cloudprice-mcp

cloudprice-mcp is a FinOps MCP server developed by Albaker-Group, with the goal of handing cloud cost decisions that were originally scattered across multiple browser tabs and half-finished spreadsheets directly to AI clients for calculation and explanation. Any MCP-compatible AI (Claude, GitHub Copilot Agent Mode, Cursor, Windsurf, Cline, Continue.dev, Zed), after loading this server, gains pricing data and analysis capabilities across AWS, Azure, GCP, and OCI.

[Purpose] cloudprice-mcp is aimed at cloud cost optimization (FinOps) scenarios, enabling AI to answer and calculate: break-even points for reserved instances/savings plans, total cost of ownership (TCO) for multi-cloud workloads, migration exit cost analysis, snapshot cost modeling, egress arbitrage, cross-cloud GPU price comparison, cross-provider LLM token price comparison, and carbon emission estimates for workloads. All amounts in the results can be traced back to public price catalogs, and are accompanied by AI-explained trade-offs, payback periods, and possible "do not migrate" recommendations.

[Main Capabilities] The toolset contains a total of 25 tools, evolving by version:

• Single-SKU queries (v0.1): query the pricing of a single cloud resource.
• Batch and workload comparison (v0.2): use the WorkloadInventory structure to compare complete workloads.
• Object storage and managed Postgres (v0.3): compare object storage tiers and managed PostgreSQL.
• FinOps decision suite (v0.6): four named tools (migration, commitment, TCO, egress arbitrage), completing the transformation from cross-cloud pricing to FinOps decisions in a single call, with inputs being a structured workload inventory (compute / storage / object_storage / databases / egress) and options (commitment, multi_az, one_time.data_to_migrate_gb). The output includes honest_gaps, explicitly listing what is not modeled.
• Egress and multi-AZ (v0.5): compare_egress models internet and cross-region egress, explicitly presenting the OCI 10 TB free allowance; multi-AZ multiplies compute costs by x2 with multi_az: true.
• Snapshots (v0.2.1): snapshot_count in storage rows and os_disk_snapshot_count in compute rows participate in pricing, giving an upper-bound estimate as snapshot_per_gb_month x full capacity x quantity x snapshot_count; real incremental snapshots are typically 20-50% of that value.
• Reserved instance/savings plan estimates (v0.2.1): compare_workload accepts a commitment parameter and applies conservative average discount tiers; storage and snapshots are not discounted.
• Multi-cloud spot bidding (available before v0.11+, including eviction trade-offs).
• Multi-cloud price history (v0.7.1+): retains weekly snapshots, can answer "how much did m5.xlarge cost in May," and supports the CLI plus the two MCP tools get_price_history and list_tracked_skus. This is the only publicly available weekly-refreshed dataset of its kind.
• Cross-provider LLM token pricing (v0.12.0+): covers 19 models from 8 providers, including Claude (4 Opus/4 Sonnet/3.5 Haiku/3 Haiku), GPT (5, 5 mini, 4o, 4o-mini, o1), Gemini (2.0 Flash, 1.5 Pro/Flash), Llama (3.1 8B/70B/405B, 3.3 70B), Mistral Large 2, DeepSeek V3/R1. Provides compare_token_pricing (manually verified catalog) and lookup_extended_model_pricing (extended catalog from LiteLLM, about 2000 model/provider combinations, covering Together AI, Fireworks, Groq, Replicate, Perplexity, etc.).
• Multi-cloud GPU pricing (v0.11.0+): covers NVIDIA T4, A10, A10G, L4, L40S, V100, A100, H100, returning the absolute hourly price winner, per-GPU efficiency winner, over-provisioning flags, and GPU memory differentiation across the four major clouds. OCI 8x H100 is about 19 cheaper than the same hardware on AWS/Azure%.
• Cost Drift Sentinel (v0.9.0+): stateless design, baseline JSON can be stored anywhere, computed as a pure function of (directory, workload, baseline); workload hash protection, SKU-level attribution, configurable threshold (default 5%), and a GitHub Action template at examples/cloudprice-watch.yml.
• Carbon-aware FinOps (v0.10.0+): the same query returns USD and kg CO2e. Modeled items include per-cloud PUE, regional grid carbon intensity, renewable matching, ARM vs x86 power consumption difference (ARM is about 30% lower per vCPU), and both location-based and market-based carbon values are provided. Unmodeled items (always disclosed): embodied carbon, GPU/network/storage power consumption, time-varying grid fluctuations, real-time 24/7 CFE matching.
• Price anomaly detection: z-score and percentage change statistical detection on weekly snapshot archives.
• CFO-level FinOps decision report generation: Markdown report containing executive summary, cost tables, carbon, audit trail, and honest gaps.
• OCI Always Free tier (4 OCPU compute, 20 GB object storage, 10 TB egress) is shown as a $0 line item where applicable.

