Skip to content
CloudPouch Docs

Agentic FinOps - the CloudPouch MCP server

Agentic FinOps is CloudPouch’s name for letting an AI assistant answer questions about your real AWS costs from CloudPouch’s own data. It runs through a Model Context Protocol (MCP) server inside CloudPouch, on your own machine. It is disabled by default, listens only on 127.0.0.1 (port 52700 by default, which you can change in Settings), requires a locally generated token, and never changes anything in your AWS account.

Once it is connected, you ask your assistant plain-language questions — “what did I spend on RDS last month?”, “what changed between June and July?”, “what savings did the last scan find?” — and the answer comes from CloudPouch’s own data about your account instead of from a model’s general guess.

What an assistant can reach

Section titled “What an assistant can reach”

CloudPouch exposes twelve MCP tools, and they come in two kinds.

Nine data tools read CloudPouch’s local data — the cost and savings data CloudPouch has already fetched and stored on your machine. They report account status, known AWS profiles, cost summaries, month-to-month cost comparisons, savings summaries, saved Savings Scan history, and individual savings opportunities. They never call AWS themselves.

Three action tools make CloudPouch do work. They can refresh CloudPouch’s local Cost Explorer data for a profile and period, start a Savings Scan, and poll that scan. A scan started this way runs visibly in the CloudPouch window, through the same machinery as the in-app button, and the finished run lands in History like any other.

An assistant can therefore trigger data collection and change CloudPouch’s local state, which is why “read-only” is the wrong word for the MCP server as a whole. What it never does:

CloudPouch’s MCP server never changes anything in your AWS account. It never creates, modifies, or deletes an AWS resource. It never exposes your AWS credentials to the assistant, and it never builds its own AWS client. The action tools drive CloudPouch’s own data collection and nothing else. No MCP tool can apply a fix, remediate a finding, or act on an opportunity — that stays your decision, made through your own change process, exactly as it does in the CloudPouch interface.

The per-tool detail is in the MCP tools reference.

Security properties

Section titled “Security properties”
  1. Disabled by default. The MCP server does not run until you turn it on in Settings.
  2. Local only. CloudPouch listens only on this machine’s loopback interface, 127.0.0.1, so the MCP server is not reachable from another machine or over a network. Loopback binding on its own does not isolate the server from other user accounts on the same machine; the generated token is what authorizes a request.
  3. Authenticated. Requests require a locally generated token, shown masked in Settings.
  4. Protected against DNS rebinding and origin abuse.
  5. Tied to CloudPouch’s lifecycle. The server runs only while CloudPouch runs. With the CloudPouch window closed, or sitting on the Login screen, the action tools answer with a structured error telling you to open CloudPouch and sign in.
  6. No AWS mutation, no credential exposure, no separate AWS client.

Because the MCP server is served by CloudPouch itself, it adds no outbound connection to a CloudPouch server. See Desktop app vs SaaS for the rest of the local-first picture.

Where the data goes after your assistant reads it

Section titled “Where the data goes after your assistant reads it”

This is the part worth being explicit about. Connecting the MCP server makes your local AWS cost and savings data available to whatever assistant you configure — spend by service, month-to-month movements, AWS profile names, and savings opportunities whose detail names the AWS resources involved.

CloudPouch’s local-first guarantees cover CloudPouch. They do not extend to the assistant’s own handling of the data it receives. If the assistant you connect sends its context to a hosted model, your cost and savings data goes wherever that assistant sends it, under that vendor’s terms — not CloudPouch’s. Treat this as the same decision you make for any other MCP server you connect to an assistant, and pick the assistant accordingly.

Free and Pro over MCP

Section titled “Free and Pro over MCP”

The data tools follow the same Free/Pro boundary as the rest of CloudPouch, tool by tool. That boundary is enforced at a shared internal boundary rather than by hiding buttons, so an assistant on Free cannot read detail that the Free interface withholds. A Free caller of a Pro-only tool gets an explicit upgrade message rather than a bare refusal.

The three action tools sit outside that split: they work on every activated plan, Free included, in parity with the in-app buttons. A session with no activated plan gets a “not activated” answer instead.

The full per-tool table is in the MCP tools reference; for what each plan includes elsewhere in CloudPouch, see Free and Pro plans.

What a refresh over MCP costs

Section titled “What a refresh over MCP costs”

AWS bills Cost Explorer per API request, and a forced refresh through MCP is billed like any other Cost Explorer query — up to three requests per call. An unforced refresh goes through the same single-flight Cost Explorer queue the Billing view uses: if the local data is already fresh it serves cache and never pays twice. Do not assume a forced refresh is free. See Data refresh and cache for how the local cache behaves.

Turning it on

Section titled “Turning it on”

Enable the server in Settings → MCP Server, which is also reachable from the MCP Server item in the header gear menu. From there, Settings → AI assistants holds the guided setup for each supported assistant.

Step-by-step: Connect CloudPouch to an AI assistant.

Section titled “Related pages”