cloud-finops-mcp

MCP server exposing the OptimNow Cloud FinOps skill (reference library + named-pattern playbooks) as queryable tools for any MCP-aware client (Claude Code, Cursor, Codex CLI, Windsurf, Aider, Cline, etc.).

The skill itself ships in canonical Claude Agent-Skills format and is also installable via the cross-tool installer (./install.sh) for direct context injection. This MCP server is the enrichment path: instead of loading the full skill into the prompt, the agent calls tools to discover, filter, and fetch only what it needs.

What the server exposes

Six tools, all read-only, split across two surfaces.

References - long-form provider and discipline files (~300-500 lines each):

The reference faceted query supports any combination of:

• domain - FinOps Framework domain (e.g. Optimize Usage & Cost, Quantify Business Value)
• capability - FinOps capability (matches both primary and secondary)
• phase - Inform, Optimize, Operate
• persona - matches both primary and collaborating personas
• maturity - Crawl, Walk, Run

Playbooks - small named-pattern runbooks (~80-130 lines each):

The playbook faceted query supports:

• scope - aws, azure, gcp, or cross-cloud
• service - provider service (e.g. AWS NAT Gateway); exact-match
• waste_category - orphaned, idle, overprovisioned, commitment-mismatch, schedule-blindness, modernization, ai-ml-inefficiency, egress
• confidence - obvious, likely, possible (OptimNow three-tier model)

All filters across both surfaces AND together. String matches are case-insensitive and exact (no substring matching).

When to use which surface:

• A playbook answers "how do I detect/fix this specific pattern?" (zombie NAT, snapshot sprawl, idle ELB). It includes problem statement, symptoms, a detection query (CUR / KQL / BigQuery SQL / CLI), fix steps, and the anti-pattern.
• A reference answers anything broader: billing mechanics, commitment strategy, allocation methodology, persona-specific framings, or cross-pattern reasoning.

Install

pip install cloud-finops-mcp

Or run without installing via uv:

uvx cloud-finops-mcp

Configure your MCP client

After install, point your client at the cloud-finops-mcp console script.

Claude Code

Project-level (.mcp.json at the repo root) or user-level (~/.claude/mcp.json):

{
  "mcpServers": {
    "cloud-finops": {
      "command": "cloud-finops-mcp"
    }
  }
}

Restart Claude Code, then run /mcp to confirm the server is connected.

Cursor

~/.cursor/mcp.json:

{
  "mcpServers": {
    "cloud-finops": {
      "command": "cloud-finops-mcp"
    }
  }
}

Codex CLI

~/.codex/config.toml:

[mcp_servers.cloud-finops]
command = "cloud-finops-mcp"

Windsurf

~/.windsurf/mcp.json:

{
  "mcpServers": {
    "cloud-finops": {
      "command": "cloud-finops-mcp"
    }
  }
}

Any other MCP client

The server speaks MCP over stdio. Point any compatible client at cloud-finops-mcp (or python -m cloud_finops_mcp).

Example tool calls

Agent prompt: "Use the cloud-finops MCP to find references for the Optimize phase aimed at Engineering."

Calls find_references(phase="Optimize", persona="Engineering") and gets back the filtered subset (AWS, Azure, GCP, Bedrock, Databricks, etc.) without loading the full skill into the prompt.

Agent prompt: "Pull the AWS reference."

Calls get_reference(name="finops-aws") and gets back the full markdown body (~300 lines) instead of the entire knowledge base.

Agent prompt: "Show me the obvious-confidence AWS waste playbooks."

Calls find_playbooks(scope="aws", confidence="obvious") and gets back the list of high-signal AWS patterns (zombie NAT gateway, orphaned EBS volumes, etc.).

Agent prompt: "Walk me through the zombie NAT gateway pattern."

Calls get_playbook(name="aws-zombie-nat-gateway") and gets back the ~90-line runbook (problem, symptoms, detection query, fix, anti-pattern, see-also).

When to use this vs the installer

The two paths are complementary. You can install both.

Development

git clone https://github.com/OptimNow/cloud-finops-skills.git
cd cloud-finops-skills/mcp_server
python scripts/sync_references.py        # populate src/cloud_finops_mcp/data/
pip install -e ".[dev]"
pytest

Versioning

The PyPI package version tracks the skill release. The trigger is a .claude-plugin/plugin.json version bump reaching main, not a hand-cut tag: the auto-tag-on-plugin-bump workflow reads the new version, creates the matching vX.Y.Z tag, and publishes both the skill release zip and a new cloud-finops-mcp wheel, so the bundled references match what the rest of the repo ships. Versions must be full three-part semver (v1.27.0); the workflow rejects anything else.

Because a plugin.json bump publishes, content PRs never touch it. Version bumps live in dedicated release PRs that move plugin.json, .claude-plugin/marketplace.json metadata.version, and this package's pyproject.toml together. See the release-train rule in the repo's CLAUDE.md.

License

CC BY-SA 4.0 - same as the parent skill. Credit OptimNow.