Costory FinOps MCP server

Costory FinOps MCP: agent skills and plugin

The Costory FinOps MCP is a hosted Model Context Protocol server that lets Claude, Cursor, VS Code, Codex, or Dust answer questions about your cloud and AI spend.

Feeding raw AWS or GCP billing lines into a prompt does not work. Costory acts as a context layer: it normalizes billing from AWS, GCP, Azure, Snowflake, Datadog, OpenAI, and Anthropic into one schema, allocates shared and untagged cost from real usage metrics, and correlates spend with deploy events and incidents. The assistant then calls structured tools against data that is already allocated and explained.

This repository holds the agent skills that sit on top of those tools: the workflows that turn "why did prod cost jump last week" into the right sequence of tool calls.

• Full MCP documentation: docs.costory.io/features/mcp
• Endpoint: https://app-api.costory.io/mcp
• Auth: OAuth, no IAM credentials, no Docker, no local server

Connect the MCP

You need a Costory workspace with billing data connected. A 15-day trial is available.

Claude Desktop / Claude Code / Cursor / VS Code: add a custom connector pointing at https://app-api.costory.io/mcp, then complete the OAuth login in the browser window that opens. Per-client walkthroughs with screenshots are in the MCP docs.

This repo ships an .mcp.json you can copy:

{
  "mcpServers": {
    "costory": {
      "type": "http",
      "url": "https://app-api.costory.io/mcp",
      "oauth": { "callbackPort": 8080 }
    }
  }
}

For clients without native remote-MCP support, proxy it with mcp-remote:

{
  "mcpServers": {
    "costory": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://app-api.costory.io/mcp"]
    }
  }
}

MCP tools reference

The server exposes tools in five groups. Names and payloads are versioned; the API reference is the source of truth.

Write tools act only inside your Costory workspace. Query scoping follows the calling user's workspace role.

FinOps skills

Skills are the workflow layer: each one encodes how to sequence the tools above for a class of question, so the assistant does not have to rediscover it.

Shortened. The full README is on GitHub.