Every entry ends with a command for your terminal or a prompt to paste into Claude Code. Copy, paste, done.
Next up
Planned, not built yet. No date until it works.
Planned
Guardrails: block risky commands before they run
Rules that stop a Claude Code tool call before it runs, such as rm -rf outside the project, a force push to main, reading .env or ~/.ssh, or curl piped into a shell. The rules are checked on your machine; only the verdict is sent ("blocked: force-push"), never the command itself. Planned for Pro.
Today: every tool call shows up by name and result, and changes to your Claude Code hooks and settings show up as configuration changes. Which rule do you want first? Tell @wuebbe on X.
Planned
Your time zone
Briefing hour and dashboard times in your own time zone, including US time zones.
Today: times are Central European Time. Pick the briefing hour in the settings accordingly; 13:00 CET is 7 AM Eastern.
Planned
OpenAI Codex
Sessions and tool calls from the Codex CLI, the same way as Claude Code.
Today: wrap a Codex run you start from a script or cron. You get start, end, duration, exit code and missing runs.
Two new alerts for Pro, through the same channels as failures. Configuration: an MCP server or plugin added or removed, permission rules loosened or tightened, settings, .mcp.json or CLAUDE.md changed. New behavior: an agent uses a tool it has not used before, say a backup job that suddenly calls WebFetch. Both come from data that was already sent: counts, checksums and tool names, never file contents or commands. New agents learn for three days and 20 tool calls before they can trigger a new-tool alert. Each agent's page lists its signals.
How to use it
>_Paste into your terminal
npx --yes https://agenttracking.co/downloads/agenttracking-0.3.0.tgz connect claude
Configuration snapshots need CLI 0.3.0: run the command above once more on each machine. Then switch to Pro under Your agents → Settings.
Agent monitoring now has two plans. Free: 5 agents with the daily briefing and email alerts. Pro: 50 agents, Telegram, Slack, Discord and webhook alerts, OpenTelemetry and task checks, for $9 a month or $90 a year. Agents you already have keep reporting on either plan; only new agents above the limit are refused.
How to use it
Your plan and the upgrade are in the dashboard under Your agents → Settings.
The home page now leads with agent monitoring: why agents fail without crashing, how that compares to checking by hand, exit codes and LLM tracing, and one command to connect. Website analytics moved to their own page. Every changelog entry ends with a command or a Claude Code prompt.
Website analytics: /websites
Markdown versions for agents: send Accept: text/markdown to / or /websites
llms.txt and llms-full.txt describe both sides
How to use it
>_Paste into your terminal
npx --yes https://agenttracking.co/downloads/agenttracking-0.3.0.tgz connect claude
✳Or paste into Claude Code
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.
Every Claude Code session now has an event-by-event timeline. Named checks report whether a task actually produced the expected result, by version. Agents that speak OpenTelemetry can send traces directly.
Session timeline: start, tool calls, notifications and end, with tool names and results only
Task checks: wrap a check script with --check and --version; its exit code decides passed or failed
Configuration snapshots on session start and config changes: fingerprints and counts, contents stay local
OTLP/HTTP JSON trace receiver with parent relationships, status and service versions
Optional bpftrace observation of one Linux process: operation types and hashed IDs, no filenames or payloads
How to use it
>_Paste into your terminal
# Update: run connect again, the hooks then point at 0.3.0
npx --yes https://agenttracking.co/downloads/agenttracking-0.3.0.tgz connect claude
# A named check for a task
npx --yes https://agenttracking.co/downloads/agenttracking-0.3.0.tgz wrap --name evaluation --check booking --version v1 -- ./check-booking.sh
# Preview what the configuration snapshot would send
npx --yes https://agenttracking.co/downloads/agenttracking-0.3.0.tgz snapshot
✳Or paste into Claude Code
Update my agenttracking.co connection to CLI 0.3.0 and add a task check.
1. Run: npx --yes https://agenttracking.co/downloads/agenttracking-0.3.0.tgz connect claude
2. Run: npx --yes https://agenttracking.co/downloads/agenttracking-0.3.0.tgz snapshot and explain in two sentences what it would send.
3. Ask me which task I want to check. Write a small script that exits 0 only if the expected result is really there and nonzero otherwise, then run it with:
npx --yes https://agenttracking.co/downloads/agenttracking-0.3.0.tgz wrap --name <task> --check <check-name> --version v1 -- <script>
OpenTelemetry setup, with your token, is in the dashboard under Your agents → OpenTelemetry.
The other side of agent traffic: what your own agents did while you were away. One command connects a machine; stuck, missing and failed runs are detected by rules, not guesses. Free during the beta, up to 50 agents.
Device login with a per-machine token you can revoke
Claude Code hooks that keep your own hooks and write a backup
wrap for cron jobs: same output and exit code, --every for missing runs, --tail for masked stderr
Morning briefing at 07:00 Europe/Berlin by default; alerts by email, Telegram, Slack, Discord or webhook
How to use it
>_Paste into your terminal
npx --yes https://agenttracking.co/downloads/agenttracking-0.3.0.tgz connect claude
✳Or paste into Claude Code
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.
Connect your MCP server and see calls, errors and duration per plugin version, next to server-confirmed outcomes such as created PDFs, inquiries and bookings. Test runs and production stay separate.
How to use it
Setup with the SDK is in the docs under MCP analytics.
Observed problems now link to their evidence and a next step. Your server can confirm real inquiries and bookings, and a documented fix can be checked again with a prepared browser run.
How to use it
✳Or paste into Claude Code
Add the Agent Tracking snippet to every page of this website.
1. Find the shared layout or template that renders <head> on every page.
2. Add this tag just before </head>, with my domain in place of example.com:
<script defer data-domain="example.com" src="https://agenttracking.co/agent.js"></script>
3. Show me the diff before you save it. Do not add cookies, consent code or anything else.
Your exact snippet, with your domain, is in the dashboard after you add the site.
User agents we cannot place are no longer dropped. A nightly job suggests which ones look like AI agents; only a person adding them to the source list changes a number.