Show HN: An MCP server for understanding AWS costs | Hacker News Reader

Show HN: An MCP server for understanding AWS costs

119 pointsStratusBen23 comments

Hey all - I work at Vantage, a FinOps platform.

I know AI is peak hype right now. But it has definitely changed some of our dev workflows already. So we wanted to find a way to let our customers experiment with how they can use AI to make their cloud cost management work more productive.

The MCP Server acts as a connector between LLMs (right now only Claude, Cursor support it but ChatGPT and Google Gemini coming soon) and your cost and usage data on Vantage which supports 20+ cloud infra providers including AWS, Datadog, Mongo, etc. (You have to have a Vantage account to use it since it's using the Vantage API)

Video demo: https://www.youtube.com/watch?v=n0VP2NlUvRU

Repo: https://github.com/vantage-sh/vantage-mcp-server

It's really impressive how capable the latest-gen models are with an MCP server and an API. So far we have found it useful for:

Ad-Hoc questions: "What's our non-prod cloud spend per engineer if we have 25 engineers"

Action plans: "Find unallocated spend and look for clues how it should be tagged"

Multi-tool workflows: "Find recent cost spikes that look like they could have come from eng changes and look for GitHub PR's merged around the same time" (using it in combination with the GitHub MCP)

Thought I'd share, let me know if you have questions.

AWS have also made their own cost analysis MCP server https://github.com/awslabs/mcp/tree/main/src/cost-analysis-m...

What a great solution. They can leverage this to avoid liability for their own price quotes and solutions while simultaneously adding another layer of vender lock-in. synergy

How does this avoid liability for anything?

I don't think anyone's really stated this outright, but large companies must believe they're not liable for anything their models/AI products are producing. That must be the case for their business model to work.

A better question is how can they be held liable at all.

How does this work? https://github.com/vantage-sh/vantage-mcp-server?tab=License...

That is extra weird when thinking about the audience who might be Vantage.sh users (and thus have the ability to create the read-only token mentioned elsewhere) but would almost certainly be using it from their workstation, in a commercial context. Sounds like you're trying to keep someone from selling your MCP toy and decided to be cute with the licensing text

I'm just trying to understand licenses, but doesn't the choice of MIT contradict the inital "non-commercial purposes" as MIT says 'including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software' - Therefore, the non-commercial purposes is actually void and I can use the software to the limits of MIT defines? And because it is already MIT, they can relicense only future software but not this piece anymore?

IANAL but that is my understanding. I created an issue to get some clarification - https://github.com/vantage-sh/vantage-mcp-server/issues/31

The issue was initially closed with the following comment:

> Hey Dave, it is as it is stated. The MCP is published with dual licenses depending on your intent. > https://fossa.com/blog/dual-licensing-models-explained/

After further discussion on the ticket the license is now just MIT.

You have a couple problems that I can see:

One is the MIT license does not prohibit selling. And wrapping it in a "for non-commercial uses" clause creates a contradiction difficult, if not impossible to enforce.

So if I want to use the software I just have to create a fork on my home machine for non-commercial purposes, update the license to MIT only, and then the fork is mine to do with as I want commercially? What's even the point of this license?

On a related note, I'm not sure when it became "ok" to leave production credentials scattered across your system in configuration files. So many MCP server examples encourage this pattern, and inevitably, it's going to cause trouble at some point.

What is your preferred way to manage them?

The 1pass CLI is great! However if you aren’t using 1password as your secrets vault, I’m building an open source, vault-agnostic alternative called RunSecret (https://github.com/runsecret/rsec)

You may want to do your own Show HN about it, so folks don't have to be "MCP curious" to find out that it exists

That said, given https://github.com/runsecret/rsec#aws-secrets-manager presumably in order to keep AWS credentials off disk one would then have to have this?

    "vantage-mcp-server": {
      "command": "/opt/homebrew/bin/aws-vault",
      "args": [
      "exec", "--region=us-east-1", "my-awesome-profile",
      "--", "/opt/homebrew/bin/rsec", "run",
      "--", "/opt/homebrew/bin/vantage-mcp-server"
      ],
      "env": {"VANTAGE_BEARER_TOKEN":  "rsec://012345678912/sm.aws/VantageBearerToken?region=us-east-1"}
    }

in contrast to the op binary that is just one level of indirection, since they already handshake with the desktop app for $(op login) purposes

This is great feedback, thank you!

I agree RunSecret adds a level of indirection at this stage that op doesn’t (if you are using 1pass). This is something I plan to polish up once more vaults are supported. You’ve given me some ideas on how to do that here.

And thanks for the advice on doing a Show HN, planning to do so once a few more rough edges are smoothed out.

Vantage engineer who worked on this feature here. The security posture of MCP servers is still in its early stages (see “The ‘S’ in MCP Stands for Security” from three weeks ago [https://elenacross7.medium.com/%EF%B8%8F-the-s-in-mcp-stands...]). The recommendations above to use something like the 1Password CLI wrapper when invoking an MCP server seem sound.

That being said, an easier-to-distribute user experience would be to leverage short-lived OAuth tokens that LLM clients such as Claude or Goose ultimately manage for the user. We’re exploring these avenues as we develop the server.

If your service is running in AWS, use the native IAM identity loading from ECS, EKS, EC2, etc. If it's your laptop, set AWS_PROFILE and let the SDK load the temporary creds from ~/.aws.

If you really really really need to use static creds on your laptop, use aws-vault to export them, or ephemeral creds generated from them, into your environment.

1Password’s CLI op does a reasonably good job of this

What's the difference between connecting an LLM to the data through Vantage vs directly to the AWS cost and usage API's?

A few things.

The biggest is giving the LLM context. On Vantage we have a primitive called a "Cost Report" that you can think of as being a set of filters. So you can create a cost report for a particular environment (production vs staging) or by service (front-end service vs back-end service). When you ask questions to the LLM, it will take the context into account versus just looking at all of the raw usage in your account.

Most of our customers will create these filters, define reports, and organize them into folders and the LLM takes that context into account which can be helpful for asking questions.

Lastly, we support more providers beyond AWS so if you wanted to merge in other associated costs like Datadog, Temporal, Clickhouse, etc.

This is going to different, as resources end up getting intertwined? or is there a way to standardize it?

This is dangrous, as I've long held that inscrutible cloud billing was one of the greatest protections we had against a runaway superintelligent AI.

Now we only have poor IAM UX to fall back on.

/s

Reply on news.ycombinator.com