By hand
Open the terminal and look
Works for one or two sessions you started yourself. It stops working overnight, across machines, or when a session has been waiting on you for an hour.
Agent monitoring · Free for 5 agents
What's your agent doing right now?
An agent can fail without crashing: it stalls, waits for your “yes”, or never starts. Agent Tracking puts your Claude Code sessions, cron jobs and own agents on one list and tells you which one needs you.
For developers and teams who let Claude Code run overnight, across servers or from cron.
Pick one way. Both do the same thing.
npx --yes https://agenttracking.co/downloads/agenttracking-0.3.0.tgz connect claude
Shows a code you confirm in the browser. Tracking starts with your next Claude Code session.
Set up agenttracking.co agent monitoring on this machine. 1. Run: npx --yes https://agenttracking.co/downloads/agenttracking-0.3.0.tgz connect claude It prints a code and a link. Show me both and wait until I have confirmed the code in the browser. 2. Run `crontab -l`. For every job, suggest a new line that wraps the command like this: npx --yes https://agenttracking.co/downloads/agenttracking-0.3.0.tgz wrap --name <short-name> --every <interval matching the schedule> -- <original command> Show me the complete new crontab and only install it after I say yes. 3. Run `npx --yes https://agenttracking.co/downloads/agenttracking-0.3.0.tgz status` and tell me whether this machine is connected.
Claude runs the commands, waits for your confirmation and asks before it changes your crontab.
- Open source, AGPL-3.0
- Free for 5 agents, no card
- No prompts leave your machine
Built by @wuebbe
Today
How you probably check today
All three are reasonable, each for the right job.
Exit codes
Cron mail and failing exit codes
Good for scripts that end. An agent that is stalled or waiting has not ended yet, so there is nothing to report.
LLM tracing
AgentOps, LangSmith, Langfuse
The right tool for prompts, spans and token cost inside agents you build. They need instrumentation in your code and usually record prompts. If you need both, use both.
The bar
What watching an agent should take
- No changes to your code or your agent
- Tells stuck, waiting and missing apart, not just pass or fail
- No prompts, code or tool output leave your machine
- One message in the morning, alerts only when it matters
That is what Agent Tracking is built for: a monitor for agent runs. Claude Code sessions through hooks, cron jobs through a wrapper, your own agents through OpenTelemetry.
What it flags
What it notices for you
Stuck
A session that stopped moving
When a running session stops sending events, it is marked as suspected stuck.
Missing
A cron job that never started
With --every the monitor knows when the next run is due. If it does not come, you hear about it.
Failed
Exit codes and tool errors
Failed runs and tool calls; for cron, optionally with the last stderr lines, tokens masked.
Waiting
Claude is waiting on you
Sessions waiting for input are marked instead of sitting unnoticed for hours.
Once a day you get one briefing, at the hour you pick (times are in Central European Time for now; US time zones are next). Immediate alerts go out by email, and optionally to Telegram, Slack, Discord or a webhook.
Proof
Real runs, not a demo data set
This is the developer's own server, not a customer. We use it every day to watch our Claude Code projects and cron jobs. No customer logos yet; connect one machine and judge for yourself.
Setup
Set up in three steps
Connect Claude Code
Signs this machine in and adds hooks to ~/.claude/settings.json. A backup is written next to it and your own hooks stay. Use --project for the current folder only.
npx --yes https://agenttracking.co/downloads/agenttracking-0.3.0.tgz connect claude
Wrap your cron jobs
Your job's output and exit code stay exactly the same. --every reports missing runs, --tail sends masked error lines.
# crontab, before 0 7 * * * /opt/backup.sh # after 0 7 * * * npx --yes https://agenttracking.co/downloads/agenttracking-0.3.0.tgz wrap --name backup --every 24h --tail 20 -- /opt/backup.sh
Switch it off whenever you like
Removes our hooks and nothing else.
npx --yes https://agenttracking.co/downloads/agenttracking-0.3.0.tgz disconnect claude
Configuration snapshots, OpenTelemetry and the optional Linux observation are in the docs. →
Integrations
Connected today, planned next
Today
Claude Code
Through hooks, with one command.
Today
Cron jobs and scripts
Any command can be wrapped, including an agent you start from a script.
Today
OpenTelemetry
Anything that sends OTLP/HTTP JSON traces shows up with its trace structure.
Next
Codex, Hermes, OpenClaw
Dedicated integrations are planned. Follow them in the changelog.
What is sent, and what never is
Sent: event type, tool name and whether it succeeded, the project folder name, a host label and a session ID hashed on your machine. For cron, also exit code and duration.
Never sent: prompts, answers, tool input or output, file contents, full paths or transcripts. Events are deleted after 30 days. Every field and retention period.
Behind it
Why this exists
Agent Tracking started by measuring AI agents on websites. On October 31, 2026, Boardy replied: “now point it at their own agents”. Two days later the monitoring was running on our own server.
Pricing
Free to start, $9 when you need more
Free
$0 forever
No credit card
- 5 agents
- Claude Code, cron jobs and scripts
- Daily briefing by email
- Instant alerts by email
- 30 days of history
Pro
$9 per month
or $90 a year, two months free
- 50 agents
- Everything in Free
- Telegram, Slack, Discord and webhook alerts
- Alerts on configuration changes: MCP servers, plugins, rules, CLAUDE.md
- Alerts when an agent uses a tool for the first time
- OpenTelemetry and task checks
Self-hosted
$0 on your server
Open source under AGPL-3.0
- All of the code: CLI, ingest, dashboard, alerts
- One Node process, one SQLite file
- Your data never leaves your server
Prices in US dollars, plus tax where due. Cancel any time, effective at the end of the period. Upgrade and cancel in the dashboard. Need more agents or events? Write to hi@finalmaster.net.
FAQ
Questions
Does Agent Tracking read my prompts or code?
No. The Claude Code integration sends lifecycle events, tool names and success or failure only. Prompts, answers, tool input and output, file contents and full paths stay on your machine.
What happens to my existing Claude Code hooks?
They stay. connect claude only adds its own entries and writes a backup next to ~/.claude/settings.json. disconnect claude removes exactly those entries again.
Does it work with Codex, Hermes or OpenClaw?
Not with a dedicated integration yet. Until then, wrap any agent run you start from a script or cron, and agents with an OpenTelemetry exporter can send traces directly. New integrations are announced in the changelog.
Who is it not for?
If you need to debug prompts, compare model output or track token cost, use an LLM tracing platform. Agent Tracking deliberately never sees prompts.
Where is my data, and who is behind this?
The hosted version runs on servers in Germany, operated by FINAL MASTER LLC. Only run metadata is stored. Prefer your own infrastructure? Self-host it anywhere Node runs.
What does it cost?
Free costs nothing (5 agents). Pro is $9 a month or $90 a year (50 agents, every channel). Self-hosted it is free for good.
How do I update the CLI?
Run the connect command with the new version again. The hooks then point at the new version.
And on your website?
Agent Tracking also measures which AI assistants send visitors, which tool calls fail on your site and which inquiries your server confirms.
Get started
Connect your first agent
One command, one confirmation in the browser. From your next session on you can see what it does.
npx --yes https://agenttracking.co/downloads/agenttracking-0.3.0.tgz connect claude