Two writes happen when you run `aws configure agent-toolkit`. One drops skill directories into per-agent paths on disk [13][15]. The other configures the MCP server connection for the agent, which only takes effect after a restart [14]. The failure signatures do not overlap, which is the practical reason to know which half you are looking at.
A skill is inert text. It is a directory holding `SKILL.md`, an `.aws-skill-metadata` file carrying a version, and usually a `references/` folder [3]. Routing happens on the front matter `description`, which also states what the skill is not for so unrelated skills stay unloaded: `aws-storage` rules itself out for Athena, Spark, Redshift, EMR, Glue, Kafka, MSK, Kinesis, RDS, Aurora and DynamoDB [4][5]. A skill that never fires is therefore a text-matching problem in a file you can open, and the set is Apache-2.0 in `aws/agent-toolkit-for-aws` [7].
The server half is a process problem. The proxy runs through `uvx`, so `uv` is a hard dependency, and nothing in the setup output says so [8][9]. `/mcp` is the check, and a healthy answer reads `aws-mcp - connected - 8 tools` [18]. Eight is a count of connected tools, and what those tools are permitted to do against a live account goes unstated; the walkthrough's first real task against the account is a billing read for July 2026 grouped by service [19].
The halves are built to lean on each other. One behavioural rule in `aws-storage` forbids recalling figures, requires a pricing-page citation for any cost claim, and requires the agent to name a value it could not verify rather than guess [6]. The walkthrough describes that rule as pairing with the server's documentation tools: the skill demands verification, the server supplies the means [22]. Press Esc at the first-run prompt and the server stays disabled with no further prompting [17], while the skill still loads and still demands verification. By the rule as written, the agent should then declare the unverified value instead of inventing one [6]. The text requires that; the walkthrough never runs the case.
The install writes a full copy per agent rather than symlinks [15]. On that machine the wizard detected four agents, preselected everything, and offered 23 default skills [13][14], so accepting the defaults puts 92 skill directories on one laptop [20]. Each copy carries its own version stamp [3]. A fix you read in the repo is a fix you install four times.
Those numbers are one run: macOS, Claude Code, aws-cli 2.36.40 [21]. 23 is the size of the `aws-core` category at that moment [15], and it stops describing your install the first time you run `search-skills` and add something finer, like `amazon-dynamodb` for access-pattern and GSI work [16]. The region pin transfers regardless. Agent Toolkit is served from us-east-1 only, official examples omit the `--region` flag, and a us-west-2 default carries you through agent detection and the selection screen before dying at the fetch with `AgentToolkit is only available in us-east-1` [11][12]. A box on CLI 2.34 needs a full package reinstall before any of that, because there is no self-update [10]. So when the agent gets AWS wrong, the first question is which half you are debugging: a description that did not match, or a proxy that never started.
