Cloudcostsmcp

OpenCloudCosts MCP

Anchor AI FinOps to real, live cloud pricing.

Your LLM's cloud pricing knowledge was frozen at training cutoff. Cloud pricing wasn't.

opencloudcosts is an MCP server that gives your AI assistant live, structured access to AWS, GCP, and Azure pricing — 16 tools it can call directly, with results it can reason over rather than guess at. Ask it to compare a full multi-resource workload across all three clouds simultaneously. Ask it for your actual post-discount effective rates from Reserved Instances, Savings Plans, or Enterprise Discount Programs. Ask it to fan out across every region and return the cheapest option. None of that is possible from training data alone, and none of it is possible from a single-cloud pricing calculator.

The Problem

Ask any AI assistant what an m5.2xlarge costs in ap-southeast-2, whether AWS or GCP is cheaper for a three-tier web app, or what your effective hourly rate is after Savings Plans. You will get a confident answer. It will probably be wrong.

Cloud pricing is a poor fit for static model weights: instance families are added and retired, regional pricing diverges (the same instance type can differ 20–40% across regions), spot markets fluctuate, and commitment discounts are by definition unique to each account. A model answering from training data cannot know your negotiated EDP rate. It can barely reliably recall last year's list price. The problem is not that LLMs are bad — it's that real-time, account-specific pricing data is structurally outside what a model can know.

opencloudcosts fixes this by giving your AI assistant 16 MCP tools backed by live provider APIs. Instead of hallucinating numbers, the model calls a tool and gets a precise answer.

A Concrete Scenario

Your team asks: *"We're evaluating whether to migrate this workload from AWS to GCP. What does the full stack cost on-demand vs. three-year committed, across all three clouds, so I can make the case to leadership?"*

Without this tool, the model interpolates — possibly from pricing that is a year out of date — across three providers, multiple resource types, and at least four pricing tiers. The numbers will be plausible and wrong.

With opencloudcosts, the model calls compare_bom with the workload spec (compute + storage + database). The tool fans out 8 concurrent provider calls across AWS, GCP, and Azure simultaneously, prices each resource category at public and committed rates, and returns a per-provider, per-term breakdown with savings analysis — in a single tool call.

No spreadsheet. No switching between three provider calculators. No manual SKU matching. No training-data approximation.

Why Not the Obvious Alternatives

Asking a model without tools — Training data has a cutoff. Cloud pricing changes constantly, varies 20–40% across regions for the same instance type, and includes account-level commitment discounts that are invisible at inference time. The model will approximate, confabulate, or recall stale numbers. There is no structured output, and the model cannot access your contracted rates under any circumstances.

Cloud pricing calculators (AWS / GCP / Azure) — Each covers exactly one cloud. They are UI-only with no API surface callable from an AI assistant. Cross-cloud comparison requires manually reproducing the same architecture three times across three separate calculators and reconciling exports by hand. They have no concept of unit economics and produce no programmatic output for agentic workflows.

Infracost — Excellent at estimating cost diffs against Terraform plans. It requires IaC files as input — it cannot answer "what does an n2-standard-8 cost in europe-west4?" without a Terraform plan in hand. It is not an MCP server and is not callable from a conversational AI context.

Calling provider APIs directly — AWS bulk pricing files are multi-GB and require targeted API access patterns to avoid downloading them in full. GCP's Cloud Billing Pricing API v1beta requires a multi-source ADC credential chain and provider-specific IAM. Azure's Retail Prices API needs pagination logic and SKU matching. Every provider uses different region naming conventions, SKU formats, and data schemas. opencloudcosts normalizes all of this behind a uniform MCP tool interface and handles credential chains, caching, and retry logic so the model does not have to.

Capabilities at a Glance

Coverage: Three Clouds, 16 Tools

Fourteen tools are fully functional. Two entries are compatibility stubs: search_pricing (a deprecated redirect from v0.8.x, kept for backward compatibility) and get_spot_history (registered but not implemented in the Go server — returns a structured "not available" response).

Credential requirements by provider

Azure Reserved VM pricing (1-year and 3-year terms) is available via the public Retail Prices API — no credentials needed. compare_bom returns committed-term Azure pricing with no setup beyond the binary.

