Skip to main content
Glama
cloudsealed-mcp

cloudsealed-mcp

Official

cloudsealed-mcp

MCP server that gives AI agents (Claude Code, Claude Desktop, Cursor, etc.) direct access to three deterministic CloudSealed analysis tools:

  • cloudsealed_analyze_billing_waste — cost anomaly detection over a cloud billing export (AWS/GCP/Azure/generic), using cloudsealed-jit's rolling-median + MAD baseline. Runs locally, no network call.

  • cloudsealed_score_architecture_risk — deterministic, auditable architecture risk scoring (single point of failure, excessive coupling, scalability gap) from a declared system inventory, backed by Predictive-ML-Core.

  • cloudsealed_correlate_cost_and_risk — runs both engines and cross-references them, ranking systems that are both costly and high architecture risk ("double jeopardy"). No cloud-native tool does this: cost anomaly detection and architecture review are separate products even within one cloud, so nothing tells you "this cost spike is on a system that's also a single point of failure." Vendor-neutral, runs on data you already exported.

All tools are read-only: they never write files, and the only network call any of them makes is to the Predictive-ML-Core service you point it at.

License: MIT

Configuração Mágica para IAs (Claude Desktop & Cursor)

O verdadeiro poder do CloudSealed MCP é usá-lo dentro das suas ferramentas de IA do dia-a-dia.

No Cursor IDE

Vá em Settings > Features > MCP e clique em + Add New MCP Server.

  • Type: command

  • Command: uvx cloudsealed-mcp Isso fará o Cursor analisar a arquitetura do seu código em C#/Python conversando com você no chat!

No Claude Desktop

Adicione ao seu arquivo claude_desktop_config.json:

{
  "mcpServers": {
    "cloudsealed": {
      "command": "uvx",
      "args": ["cloudsealed-mcp"]
    }
  }
}

Install

# recommended: no local install, uvx fetches and runs it on demand
uvx cloudsealed-mcp

# or, from source until this is published to PyPI
pip install git+https://github.com/cloudsealed/cloudsealed-mcp

Related MCP server: Azure FinOps Elite

Configure your MCP client

Add to your client's MCP config (.mcp.json for Claude Code, claude_desktop_config.json for Claude Desktop, Cursor's MCP settings, etc.):

{
  "mcpServers": {
    "cloudsealed": {
      "command": "uvx",
      "args": ["cloudsealed-mcp"]
    }
  }
}

Restart the client, and both tools become available to the agent.

cloudsealed_score_architecture_risk needs a running Predictive-ML-Core

cloudsealed_analyze_billing_waste works out of the box — the analysis engine is a pure Python dependency, no server involved.

cloudsealed_score_architecture_risk calls the Predictive-ML-Core HTTP API. By default it looks for one at http://localhost:8092. Start one with:

docker run -p 8092:8092 cloudsealed/predictive-ml-core

To point at a different deployment (self-hosted or otherwise), set:

export PREDICTIVE_ML_CORE_URL="https://your-deployment"
export PREDICTIVE_ML_CORE_API_KEY="..."   # only if that deployment requires one

Example prompts

  • "Here's our AWS Cost and Usage Report for last month — find the cost anomalies and tell me what to fix first." (paste the CSV; the agent calls cloudsealed_analyze_billing_waste)

  • "We have a checkout-api (CRITICAL, public-facing, no declared auth), an orders-db (CRITICAL), and a third-party payment-gateway. What's our biggest architecture risk?" (the agent calls cloudsealed_score_architecture_risk)

  • "Here's our billing export and our system inventory — which service is both burning money and a reliability risk? Prioritize our cloud work by cost AND risk together." (the agent calls cloudsealed_correlate_cost_and_risk; name systems to match billing service names to get the linked view)

Why deterministic engines, not another LLM call

Both underlying engines score with explicit, auditable rules — not a model. Every anomaly and every risk score traces back to a specific rule and a stated rationale (see JIT's METHODOLOGY.md and Predictive-ML-Core's METHODOLOGY.md). That means an agent calling these tools gets a reproducible, explainable answer instead of a second opinion from another LLM.

FAQ

What's an MCP server, and why not just ask the agent to write the analysis code itself? MCP (Model Context Protocol) lets an agent call a real tool instead of generating code from scratch each time. The underlying engines here are deterministic and already tested/benchmarked — an agent calling them gets a reproducible answer, not a fresh guess with its own bugs.

Which clients support this? Any MCP client: Claude Code, Claude Desktop, Cursor, Windsurf, and others that read a standard mcpServers config block.

Is there a hosted version, or do I need to run anything locally? cloudsealed_analyze_billing_waste needs nothing but the uvx invocation. cloudsealed_score_architecture_risk needs a Predictive-ML-Core instance reachable at PREDICTIVE_ML_CORE_URL — self-hosted via Docker by default (see above), or your own deployment.

Does this send my billing data anywhere? No — the billing analysis runs entirely in-process. Only the architecture tool makes a network call, and only to whichever PREDICTIVE_ML_CORE_URL you configure.

