Agent unit economics · Margin protection

Know what your AI costs and whether it pays off.

Every agent run adds token and tool costs. Your subscription price may stay the same. See what each customer costs to serve, so you know what to optimize, limit, or charge for.

Agent workflow metadata only · No prompts or responses required · No proxy

Start free trial

14 days · No credit card required

See an example now

02 lines to install↗ no proxy∅ no prompts stored✓ fail-open telemetryOpenAI · Gemini · Anthropic · Perplexity · MCP

The silent failure mode

Usage looks healthy. Customers are happier. The economics are getting worse.

A product team adds longer context, stronger retrieval, MCP tools, and a premium model for difficult requests. Quality improves and revenue stays fixed—but every tool can trigger more planning, retrieval, validation, retries, and model calls.

Longer contextMore retrievalMCP fan-outPremium routingTool retries

Illustrative customer$50 / month plan

At launchAI cost-to-serve

$10+$40 profit

Six months laterAI cost-to-serve

$80−$30 loss

The provider bill shows $80 of spend. Ganivra shows which customer caused it, which workflow changed, and what to do next.

01Which customers are profitable?

02Which features consume margin?

03Which model or prompt change hurt economics?

04When should pricing, limits, or routing change?

MCP cost attribution

The tool call is only one line on the trace. The cost can keep growing after it returns.

MCP gives agents access to data and actions. It also creates hidden cost paths: a single tool decision can trigger retrieval, follow-up generation, validation, and retries. Ganivra joins the server, tool, and every downstream model call under one execution.

See cost by server + toolTrace downstream model callsCatch loops and retriesPrice the full customer action

One customer request$0.212 total

01
Agent plansGPT-5
$0.040
02
search_accountsMCP · crm-server
$0.002
03
Agent summarizesGPT-5
$0.110
04
Validation retryGPT-5 mini
$0.060

0.2¢ direct tool cost became 21.2¢ of execution COGS. The protocol is not the problem; invisible fan-out is.

High-cost execution detected

One MCP lookup triggered three model calls. At scale, that pattern erases $11K/month in margin.

The search_accounts tool itself cost only 0.2¢. Planning, follow-up generation, and a retry pushed this support execution for Northstar Health to 31¢—3.4× the workflow baseline.

MCP: search_accountsaccount: Northstar Health3 model calls · 1 tool call

Inspect the four linked steps →

Typical answer$0.09

This execution$0.31

At 50k answers / month$11,000monthly exposure if the extra 22¢ becomes a pattern

Unit economics snapshotIllustrative · with revenue metadata

AI revenue

$6,420Example revenue metadata

AI cost

$1,842.67Model + MCP-linked COGS

Gross margin

71.3%$4,577 after AI cost

Cost / execution

$0.048Workspace average

Below target margin

3Illustrative customers

Prompt-level unit economics

Know which prompt versions make or lose money.

Illustrative product direction · with revenue metadata

Margin regressionticket_resolution / prompt:v17

+28%cost per execution

Prompt v17 added more context and longer outputs. Cost rose from 4.8¢ to 6.1¢ while gross margin fell eight points.

Prompt versionCost / executionGross marginChange

v17Current$0.06163%−8 pts

v16Baseline$0.04871%—

v15Previous$0.05268%−3 pts

No prompt content stored. Ganivra attributes cost using customer-supplied prompt_id and prompt_version metadata.

The profitability gap

Your AI bill cannot tell you whether your AI product is profitable.

Provider spend is a total. Margin is a decision.

Provider bill saysGanivra reveals

GPT-5 spend: $1,054Prompt v17 raised cost 28% and cut margin 8 points

Total AI cost: $1,843Northstar Health costs 2.7× average

Requests ↑ 18%Retry behavior adds $11K/month exposure

1 MCP tool call3 downstream model calls drove 99% of execution cost

429M tokensThree customers fell below target margin

What you spentWhere margin is won or lost

AI agents across industries

Find where your AI workflows lose margin.

Model + MCP events · one API

01 · Customer support

Protect fixed-fee margins

Inspect repeated calls and growing token usage in support workflows. See which customers cost more to serve than their contracts earn.

Explore customer support →

02 · LegalTech & compliance

See the cost of every document workflow

Compare model and prompt-version costs, including reported cached tokens. Identify expensive extraction workflows before changing models or prompts.

Explore LegalTech →

03 · Financial services

Price the full research workflow

Link model calls and MCP tool steps to one research report or audit execution. See the tracked cost beyond the initial query.

