Introducing Agent Toolkit for AWS

Added: 2026-05-09

[00:00:00]Your coding agent is smart, but AWS ships new services faster than any model's training data. And even smart models make mistakes on multi-step AWS workflows without the right guidance.

[00:00:11]They'll skip a critical AM permission or call an API that changed 6 months ago.

[00:00:17]Agent Toolkit for AWS closes [music] those gaps. Live AWS context and step-by-step guidance working with whatever coding agent you already use.

[00:00:25]Today, I want to show you what it actually [music] does. We'll deploy a Python lambda behind API gateway and you'll see the agent use the toolkit end to end. But first, some quick context on the pieces. The toolkit has four parts.

[00:00:38]The AWS MCP server is a managed endpoint your agent connects to through the model context protocol. It gives your agent access to current AWS documentation, [music] not just training data, the actual docs searched in real time. and it lets your agent make authenticated AWS API calls using your AM credentials.

[00:00:58]The calls go through the manage server.

[00:01:00]They get validated and then they all land in cloud trail. If you're on a team or if you're a platform engineer trying to figure out how to get developers to use coding agents without giving the agent full write access, there are AM condition keys that let you write policies specific for agent initiated actions. So you can give the agent readonly access through the MCP server even if the developer's own role can write. [music] I'll come back to this after the demo. The second piece is agent skills. A skill is a runbook in markdown. Step-by-step instructions for a specific AWS task written by someone who actually ran the workflow and found where agents get stuck. The Lambda plus API gateway skill, for example, includes the [music] add permission step. That's the step that grants API gateway permission to invoke the function and it's the one agents most consistently miss. Without it, the API exists and the function exists, but they can't talk with each other. Skills are compact.

[00:01:55]It's a few thousand tokens, not the tens of thousands you'd get from pasting a full docs page into context. The agent discovers skills through the MCP server, loads the one it needs, follows it, and drops it when the task is done. Third piece, plugins for Claude Code and Codeex. A plug-in bundles the MCP server config and a set of skills into one install. For every other agent, cursor, Kira, Windsurf client, Claude Code, or any other MCP compatible client, you add the AWS MCP server to your MCP config file [music] directly. And the fourth piece is rules files. Project level config that tells your agent how to use AWS in your project. Things like use the MCP server for API calls or search for a skill before starting a task. You drop them in your repo and they apply to every session. The toolkit itself is free. You just pay for whatever AWS resources the agent creates. All right, let's watch this work. The task deploy a Python lambda behind an API gateway HTTP API. Return hello on get/ US East1.

[00:02:57]First thing the agent does searches for relevant skills and documentation through the MCP server, not its training data, live [music] AWS context. It finds a skill called connecting Lambda to API gateway and loads it. Now the agent has a tested playbook for this exact deployment. One thing I like about how this works, the agent isn't just mechanically following the skill. It notices the skill is written for REST [music] APIs. And since we asked for an HTTP API, which is newer, simpler, and cheaper, it adapts. The skill gives structure, but the agent handles the judgment. From there, it works through the steps. It creates the lambda function code locally and zips it.

[00:03:36][music] Creates the AM execution role with the basic execution policy. Waits for AM propagation. Then uploads the zip and creates the Lambda function. After that, it creates the HTTP API with the Lambda integration wired in. And here's that add permission step I mentioned earlier, the one that agents often skip.

[00:03:56]The skill has it, so the agent runs it.

[00:03:59]API gateway now has permission to invoke the function. [music] Then one more refinement. The agent notices the API was created with a catchall route and scopes it down to just get/ikely asked.

[00:04:10][music] Call the endpoint. Hello. So it was just 2 minutes and 35 seconds from prompt to working endpoint. [music] Each one of the API calls ran through the MCP server on my AM credentials. Everyone is in cloud trail. And if I've had a policy restricting what agent initiated requests can do, those restrictions would have applied to every step. The skill is the part that stands [music] out to me. The agent didn't just follow it rigidly. It adapted it for HTTP APIs, then caught the catch all route and scoped it down. But it also didn't miss add permission, which is the step agents consistently skip when they're working from training data alone. That combination of structure and judgment is what makes this different from just pasting docs into context. [music] And everything was auditable. Every call, every parameter is in cloud trail with my AM identity attached. The MCP server runs in two regions right now. [music] US East1 and EU Central 1. That's where the server lives, but your agent can operate on resources in any [music] region. A couple of quotas worth knowing. You get 27 concurrent [music] connections per account per region, and that's a hard cap. Three requests per second [music] sustained for one person.

[00:05:21]That's fine. If you're rolling this out to a larger team, watch the Cloudatch usage metrics. The 27 connection limit will surprise you faster than you'd expect. And on the team note, my recommendation is to start read only.

[00:05:34]Use the AM condition key to block write operations through the MCP server. Let people get comfortable with the agent searching docs and loading skills before you open up right access. You can always loosen it later, but going the other direction is always harder. Installing takes just a minute. There are two parts depending on your agent. If you're on Claude Code or Codeex, it's a couple of commands to install the plugin. For every other agent, Cursor, Kira, Windinssurf client, Claude desktop, or any other MCP compatible client, add the AWS MCP server to your MCP config file.

[00:06:08][music] It's the same server just configured directly. Either way, you need AWS credentials configured on your machine. [music] The agent authenticates with your existing AM identity. AWS login or AWS configure SSO both work.

[00:06:22]Links to the product page and the docs are in the description. Try it with something you've been meaning to deploy and see how your agent behaves when it's got current docs [music] and a skill to follow. See you next time.