AWS Agent Toolkit: install the rules file
26/6/2026 · 10:22

AWS Agent Toolkit: install the rules file

AWS’s official toolkit gives coding agents AWS MCP access, curated skills, plugins, and behavior rules — useful for real AWS work, but still early around Windows, profiles, Terraform, and ARM64 support.

Today's Trending Agent Skills
Today's Trending Agent Skills@NeoDrop Official

Vistazo a la investigación

Install aws/agent-toolkit-for-aws if your coding agent touches real AWS architecture. Skip it if you only need generic cloud explanations, or if your current workflow depends on Windows non-English setups, local AWS profiles, Terraform-first guidance, or ARM64/Graviton remote execution.
The reason to look at it today is not just the star count. The repository reached 1.3K stars, 119 forks, 94 commits, and 0 releases in the June 26 snapshot, after gaining about 200 stars in roughly 24 hours. 1 The sharper signal is that AWS packaged the toolkit as a full stack: a managed AWS MCP Server, curated skills, plugins, and a rules file that tells the agent when to use them. 1
Cargando tarjeta de contenido…

Verdict

Install today if: you use Claude Code, Codex, Cursor, Kiro, or an MCP-capable agent for AWS design, deployment, operations, data engineering, or security review. AWS says the toolkit helps AI coding agents "build, deploy, and manage applications on AWS," and the current package includes AWS-supported MCP servers, skills, plugins, and rules. 1
Do not install it as a magic AWS brain. The best part is the behavior layer. The rules file tells the agent to prefer the AWS MCP Server for AWS interactions, check for relevant skills before starting, verify documentation when uncertain, follow the Well-Architected Framework, and resolve secrets through Secrets Manager rather than direct API calls. 2
Esin Saribudak, an AWS technical educator writing in TheNewStack, put the core issue plainly: 3
"There's a difference between 'can discover skills' and 'will proactively load them before answering.' The rules file bridges that gap."
That is the install decision. If you add only tools and skills, your agent may still answer from old model memory. If you add the rules file, the agent has an explicit instruction to look for AWS-specific guidance before it writes code or architecture advice.

What you get

The toolkit has three layers. The AWS MCP Server is a managed service with access to 300+ AWS service APIs through a single authenticated endpoint, sandboxed Python execution, documentation search, CloudWatch metrics, IAM condition keys, and CloudTrail audit logs. 1 The skills layer is split into core skills and specialized skills; the specialized catalog covers Analytics, Database, EC2, Migration & Modernization, Networking & Content Delivery, Operations, Security & Identity, Serverless, Storage, and Web & Mobile development. 4 The rules layer is the small file that changes when the agent reaches for those resources. 2
The plugin set is currently four buckets: aws-core for core AWS development patterns including CDK, CloudFormation, serverless, containers, storage, observability, billing, SDK use, and deployment; aws-agents for Amazon Bedrock AgentCore; aws-data-analytics for S3 Tables, Glue, Athena, data lakes, and ETL; and aws-agents-for-devsecops for DevOps and security-agent workflows such as incident investigation, code review, vulnerability scanning, and penetration testing. 5
AWS also positions Agent Toolkit for AWS as the successor to the older awslabs/agent-plugins project. AWS says the older Labs project will keep accepting contributions, but recommends the new toolkit because it adds IAM condition keys, CloudWatch and CloudTrail observability, and validated workflows. 1

Install path

The cleanest install path depends on your agent client. AWS documents these entry points: Claude Code can install aws-core through /plugin install aws-core@claude-plugins-official; Codex can add the marketplace with codex plugin marketplace add aws/agent-toolkit-for-aws and then use /plugins; Cursor can import the marketplace from Settings → Plugins → Team Marketplaces; Kiro uses MCP Server JSON plus npx skills add; other compatible agents can install the skills with npx skills add aws/agent-toolkit-for-aws/skills. 1
For direct skills installation, the reusable command is:
npx skills add aws/agent-toolkit-for-aws/skills
For Codex plugin setup, start here:
codex plugin marketplace add aws/agent-toolkit-for-aws
For Claude Code, install the core AWS plugin from the official plugin marketplace:
/plugin install aws-core@claude-plugins-official
DevelopersIO tested the AWS CLI integration and found that aws agent-toolkit is available from AWS CLI v2.35.9, with commands for skill installation, listing, search, update, and removal. 6 DevelopersIO also found the CLI detects 10 agents by checking local configuration directories: Claude Code, Cline, Codex, Cursor, Gemini CLI, Kiro, OpenClaw, OpenCode, Pi, and Windsurf. 6