Explore financial services →

04 · Coding agents & DevTools

Inspect expensive agent runs

Follow model steps, token usage, and reported retries through a coding workflow. Find where repeated generation and testing add cost.

Explore coding agents →

05 · Multi-tenant B2B SaaS

Know your cost per customer

Use existing tenant IDs to compare usage, revenue, and margin. See where customer pricing or usage limits need a closer look.

Explore B2B SaaS →

06 · Regulated industries

Share usage facts, not sensitive content

Send model names, tokens, latency, and pseudonymous IDs. No prompts, responses, PHI, or customer contact details required. No Ganivra proxy.

Explore regulated enterprises →

Breakdowns

Where cost-to-serve accumulates

Illustrative B2B data · metadata only

By model

Cost

gpt-5$1,054.80

gpt-5-mini$438.52

gpt-4.1$244.10

gpt-4.1-mini$105.25

By MCP tool · linked COGS

Cost

search_accounts$388.12

retrieve_policy$274.45

update_ticket$189.90

check_availability$96.22

By workflow

Cost

inbound_support_call$741.28

ticket_resolution$506.81

document_processing$367.43

sales_follow_up$227.15

By feature

Cost

voice_agent$698.02

support_copilot$467.36

workflow_automation$401.12

ticket_triage$276.17

By account

Cost

Northstar Health$315.48

Atlas Logistics$248.91

Meridian Financial$196.20

Summit Retail$164.87

Margin risk queue

Executions eroding margin

5 shown

Execution	Workflow	Account	Calls	Tokens	Cost	Latency	Timestamp
exec_demo_72f1	inbound_support_call	Northstar Health	4	22K	$0.31	7.20s	Aug 07, 2026 · 12:00 UTC
exec_demo_91ac	ticket_resolution	Atlas Logistics	5	29K	$0.284	11.92s	Aug 07, 2026 · 11:42 UTC
exec_demo_0bd4	document_processing	Meridian Financial	4	22K	$0.247	8.64s	Aug 07, 2026 · 11:18 UTC
exec_demo_c822	sales_follow_up	Summit Retail	6	19K	$0.219	14.32s	Aug 07, 2026 · 10:51 UTC
exec_demo_f10e	inbound_support_call	Northstar Health	3	9.7K	$0.183	5.84s	Aug 07, 2026 · 10:37 UTC

Frequently asked questions

What teams ask before instrumenting.

Straight answers about data, integrations, and getting started.

01What is Ganivra?

Ganivra is an AI and MCP unit economics platform for teams building agentic products. It connects model and tool cost to the customer, feature, workflow, prompt version, MCP server, and MCP tool that caused it.

02How does Ganivra calculate AI unit economics?

Ganivra attributes each paid model call, MCP tool step, retry, token, and workflow step to product metadata such as customer, feature, workflow, prompt ID, server, and tool. With revenue metadata, teams can compare the full execution cost-to-serve with revenue and gross margin.

03Why does MCP cost tracking matter?

An MCP tool call may be free or inexpensive by itself, but it can trigger planning, retrieval, validation, retries, paid APIs, and additional model calls. Ganivra joins those steps under one execution so teams can see the true cost of each server and tool instead of only the direct tool fee.

04Does Ganivra store prompts or responses?

No. Ganivra is designed to collect execution metadata rather than prompt or response bodies. Prompt-level economics use customer-supplied prompt IDs and version labels.

05Which AI providers does Ganivra support?

The Python SDK automatically instruments OpenAI Responses calls and can record MCP tools with ganivra.mcp_call. The same REST API accepts model and MCP events from any backend, with versioned catalog pricing for OpenAI, Google Gemini, Anthropic Claude, and Perplexity Sonar. Unknown costs remain explicitly unpriced until a catalog, contract, or provider-reported rate is available.

06How do I get started?

Start a 14-day full-product trial, copy your SDK key, and instrument one real workflow. Your first execution will appear in the dashboard with attributed cost.

Start your full trial

Your first workflow is enough to start finding value.

Connect one workflow to inspect its model and tool costs. Add customer IDs to see who drives the spend, then connect revenue to understand margin. No sales call or credit card required.

01 See the cost of your first recorded execution02 Find expensive customers and workflows03 Connect revenue to compare cost and margin

Initialize the Python SDK in two lines

import ganivra
ganivra.init(api_key="your_key")

Start with one workflow

See your first real execution

No credit card

01Create a workspace

02Copy your workspace API key

03Connect via SDK or REST and send usage

Start free 14-day trial

Already have an account? Sign in

Python SDK · REST APIOpenAI · Gemini · Claude · Sonar · MCPMetadata only · Fail-open