finops

AI FinOps for the MCP era: why per-call cost attribution is becoming table stakes

Cloud FinOps took 5 years to mature. AI FinOps is repeating the same arc — but the unit of cost is now the MCP tool call, not the EC2 instance.

Andrei Sirbu

July 1, 2026 · 3 min read

If you remember the early days of cloud, you'll recognize the pattern. Around 2014-2018, companies got hit with surprise AWS bills because they had no per-team, per-environment, per-feature attribution. The solution arrived in waves:

2015: Cost & Usage Reports + custom dashboards (Vantage, CloudHealth)
2017: Tag-based showback / chargeback (every infrastructure team's standard)
2019: Real-time anomaly detection (alerts on spike-vs-baseline)
2021: Unit economics (cost per request, cost per active user)

Each wave needed new tooling. None of them existed in 2014. By 2022 you couldn't run a serious cloud team without them.

AI agents are now where cloud was in 2015.

The unit shift

Cloud FinOps measures $ per EC2-hour or $ per S3-GB-month. The unit is a resource consumed over time.

AI FinOps in the LLM-only era (2023-2024) measured $ per chat completion. That worked when one prompt = one cost event. Tools like Helicone, Langfuse and PostHog built their dashboards around that unit.

But agentic workflows broke this model. A single user prompt now triggers 30-80 MCP tool calls in the background. The unit needs to shift to:

$ per MCP tool call
× attributed to:
    - which MCP server
    - which agent / IDE
    - which project / customer
    - which prompt / session

That's the granularity that lets you do the things FinOps did for cloud:

Showback: "Customer ACME's agent activity cost $312 this month — bill it back to their account."
Anomaly detection: "Playwright cost jumped 5× yesterday because someone shipped an agent that scrapes 200 pages instead of 20."
Unit economics: "Our AI feature costs $0.04 per query — at $4/month subscription that's profitable above 10 queries."
Optimization triage: "playwright/browser_navigate is 40% of our spend. Fix that first."

None of these are possible with LLM-call-level data alone.

Why MCP makes this finally tractable

Pre-MCP, every agent integration was custom code calling provider SDKs. There was no standard observation point. You either instrumented every codebase by hand (expensive, brittle) or you put a proxy in front of OpenAI/Anthropic (limited — misses tool calls).

MCP changed this. Every MCP server speaks the same JSON-RPC dialect over stdio or HTTP. One proxy can observe every tool call, regardless of which server, which client, which provider. It's the FinOps equivalent of EC2 finally having consistent tags.

What "good" looks like in 2026-2027

Companies that scale AI agent usage will adopt the same FinOps maturity ladder cloud did, just on a 2-3x faster timeline:

Cloud FinOps year	Practice	AI FinOps equivalent
2015	Custom CUR dashboards	Per-tool cost dashboard
2017	Tag-based showback	Per-customer attribution
2019	Real-time anomaly alerts	Tool-cost spike alerts (3× baseline)
2021	Unit economics	$ per agent task / per user query

My prediction: in 18 months, the question "what does it cost when a customer asks our agent X?" will be as foundational for AI-product teams as "what does it cost to serve a page?" was for cloud-native teams in 2018. The teams without per-tool attribution will be flying blind.

How MCPSpend fits

We're building the AI FinOps stack at the MCP layer specifically — not LLM-call observability (Helicone does that better), not generic event analytics (PostHog does that better). Just the missing layer: every MCP tool call, attributed to its tool, server, project, end-customer.

If you're running agentic workflows seriously, you want this layer regardless of whether it's us providing it.