MCP, skill or Claude alone for your FinOps advice?

il y a 10 minutes

4 min de lecture

On Claude, install the Cloud FinOps plugin, which loads the skill. In my November 2026 test it answered from the library on 16 of 16 questions with Claude Opus 5.5 and on 13 of 16 with Claude Sonnet 5.5. Use the MCP (Model Context Protocol) connector for an assistant or an agent that cannot load a Claude skill: alone, it scored 15 and 7. Claude alone gave sound advice on a general question, and on an Azure renewal question it missed a price rise that the skill reported.

On 11 November 2026 I put one question to Claude Opus 5.5 twice: "Our AWS bill has NAT gateway charges but I do not think anything uses them - what should I do?" The first time Claude answered alone. The second time it had the Cloud FinOps skill, the open library of guides and runbooks that OptimNow publishes and that installs in Claude as a plugin.

The NAT gateway question, alone and then with the plugin (46 seconds).

Claude alone and the skill on a NAT gateway question: 4 differences

Claude alone gave sound advice. With the skill, Claude read the runbook first and the answer differs in 4 places:

one Athena query that checks every gateway in one pass, where Claude alone starts in Cost Explorer
2 months of observation, where Claude alone asked for 30 days
a price that carries its date and its source
a sentence saying that it cannot see the account

Each version ran once: an example, not a measurement.

The same NAT gateway question. Claude alone: read the bill line items in Cost Explorer. With the Cloud FinOps plugin: one Athena query that checks every gateway.

On an Azure renewal, Claude alone missed a price rise

The second question concerned a renewal: "We renew our Azure reservations in January for VMs in Dublin and Amsterdam. What should we know before we sign?" Claude alone searched the web and reported one change on 1 February 2027, the end of reservation exchanges. With the skill, it reported 2 changes on that date. The second is a regional price rise: 17% in North Europe (Dublin) and 9% in West Europe (Amsterdam). Both runs of each version agreed. Claude alone did cite the price rise the same day, when the question was a broader one about 2027.

The Azure renewal question, alone and then with the plugin (46 seconds).

The same Azure renewal question. Claude alone advises treating the renewal as a new purchase at current prices. With the Cloud FinOps plugin, the answer gives the regional price rise of 1 February 2027: 17% in North Europe and 9% in West Europe.

The MCP connector: the same library, fetched on demand

MCP (Model Context Protocol) is the protocol through which an assistant calls outside tools. The Cloud FinOps connector serves the same library that way, one guide or runbook per call, and the model has to decide to make the call. In August 2026 the model skipped that call on symptom questions: Skills inject, MCP tools decide reports the 33 questions behind that finding. Since then, version 1.50.0 added a search on the user's own sentence and version 1.51.0 removed the interactive views that kept Claude from calling the search tools.

I tested the connector again on claude.ai on 10 and 11 November 2026, with 16 questions from the hard end of my test set, one run each. Through the connector alone, Claude Opus 5.5 answered 15 from the library and Claude Sonnet 5.5 answered 7. The skill scored 16 and 13. The August post left one hypothesis open: with both installed, the connector might take over from the skill. It did not: Claude did not call the connector once in 34 chats. The method and the results are in the repository's documentation.

Bar chart: answers drawn from the library, of 16 questions. Claude Opus 5.5: skill 16, MCP connector alone 15. Claude Sonnet 5.5: skill 13, MCP connector alone 7. With both installed, Claude called the connector in 0 of 34 chats.

Plugin or MCP connector: what to install

On Claude, install the plugin and stop there: adding the connector changed nothing in my test. Use the connector for an assistant or an agent that cannot load a Claude skill. Claude alone remains a good first answer to a general question. These results come from 2 questions and one 16-question test, all on claude.ai, and they say nothing about other assistants.

Before your next renewal, ask the question twice, with the plugin off and then on, and compare the 2 answers line by line. Each of mine took under a minute.

Frequently asked questions

Is Claude alone good enough for FinOps advice?

On a general question it was in my example: asked about NAT gateway charges that nothing seemed to use, Claude Opus 5.5 alone gave sound advice. On a question that turns on a dated fact it was not: asked about renewing Azure reservations in Dublin and Amsterdam, it did not mention the regional price rise of 1 February 2027, in 2 runs of 2. These are 2 questions, not a benchmark.

Should I install both the Cloud FinOps plugin and the MCP connector in Claude?

The plugin is enough. On claude.ai on 10 November 2026, with both installed, Claude did not call the connector in any of 34 chats, and every answer that used the library took it from the skill. Adding the connector changed nothing.

How does the Cloud FinOps MCP connector alone compare with the skill?

Alone on claude.ai, the connector answered from the library on 15 of 16 questions with Claude Opus 5.5 and on 7 of 16 with Claude Sonnet 5.5, against 16 and 13 for the skill (one run per question, 10 and 11 November 2026). The connector depends on the model deciding to call a tool, so its result varies with the model.

Who is the Cloud FinOps MCP connector for?

For an assistant or an agent that cannot load a Claude skill, and for sessions where the model should search the library one guide at a time. It is listed in Claude's connector directory and it has no access to your cloud accounts.

The library is free and open (CC BY-SA 4.0):