One agent stuck in a loop can burn $1,400 on a quiet Sunday. Inferbooks traces every model call, agent run and seat licence to a person, team and task, down to the minute, and warns you while it is still happening.
One Sunday stands out. October 11 cost $1,489, nearly four times a normal weekday. Click the spike to zoom in.
Reads what your tools already send
What it answers
Six answers from Acme Inc, the demo company Inferbooks ships with. Each one is a click away in the product.
A research run for Initech looped for 47 minutes: nearly 2,000 model calls, each re-sending a context that kept growing. Spend climbed from $18 to $43 a minute.
The agent's monthly budget crossed 50% at 16:29 and 80% at 16:45, with the loop still running. Alerts go to Slack or Teams.
The support copilot's cost per task rose from $0.04 to $0.07 between October 8 and 12, about $171 extra, then fell back.
52 of 317 seats went unused for 30 days: 19 ChatGPT Enterprise, 12 Microsoft 365 Copilot, 11 GitHub Copilot Business and 10 more.
44 people hold both ChatGPT Enterprise and Claude, and 28 hold both GitHub Copilot and Cursor. Keeping only the tool each person uses more saves $2,794 a month.
11% of provider billing came from API keys nobody owns yet. The identity map lists them so you can assign each one once.
Acme Inc · 150 people · 14 teams · 90 days of generated sample data
The product
Overview, minutes, tasks, seats and budgets all read the same ledger, so finance and engineering finally look at one bill.
API usage and seat licences in one total, split by team every day. Insights point at what changed and where to save.
Drag across any chart to zoom: buckets refine from days to hours, five minutes and single minutes. Split by agent and the culprit stands alone.
Each task gets a timeline with one lane per model, tool and agent, and a cost curve. A steep climb is a loop or a context that keeps growing.
Seats from ChatGPT, Claude, Copilot, Cursor and Microsoft 365 sit next to API spend, with daily activity for every person.
Monthly limits for a team, an app or a kind of work, with a forecast to the end of the month and alerts at 50, 80 and 100%.
Who it's for
How it works
Inferbooks reads the telemetry your tools already emit and the usage your vendors already report. Requests go straight to the providers, exactly as they do today.
Claude Code, Codex, Gemini CLI and your own apps send OpenTelemetry. Roll the settings out with your device management.
OpenAI, Anthropic, GitHub Copilot, Cursor, ChatGPT Enterprise and Microsoft 365 report bills, seats and who used them.
Keys, emails and agent names map to people and teams, so each model call, tool call and seat lands on a person and a task.
Budgets with alerts, insights on what changed, and a list of seats to reclaim before the next renewal.
Security
Run it on your own servers when usage data cannot leave your company, or let us host it when you would rather not run anything.
Run it in your own infrastructure: your cloud, your data center or your Kubernetes cluster, with your data in a ClickHouse database you control. Or let us host it, with nothing to install or operate: a good fit for smaller teams.
Who, which model or tool, tokens, cost and timing. Prompts, responses and code are never collected.
Sign in through your identity provider with OpenID Connect, limited to the email domains you allow.
Admins, analysts and members each see what their role allows, and every change lands in the audit log.
Inferbooks can be down for an hour and nobody's AI tools notice. It catches up from the vendors' reports.
OpenTelemetry in, CSV out. Your data is plain rows in your own database, ready for any BI tool.
FAQ
Anything else? Write to info@inferbooks.com.
No. Inferbooks never sits in the request path, so it cannot add latency or take your tools down. It reads OpenTelemetry from your tools and usage from your vendors' admin APIs.
No. Inferbooks stores metadata only: the person, team, task, model or tool, tokens, cost and timing. Prompt and response content is never collected.
Usage from Claude Code, Codex, Gemini CLI and any app instrumented with OpenTelemetry. API bills from OpenAI and Anthropic. Seats from GitHub Copilot, Cursor, ChatGPT Enterprise, Claude Enterprise and Microsoft 365 Copilot, plus CSV for anything else.
Wherever suits you. Self-hosted: in your own infrastructure (your cloud, your data center or your Kubernetes cluster), with your data in a ClickHouse database you control. Hosted: we run it for you, with nothing to install or operate, a good fit for smaller teams without their own infrastructure.
From the emails, API keys and agent names in the telemetry, matched to your directory. Anything unmatched shows up in one identity map where you assign it once.
We walk through Inferbooks on the sample company, look at the AI tools your teams use, and, if it fits, plan a pilot, self-hosted or hosted by us.
Book a demo. We'll walk through Inferbooks on the demo company, then plan a pilot, on your own servers or hosted by us.