AWS Agent Toolkit hands Claude Code curated AWS skills and live API access

What the toolkit actually installs

A hands-on walkthrough published on dev.to, the first part of a series, works through AWS's Agent Toolkit inside Claude Code and lands on a clean framing: the toolkit gives a coding agent two things it normally lacks — vetted AWS knowledge, and a route to calling AWS services for real. Setup is one command, but what lands on the machine is two independent pieces.

According to the post, the first piece is Agent Skills: curated guidance on service selection, tested procedures and troubleshooting, kept as files on disk. Each skill is a directory holding a SKILL.md and usually a references folder, and the full set is open source under Apache-2.0 in AWS's agent-toolkit-for-aws repository, so it can be read before anything is installed. The second piece is the AWS MCP Server — AWS-managed, reached through a local proxy — which lets the agent call AWS APIs and search current documentation. Neither piece requires the other.

Two design details stand out. Skill descriptions state what a skill excludes as well as what it covers — the aws-storage skill rules out SQL query engines, ETL, streaming and managed databases — which stops unrelated skills being pulled into context. Skill bodies also carry behavioural rules: verify current figures instead of reciting them, cite the pricing page for any cost claim, and name any figure that could not be verified rather than guessing. The rules create an obligation to check; the server's documentation tools give the agent a way to meet it.

Setup and the traps in it

The install command is aws configure agent-toolkit --region us-east-1, and the region flag matters more than the official examples imply. The toolkit is served only from us-east-1. The author's default region was us-west-2, and without the flag the wizard ran through agent detection and the selection screen, then errored out when fetching skills.

The prerequisites are AWS CLI 2.35.0 or later (there is no self-update; the package must be reinstalled to upgrade), Node v22+ for Claude Code, and uv. That last one is the quiet trap: the dev.to author reports the setup output never mentions uv, yet the MCP proxy runs through uvx and the server will not connect without it.

The wizard detects installed agents — the run showed Claude Code, Codex, Cursor and Kiro — and installs 23 default skills spanning IaC, core services, databases, networking, storage, security, observability, messaging, SDKs and cost management, as full copies per agent rather than symlinks. Finer-grained skills exist beyond the defaults: searching via aws agent-toolkit search-skills surfaced an amazon-dynamodb skill covering access patterns, partition key and GSI selection, single- versus multi-table design and cost estimation — more depth than the domain-level aws-database skill offers.

The verification run, and where it broke

After restarting the agent, /mcp in Claude Code reported aws-mcp connected with eight tools. One silent failure mode: on first run in a project, Claude Code asks whether to enable the MCP server, and pressing Escape rejects it — the server then stays disabled with no further prompt.

The read-path test asked for July 2026 billing details grouped by service, with sorting and totalling done inside the script. The agent reached for the aws___run_script tool and displayed the full script before running it, and the arithmetic genuinely ran inside the script rather than in the model's reply. The script still failed: it assumed call_boto3 returns a wrapper containing api_calls and return_value fields, neither of which exists, so its own assertion could never pass.

Why it matters

Agent-driven cloud work needs both halves of this toolkit: knowledge that stays current, and access that is real. Pairing them, with rules that push agents toward verifying numbers rather than recalling them, targets the most common failure mode when agents discuss cloud services — confident, stale figures. Making the skills open source also makes the curation inspectable, which matters for teams deciding how much to trust an agent operating on a real account.

The rough edges are equally instructive. An undocumented dependency, a single-region constraint the examples gloss over, and a first verification script that failed on a wrong assumption about its own tooling are a reminder that generated AWS scripts still need a human reading them before they run.