The 16 tools by category

† Compatibility stub only — not functional for live data. See opencloudcosts-go/README.md for full parameter reference.

Performance and Reliability

Concurrency — The analysis tools are not sequential HTTP wrappers:

• find_cheapest_region and find_available_regions: errgroup + semaphore, up to 32 goroutines — queries all available regions in parallel, returns results sorted cheapest-first
• compare_bom: 8 concurrent provider calls across AWS, GCP, and Azure simultaneously
• compare_prices: semaphore of 10 concurrent region calls
• get_prices_batch: parallelized across instance types within a region

Rate limiting and timeouts — Token-bucket rate limiter at 200 req/s on the HTTP transport (OCC_RATE_LIMIT). Per-request deadline: 60s (OCC_REQUEST_TIMEOUT). Per-provider API call: 30s (OCC_PROVIDER_TIMEOUT). Graceful SIGTERM drain: 15s (OCC_SHUTDOWN_TIMEOUT).

Cache — Prices are stored in a concurrent in-memory cache (read-optimised with sync.RWMutex) with atomic JSON persistence at ~/.cache/opencloudcosts/cache.json. TTLs: public prices 24h (OCC_CACHE_TTL_HOURS), region/instance metadata 7 days (OCC_METADATA_TTL_DAYS), effective/contracted rates 1h (OCC_EFFECTIVE_PRICE_TTL_HOURS). Cache survives binary updates. 401/403 responses from billing APIs are never cached, so credential rotation takes effect immediately.

AWS pricing — EC2/EBS/RDS public pricing uses a targeted SKU API path rather than downloading the full multi-GB bulk pricing file, keeping startup fast and avoiding large network payloads.

Error isolation — Raw exception text never reaches LLM context. All tool boundaries emit structured error envelopes; full tracebacks are logged server-side only. GCP contract pricing falls back to public list prices on auth failure rather than surfacing an error into the conversation.

Validated: 234/234

opencloudcosts v1.0.0 achieves 234/234 (100%) on the LLM grounding harness, with zero XML hallucinations across the full suite.

The harness covers 234 realistic pricing questions across all three clouds: instance spot checks, cross-region comparisons, BOM estimates, multi-cloud comparisons, unit economics, AI model pricing, database pricing, storage pricing, discount summaries, egress pricing, availability queries, and network pricing.

Primary validation model: `qwen3.6-35b-128k` running locally via llama-swap — a self-hosted 35B reasoning model with no external API dependency. The harness has also been exercised against qwen3.6-35b-a3b, qwen3.5-122b-a10b@q6_k, and gemma-4-26b-a4b. Because MCP is a protocol rather than a model feature, accuracy comes from the tool returning correct live data — any MCP-capable AI assistant calls the same tool surface and gets the same structured response.

Harness progression: 109/123 (Python v0.8.x) → 169/169 (v0.9.0) → 199/199 (v0.9.2) → 234/234 (v1.0.0).

643 Go unit tests across all providers and tools verify behavioral correctness and cross-provider API parity, independent of LLM evaluation.

Use Cases

1. Price the same workload across all three clouds at once

"Price the following on AWS, GCP, and Azure simultaneously: 4 instances (8 vCPU, 32 GB RAM), 1 managed PostgreSQL database (4 vCPU, 16 GB RAM), and 500 GB block storage. Return on-demand, 1-year committed, and 3-year committed totals for each cloud with monthly and annual figures, and flag which provider is cheapest at each commitment term."

compare_bom fans out 8 concurrent provider calls and returns a per-provider, per-term breakdown with savings analysis versus on-demand. No pricing calculator does this across cloud boundaries. AWS and Azure public pricing requires no credentials; GCP requires a free API key.

2. Find the cheapest AWS region for a long-running compute workload

"I need to run a c6a.4xlarge continuously. Fan out across all available AWS regions and return the 5 cheapest, sorted by on-demand hourly rate. Show us-east-1 as a baseline."

find_cheapest_region uses a 32-goroutine fan-out across every region where the instance type is available and returns results sorted cheapest-first. Regional price deltas for the same instance type routinely exceed 20%. No credentials needed.

3. Determine your effective AWS rate after commitments

"I h