A useful first prompt

After install, test whether the rules file is actually being loaded. A practical first task is narrow enough to expose whether the agent uses AWS-specific decision tables rather than generic cloud advice:
Use the AWS Agent Toolkit rules and relevant AWS skills before answering.
Design a small authenticated CRUD API on AWS.
Compare HTTP API vs REST API, pick one, and explain the Cognito authorizer choice.
Keep the answer deployable with CDK or CloudFormation.
That example maps to Saribudak's TheNewStack test: without the rules file, the agent gave a generic API Gateway, Lambda, DynamoDB, and Cognito answer; with the 17-line rules file, the agent loaded skills first and became more specific about HTTP API, JWT authorizer with Cognito, and the "one function per route" pattern. 3
For data work, a second test should route through the data plugin:
Use the AWS Agent Toolkit data analytics skills.
Design an S3 Tables + Glue + Athena workflow for daily event data.
Show the IAM boundaries and where CloudTrail should capture agent actions.
The aws-data-analytics plugin is the relevant package for S3 Tables, Glue, Athena, data lakes, and ETL workflows. 5

Feedback so far

Public feedback is still early and mostly comes from hands-on articles rather than large user threads. Thomas Reid, a data engineer writing in Towards AWS, described the toolkit as "your own personal expert AWS solutions architect and data engineer rolled into one." 7 That is the optimistic read: the toolkit compresses architecture guidance, data-engineering patterns, and operational guardrails into something a coding agent can load on demand.
The more useful critical feedback came from DevelopersIO. Shibata found the AWS CLI command useful, but reported a CP932 encoding error in a Japanese Windows environment because the CLI opened UTF-8 agent configuration files with open() rather than AWS CLI's internal compat_open() helper. 6 DevelopersIO also reported that skill management is effectively tied to us-east-1, and that MCP server configuration did not work on a non-English Windows setup. 6

Known limitations

The GitHub issue list had 8 open issues in the June 26 snapshot. 8 The ones that should affect install decisions are concrete:
  • The aws-cloudformation skill used an invalid --change-set-id flag where --change-set-name was expected in issue #83. 8
  • Transform custom Remote execution did not support ARM64/Graviton because the prebuilt image and Batch job definition were x86_64-only in issue #64. 8
  • aws-core had confusion around --metadata AWS_REGION and local region semantics for mcp-proxy-for-aws credential refresh in issue #63. 8
  • Skill descriptions were reported as consuming too much context budget in issue #54. 8
  • Local AWS profile support through --profile or AWS_PROFILE was not supported in issue #39. 8
  • Terraform support was still a request, not a current core skill, in issue #115. 8
The license is Apache 2.0. The repository language mix in the June 26 snapshot was Python 67.1%, TypeScript 15.0%, Shell 14.3%, Java 1.8%, and JavaScript 1.8%. 1

When to install / when to skip

Install it if your agent regularly designs AWS infrastructure, writes CDK or CloudFormation, works with Bedrock AgentCore, builds Glue/Athena data pipelines, or performs AWS security and operations reviews. The toolkit now has AWS-backed guidance, a managed MCP layer, four plugin areas, and an explicit rules file that tells the agent to load AWS-specific context before answering. 1 2
Skip it for now if your workflow is Terraform-first, depends on local AWS profiles, runs on ARM64/Graviton remote execution, or needs smooth non-English Windows support. Those are not theoretical cautions; they are either open GitHub issues or DevelopersIO's reported test results. 8 6

Quick reference

Repositoryaws/agent-toolkit-for-aws 1
MaintainerAWS 1
LicenseApache 2.0 1
Snapshot metrics1.3K stars · 119 forks · 94 commits · 0 releases 1
Supported agentsClaude Code, Codex, Cursor, Kiro, and other MCP-capable agents 1
CLI detection listClaude Code, Cline, Codex, Cursor, Gemini CLI, Kiro, OpenClaw, OpenCode, Pi, Windsurf 6
Main install commandnpx skills add aws/agent-toolkit-for-aws/skills 1
Best first stepInstall the relevant plugin or skills, then add the rules file so the agent checks AWS skills before answering. 2
Cover image: GitHub OpenGraph card from aws/agent-toolkit-for-aws.

Fuentes de referencia

  1. 1AWS Agent Toolkit for AWS — GitHub README
  2. 2aws-agent-rules.md
  3. 3TheNewStack — AWS Agent Toolkit rules file
  4. 4agent-toolkit-for-aws skills directory
  5. 5agent-toolkit-for-aws plugins directory
  6. 6DevelopersIO — AWS CLI Agent Toolkit integration test
  7. 7Towards AWS — An introduction to the Agent Toolkit for AWS
  8. 8AWS Agent Toolkit for AWS — GitHub issues

Contenido relacionado

Añade más opiniones o contexto en torno a este contenido.

  • Inicia sesión para comentar.