exports.gcp-to-awsreads Terraform, application code, and billing exports, and can capture a live project throughgcloud(read-only, consent-gated). BigQuery is recorded and left for a specialist; the skill does not pick an AWS analytics service for it. It points users tobq-assessas a read-only first step.azure-to-awsandgcp-to-awscan also read aggregate AI spend when the app calls OpenAI directly, calls it through OpenRouter, or calls the Anthropic API — from the OpenAI Admin API (Usage Read), the OpenRouter API, or the Anthropic Admin API. Each is consent-gated and optional, and none replaces the code scan.heroku-to-awsdiscovers through the authenticated Heroku CLI (read-only, consent-gated) or fromProcfileandapp.json.render-to-awsreadsrender.yaml, application and billing files, and optionally captures the official Render CLI and an exact GET-only REST allowlist after explicit workspace confirmation and consent; it never uses Render MCP or Render-side mutation. - Clarify — Resolves the requirements that change the target architecture: availability, compliance, data residency, cutover tolerance, and team capacity.
- Design — Maps source resources to AWS services using deterministic mapping tables — for example Heroku dynos to AWS Elastic Beanstalk, Heroku Postgres to Amazon RDS or Aurora, Heroku Redis to Amazon ElastiCache for Valkey (the default engine; ElastiCache for Redis OSS is also selectable), Heroku Kafka to Amazon MSK, Cloud SQL to Amazon RDS, GKE to Amazon EKS, Cloud Run to AWS Fargate, Azure App Service to AWS Elastic Beanstalk, AKS to Amazon EKS, Azure SQL to Amazon RDS, and Cosmos DB to Amazon DynamoDB or DocumentDB.
render-to-awsadds ordered exactly-one selectors — for example Render static sites to Amazon S3 and CloudFront, Render services to Elastic Beanstalk or AWS Fargate according to management model, Render Postgres to one selected PostgreSQL target, Render Key Value to Amazon ElastiCache, and Render persistent disks to one storage target. - Estimate — Costs the target architecture from the plugin's bundled rate reference data and a documented cost algorithm, with an estimation schema and complexity tiers so two runs of the same workload agree. Estimates are planning figures; confirm them against the AWS Pricing Calculator before committing budget.
- Generate — Emits migration artifacts, including Terraform, gated by the
tf-best-practicespolicy check and a validated migration report. On the infrastructure route,heroku-to-aws,gcp-to-aws, andazure-to-awsalso emitbaseline.tf— an account-wide security baseline (GuardDuty, CloudTrail, IMDSv2, EBS encryption, budget alerts) — with AWS Config and Security Hub controls added when the declared compliance set includes SOC 2, PCI, HIPAA, or FedRAMP.gcp-to-aws's AI-only and billing-only routes emit monitoring or skeleton Terraform instead and skip the baseline. - Workshop mode — After estimate,
azure-to-aws,gcp-to-aws,heroku-to-aws, andrender-to-awscan reprice region, high-availability, compute, and AWS Graviton scenarios without repeating discovery. - Credits handoff — When estimate finishes,
gcp-to-awsandheroku-to-awscan writeplan.jsonand offer a link to apply for up to $1,500 in AWS migration credits. Generate writes the same file again when it completes.azure-to-awsdoes not yet — the import contract's acceptedsourcePlatformvalues are not confirmed to include Azure.
Examples
- "Migrate us off Heroku to AWS"
- "Migrate our Render Blueprint and live services to AWS"
- "Move our Azure estate to AWS — map AKS onto EKS"
- "We want to move from GCP — what would this cost on AWS?"
- "Map our Cloud Run services onto Fargate"
- "Reprice the migration with Graviton in us-west-2"
AI Stack Migration
The llm-to-bedrock skill is a focused model and SDK rewrite. It assesses the codebase, rewrites OpenAI, Gemini, or Anthropic API call sites to Amazon Bedrock, evaluates the rewritten behavior against a golden prompt set, and delivers a ready-to-review git branch with a migration report.
A full migration requires gcp-to-aws or azure-to-aws installed alongside it. Assess runs whichever sibling is present's discover, clarify, design, and estimate phases in the same session (gcp-to-aws preferred when both are installed), and installing llm-to-bedrock on its own stops there. Access-only mode does not need either sibling: it checks the models you name, walks the access step for each, and stops, with no rewrite and no branch. Install the pair for a migration:
npx skills add aws/agent-toolkit-for-aws/plugins/aws-startup-advisor/skills --skill llm-to-bedrock --skill gcp-to-aws
(Substitute azure-to-aws if your source infra runs on Azure.)
(The --skill '*' install above already includes both.)
