See what your AI coding agents spend, and what the spend pays for.
Kestrel reads the session logs of Claude Code, Codex and Cowork on your computer. It shows the tokens and the cost of each call, session, ticket and account, how much of your plan's limits is used, and the waste it can fix. The Kestrel board shows the same for a whole team, with the cost of each merged pull request.
Early access. Free during early access.
kestrel hud shows.Who it is for
- You use AI coding agents every day
- See what each session, ticket and account uses, and how much of your plan's limits is left before it resets.
- You lead an engineering team
- See what the team's agents cost, the cost of each merged pull request, and how many agents each person runs at once.
- You answer for security or compliance
- See the risks that agents create: settings that let an agent do anything, reads of secret files, secret keys in code. Kestrel flags risks. It does not make a company compliant or certify it.
On your computer
Kestrel runs on your computer. It reads the logs that the agents already keep there, and keeps only numbers and names. On its own it sends nothing anywhere.
- The status line
- At the bottom of each Claude Code session: the plan's limits, the size of the context, how long the prompt cache stays warm, today's tokens and open security risks.
- Reports
- Spend per call, session, ticket, project, model and account, at list prices. For example
kestrel summary,kestrel ticketsandkestrel accounts. - Waste, and the fix for it
- Lost prompt caches, a context that has grown too large, and a model or effort level that costs more than the work needs. Kestrel proposes a change. You apply it with one command and undo it with one command. Nothing changes on its own.
- In Claude Code and the Claude app
- A Kestrel pane beside the chat in Claude Code. A desktop extension for the Claude app, where you ask in a chat what your agents spent.
- Security checks
- The agents' settings, what the agents did, and the code they changed. On request, read only checks of an AWS account, a Vercel project or a GitHub repository. Each finding names the controls it touches in the frameworks you turn on, such as HIPAA, SOC 2, HITRUST, 21 CFR Part 11, PCI DSS and the OWASP Top 10 for Agentic Applications.
On the team board
Each person links their computers to the board. Kestrel then sends what it records, and the board joins it with the team's pull requests on GitHub.
- Spend
- The team's spend by person, repository, ticket, model and agent.
- Cost per merged pull request
- The middle value and the 90th percentile, and how much of the spend the pull requests explain.
- Parallel work
- How many agents each person runs at once, and how long agents wait for a person.
- Waste
- Restarts after a pause that send the whole context again, the mix of models and effort levels, and Kestrel's proposals added up.
- Security findings
- The open findings of every linked computer, by severity, repository and framework.
The board compares the team with its own past. It lists people in alphabetical order and does not rank them.
What Kestrel found on one computer
One person who uses Claude Code all day ran Kestrel over 30 days of logs. Their agents made 150,541 model calls, worth $16,677 at list prices.
- 73% paid for carrying context that was already sent
- 27% paid for everything else
- 89% of the spend was in subagents: agents that the main agent starts for a part of the work.
- 2,950 prompt caches were lost, worth $5,142. A provider keeps the context in a cache for 5 minutes or 1 hour. After a longer pause, the next call sends the whole context again at the higher price.
Carrying context: an agent sends its whole conversation again with every call, so each call pays again for what was sent before. These numbers are from one computer and one person. List prices are what the same use costs when every token is paid for. On a subscription this is not money paid.
What is sent, and what is not
Kestrel on its own sends nothing. When you link a computer to the board, Kestrel sends what it records there.
Sent to the board
- For each model call: the time, the agent, the model, the token counts and the cost at list prices.
- The folder, the repository, the git branch and the ticket in the branch name.
- The names of the tools an agent used, and the size of what entered the context.
- Security findings: the rule, the severity and where, such as a file and line.
- The computer's name, as the computer reports it (it often holds a first name).
- The names of tasks and of what subagents work on, while Kestrel's setting for names is on (the default).
Never sent
- The text of prompts and answers.
- The text of tool results and of files.
- What a security check found, such as a key or a number.
- Your email address.
The board's data is stored in the United States. The members of your team can see it, and Widal staff only when they must, to run or fix the board. Read what is stored, where, and who can see it.
How to start
- Sign in with GitHub. The board asks GitHub for your id, login and picture. It never asks for your email.
- Install the board's GitHub App on your organization, or on your own account if you work alone. It can read pull requests, and it cannot change anything.
- Install Kestrel on each computer and link it with
kestrel board login https://kestrel.widal.com.