[Usage] Install: pip install cloudprice-mcp && cloudprice-mcp setup. This command automatically detects Claude Desktop, GitHub Copilot Agent Mode, Cursor, Windsurf, Cline, Continue.dev, and Zed, and asks Y/N before writing each configuration. After configuration, the corresponding client must be fully restarted, and 10 tools will appear in each client. If problems occur, a diagnostic command can be run to check the Python version, installation path, configuration locations, tool registration, and command path validity. Requires Python 3.10+. For manual installation steps, see INSTALL.md.

[Data Refresh] The built-in catalog is a curated dataset of common SKUs across four clouds, automatically refreshed every Sunday by a GitHub Action: AWS uses the Pricing API (boto3, OIDC authentication), Azure uses the Retail Prices API (public, no authentication), OCI uses the public pricing API (public, no authentication), GCP uses the Cloud Billing Catalog API (requires the GCP_API_KEY environment variable, added in v0.8.0). Bedrock LLM Token pricing is refreshed via the AmazonBedrock service code of the AWS Pricing API (v0.15.0). The extended LLM catalog comes from LiteLLM's model_prices_and_context_window.json (MIT licensed, community maintained, no authentication, v0.16.0). Each refresh writes a dated snapshot JSON, and the historical archive is MIT licensed and grows with releases. Every tool result includes an as_of field.

[Use Cases]

• Compare the 3-year total cost and break-even month for 6× t3.2xlarge on AWS across On-Demand, 1-year Savings Plans, and 3-year Reserved Instances with partial upfront.
• Evaluate the exit cost and payback period of moving 5 TB of cold-tier object storage from AWS S3 to a cheaper provider (often finding that AWS Glacier Deep Archive is already tied for cheapest).
• Find the cheapest cloud under 50 TB/month of internet egress, showing 3-year migration savings (OCI about $340/month vs AWS/Azure/GCP about $4,000/month).
• Compare the full-stack monthly cost across four clouds for a three-tier SaaS workload (8 web 4/16, 12 app 8/32, 4 DB 16/64, 5 TB shared SSD, 50 TB HDD bulk, 10 TB/month egress), including multi-AZ and 1-year commitment.
• Query price history trends and price anomalies.
• Compare Claude pricing on the Anthropic API vs Bedrock, and find that the Anthropic API cache_read rate is 10x cheaper.
• Find the cheapest model for 200K context, output-heavy chat, and 10M/month output volume.
• Compare the lowest-carbon cloud for 4 vCPU/16 GB 6 instances, or compare the carbon savings of ARM vs x86.
• Place workload.json and the GitHub Action template in an IaC repository to automatically open an Issue when cost drifts.

[Limitations] Some modeled items are still on the roadmap: Reserved Instances/Savings Plans only apply a uniform tier discount, with no per-regionRegion/series details; currently only us-east pricing, us-west/eu-west planned for v0.5.1; IOPS-based storage matching (currently capacity only); backup storage costs; object storage request costs (PUT/GET); archive tier retrieval costs (Glacier-style retrieval can reach 10x storage cost); VPC peering/interconnection costs. Snapshots are upper-bound estimates. Carbon footprint covers operational emissions only, excluding embodied carbon, GPU/network/storage power consumption, time-varying grid fluctuations, and real-time 24/7 CFE matching. These limitations are explicitly disclosed via honest_gaps[].

[Category] This tool is an MCP server that provides cloud pricing data and analysis primitives for AI clients, and belongs to the AI development platform category.