When the sibling's Discover captured real spend from the OpenAI Admin API, the OpenRouter API, or the Anthropic Admin API (consent-gated, read-only), llm-to-bedrock reuses that usage profile as the cost baseline instead of extrapolating from the golden prompt set.
bedrock-migration-assessment is the provider-neutral AI engine that will take over this work from gcp-to-aws (#389). Today it scans application code for AI providers, models, gateways, workloads, and agents and their tools, or accepts a provider-neutral ai-workload-profile.json (rejecting the older GCP-shaped profile by name), can measure OpenAI, Anthropic, and OpenRouter usage through their read-only admin APIs with consent, asks the AI-only Clarify questions, and writes aws-design-ai.json in the provider-neutral design contract. It stops after Design. It does not depend on any infrastructure skill.
Model mapping is compatibility-guided rather than one-to-one parity. Validate prompts, tool-calling behavior, and evaluation metrics before cutover.
Examples
- "Migrate our OpenAI calls to Bedrock"
- "Move off the Gemini API to Amazon Bedrock"
- "Which Bedrock model is closest to what we use today?"
- "Score our prompts against Bedrock before we switch"
Startup Reference Content
Two skills serve AWS-curated content rather than performing work on your account:
knowledge-base-for-startups— AWS Activate FAQ, credits guide, programs, partner offers, sample architectures, and learn articles spanning generative AI, cloud architecture, cost optimization, security, fundraising, and go-to-market. Answers come from the bundledreferences/tree.prompt-library-for-startups— Copy-paste prompts for AI coding agents (MVP scaffolding, RAG chatbot on Amazon Bedrock, security baseline evaluation, cost anomaly detection, GPU quota requests, Amazon EKS deployment, Well-Architected review) plus downloadable installable agents.
Neither skill can look up account-specific state such as your credits balance, Activate membership, or application status. Those questions belong at AWS Startups.
Examples
- "Am I eligible for AWS Activate credits?"
- "Show me a sample architecture for a RAG application"
- "Give me a prompt to set up a security baseline"
- "Which Activate provider credits apply to us?"
Operating on AWS
The operate-on-aws skill connects a startup with no dedicated DevOps or SRE engineer to AWS DevOps Agent, which investigates incidents autonomously, finds probable root cause, and reviews pull requests before they ship. DevOps Agent is metered, so the skill states cost and access up front and asks before setting anything up.
How It Works
- Detect first — Checks for an existing DevOps Agent connection and goes straight to work if there is one. Someone already set up is never walked through setup again.
- Recommend only when it fits — Looks for evidence of a live workload or active pull requests, and recommends nothing to a team with neither.
- Safe defaults with one explicit approval — Before first-time AWS DevOps Agent setup, the skill explains the two-month new-customer free trial, applicable AWS Activate credit coverage, standard pricing, potential charges, and the account-wide read access and diagnostic permissions required for investigations. It then asks one Yes/No question covering both cost and account access, and waits for Yes before creating or connecting resources. AWS credit checks use existing permissions when available and never block setup solely because billing details are inaccessible. Release-readiness reviews are offered as a separate optional choice after setup; automated verification testing and code egress stay off unless you say Yes to them. This setup does not enable Directed Actions or create elevated infrastructure-remediation permissions; findings and proposed fixes remain recommendations. Say stop during setup and the skill halts and rolls back what it created; say stop after setup and it cancels or pauses the work, keeping the setup.
- Propose, never apply — Findings and proposed fixes are shown for approval; nothing is changed in the account without an explicit decision.
The DevOps Agent MCP server is region-specific and per-user, so the skill registers it during setup rather than the plugin declaring it. If the aws-agents-for-devsecops plugin is already installed, the skill reuses its aws-devops-agent connection instead of creating a second one, and still asks for the same approval before any first-time setup.
Examples
- "Our API has been returning 502s since the last deploy — what broke?"
- "Is this PR safe to merge?"
- "Set up AWS DevOps Agent for us — what will it cost?"
- "Pause DevOps Agent, we're burning credits"
Cost Regression Checks
The cost-regression-check skill reports what an infrastructure change adds to the AWS bill on the pull request that adds it, before apply. It is for a company that pays from a fixed monthly ceiling, often funded by credits, where a new NAT gateway or database minimum can take a visible share of the budget before the product has users.
How It Works
- Two numbers, not one — Each pull request gets the fixed monthly floor the change adds at zero traffic and the change in cost per unit of your own usage, so a blended estimate cannot hide which one moved.
- Thresholds from your ceiling — Absolute thresholds derived from the monthly ceiling live in a reviewed file in the repository, with an owner, rather than a percentage of last month's bill.
- List price before credits — Every number is the public on-demand price, so the comment describes the bill that arrives when credits run out.
- Set up once, no model per pull request — The skill adds a CI job and a repository-owned cost script that runs after plan, synth, or change set. It starts advisory and blocks only after estimates are checked against the bill.
It works with Terraform, AWS CDK, and AWS CloudFormation, and defers to the aws-core plugin's skills for pricing data, IaC mechanics, and billing comparisons, so install aws-core alongside it.
Examples
- "What will this Terraform PR add to our monthly AWS bill before we apply it?"