Development

pip install -e ".[dev]"
python -m py_compile src/cloudsealed_mcp/server.py

License

MIT. See LICENSE.

Available Tools

2 tools
cloudsealed_analyze_billing_wasteA
Read-onlyIdempotent

Detect cost anomalies in a cloud billing export (AWS/GCP/Azure/generic).

Models the expected daily spend for each day as a rolling-median baseline times a day-of-week factor, then flags days whose actual spend deviates from that baseline by a robust (median-absolute-deviation-based) modified z-score. This is resistant to the "masking effect" that causes textbook mean+standard-deviation detectors to miss anomalies once a few large spikes have inflated the standard deviation. It does NOT call any cloud provider API — the caller must already have exported the billing data to a CSV/text string and pass its contents directly.

Args: params (AnalyzeBillingWasteInput): Validated input containing: - csv_content (str): Raw billing export text (see field description for supported provider formats). - analysis_type (AnalysisType): 'waste-audit' (default), 'cost-forecast', or 'efficiency'. - response_format (ResponseFormat): 'markdown' (default) or 'json'.

Returns: str: Markdown report, or a JSON object with this schema: { "anomalies": [ {"date": str, "expectedCost": float, "actualCost": float, "deviation": float, "zScore": float, "severity": "LOW"|"MEDIUM"|"HIGH"|"CRITICAL", "description": str} ], "metrics": {"averageDailyCost": float, "stdDeviation": float, "sharpeRatio": float, "wastePercentage": float}, "recommendations": [ {"title": str, "description": str, "potentialSavings": float, "effort": "LOW"|"MEDIUM"|"HIGH"} ], "summary": str }

Error response: "Error: <message>" when the CSV cannot be parsed
(e.g. no recognizable date/cost columns).

Examples: - Use when: "Why did our AWS bill spike last month?" -> paste the CUR export contents as csv_content. - Use when: "What will we spend next month at this rate?" -> analysis_type="cost-forecast". - Don't use when: you need architecture/reliability risk instead of cost — use cloudsealed_score_architecture_risk.

Inputs: params (required). Returns: result

cloudsealed_score_architecture_riskA
Read-onlyIdempotent

Score architecture risk from a declared system inventory.

Scores single-point-of-failure, excessive-coupling, and scalability-gap risk (0-100 each) for every declared system using explicit, weighted rules — not a trained model. Every score ships with a rule-by-rule breakdown so the reasoning is auditable, not a black box. Calls the Predictive-ML-Core production HTTP service (or a self-hosted instance if PREDICTIVE_ML_CORE_URL is set).

Args: params (ScoreArchitectureRiskInput): Validated input containing: - company_name (str): Name of the company/project. - systems (list[SystemInput]): Declared inventory — each with name, type (APPLICATION|DATABASE|API|THIRD_PARTY_SERVICE), criticality (LOW|MEDIUM|HIGH|CRITICAL), public_facing, and optional data_sensitivity/auth_method. - historical_metrics (Optional[HistoricalMetricsInput]): Observed latency/throughput, improves the scalability-gap score. - response_format (ResponseFormat): 'markdown' (default) or 'json'.

Returns: str: Markdown report, or a JSON object with this schema: { "predictions": [ {"systemName": str, "riskScores": {"singlePointOfFailure": int, "excessiveCoupling": int, "scalabilityGap": int}, "scoreBreakdown": {...rule-by-rule points and rationale...}, "findings": [{"title": str, "severity": str, "description": str, "remediation": str}], "recommendations": [{"title": str, "description": str, "effort": str}]} ], "architectureSummary": str, "overallArchitectureScore": int }

Error response: "Error: <message>" if the service is unreachable or
rejects the request (e.g. empty systems list).

Examples: - Use when: "Is our checkout service a single point of failure?" -> declare it with criticality=CRITICAL, type=API. - Use when: "Which of these services should we harden first?" -> declare the whole inventory and compare riskScores. - Don't use when: you need cost/billing analysis — use cloudsealed_analyze_billing_waste.

Error Handling: - Returns "Error: Request timed out..." if the service doesn't respond within 30s. - Returns "Error: ..." with the upstream message on 4xx/5xx responses.

Inputs: params (required). Returns: result

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.7/5.0

    Scored across 2 tools

    Disambiguation5/5

    The two tools address completely different domains—architecture risk scoring and billing waste analysis. Each description explicitly references the other as the tool not to use for the wrong scenario, leaving no ambiguity about their distinct purposes.

    Naming Consistency5/5

    Both tools follow the consistent pattern `cloudsealed_<verb>_<noun>` using snake_case, with verbs 'score' and 'analyze' clearly indicating actions. Naming is uniform and predictable.

    Tool Count3/5

    With only two tools, the server feels thin. While each tool is substantial and covers a distinct need, the overall scope is narrow for a server named 'cloudsealed-mcp', making it borderline.

    Completeness3/5

    The two tools cover their specific analysis tasks well, but the server lacks supporting operations such as fetching cloud data or handling remediation. The billing tool requires the user to supply CSV data, and there's no way to act on the findings, leaving notable gaps.

    Related MCP Connectors

    Related MCP Servers