- "Add a cost check to every pull request that flags new fixed costs against our $3,000 monthly ceiling"
- "Set up Infracost in CI so the bill owner sees the cost of each infrastructure change"
Self-Hosted LLM Batch Inference
The self-hosted-llm-batch-inference skill builds a repeatable offline inference path for a self-hosted 1B to 8B generative model, for a startup that wants GPU economics without running an ML platform.
How It Works
- One Batch job per corpus run — AWS Batch on Amazon ECS Managed Instances Spot runs vLLM on one GPU, with model weights staged in Amazon S3 and the corpus split into deterministic shards.
- Spot without babysitting — Bounded retries and resumable shards recover from Spot interruption automatically.
- Nothing idle — Sizing comes from the measured prompt and model footprint, networking avoids an idle cost floor, and GPU capacity scales in to zero when the job ends.
- Verified, not assumed — A run succeeds only after output records reconcile with input, and verification confirms the GPU capacity is gone.
It is not for online endpoints, training, embeddings, or multi-GPU inference. For moving managed-API model calls to Amazon Bedrock, use llm-to-bedrock. It defers to the aws-core plugin's skills for container, IAM, storage, and cost mechanics, so install aws-core alongside it.
Examples
- "Run nightly batch inference with Llama 3.1 8B over our document corpus on Spot GPUs"
- "Set up AWS Batch to classify a million support tickets with a self-hosted model"
- "Our offline LLM jobs keep losing work to Spot interruptions. How should we structure them?"
Supported Environments
Using the plugin in your local compute
In your local environment, configure AWS credentials and set your target region to get started.
Prerequisites
- An AWS account
- Local AWS credentials and config
- uv (for the MCP server and the locked
render-to-awsPython tools; an unfinished Render Discover halts before capture without it, later phases on a run with completed Discover output do not need it except Generate's Blueprint semantic revalidation and final report validation, and ambient Python never replaces it) - Python 3.12 (required by the locked
render-to-awsscript project; other bundled validators require Python 3) - Git (the
llm-to-bedrockflow delivers its rewrite on a branch) - The Heroku CLI, authenticated, only if you want live discovery in
heroku-to-aws - Render CLI 2.35.0, authenticated by a user-supplied
RENDER_API_KEYor an existing saved CLI token, only if you want consented live discovery or Blueprint semantic validation inrender-to-aws; the skill never starts browser login. Prefer an early-access workspace-scoped API key with read access only to the selected discovery resources. Saved CLI tokens and unscoped API keys remain supported fallbacks; secret-value and write access are never required. - The Azure CLI, authenticated, only if you want live discovery in
azure-to-aws - The gcloud CLI, authenticated, only if you want live discovery in
gcp-to-aws - Terraform, only if you want to run
fmt,init, orvalidateover generated Terraform
Authentication and Authorization
Configure AWS credentials using one of the following methods:
- AWS CLI — Run
aws configure(IAM credentials) oraws sso login(IAM Identity Center) - Environment variables — Set
AWS_ACCESS_KEY_ID,AWS_SECRET_ACCESS_KEY, andAWS_SESSION_TOKEN. See Configuring environment variables for details.
Documentation search, regional availability lookups, and the reference-content skills need no AWS credentials. Credentials are needed when a skill inspects or provisions resources in your account. The relevant IAM action namespaces are:
sts- Caller identity and account context checksbedrock,bedrock-runtime- Model availability and prompt evaluation for the Bedrock migration flowbedrock-mantle- Calls to the Bedrock Mantle endpoint (Projects, Workspaces and other Mantle-only features, and a few Mantle-only models). It is a separate IAM service prefix, so abedrock:*grant does not cover it; the Bedrock migration flow only needs it when a plan uses a Mantle pathbedrock-websearch- Bedrock Web Search tool calls, only when a plan uses Web Search. It is a separate prefix from bothbedrockandbedrock-mantle, so a custom policy limited to those two can authorize inference while the tool's permission checks fail. GrantInvokeSearchandInvokeFetch, and addExternalWebAccessonly if you enable retrieval from the live web. AWS managed policies may already include these actionsec2,rds,s3- Read-only discovery of existing compute, database, and storage resourcesiam- Reviewing roles and policies referenced by a design or security baselinecloudformation- Inspecting and deploying generated infrastructure
Start with read-only credentials for discovery, advice, and estimation, and scope write permissions to the resources your workload actually uses.
Configuration
- Set
AWS_DEFAULT_REGIONto your preferred AWS region (e.g.,us-east-1). See Configuring environment variables for details.
Customizing Skills for Your Organization
The skills in this plugin follow AWS best practices, but they are fully customizable. You can fork the repository and modify any SKILL.md to reflect your organization's standards, naming conventions, approved services, or internal tooling. Workspace-level skills take precedence over global skills, so teams can maintain their own versions without affecting other users.
Related Resources
- AWS for Startups
- AWS Activate
- Amazon Bedrock User Guide
- Amazon Bedrock AgentCore
- AWS Well-Architected Framework
- AWS Pricing Calculator
- Agent Skills open standard — Anthropic
- Agent Toolkit for AWS
License
This project is licensed under the Apache 2.0 License.