$47,000. Three days, or ninety. Nobody agrees, because nobody was measuring.
Search the story and you'll find three versions. Twenty-three unattended subagents burning it in three days. One engineer racking it up over ninety, then getting a hard question from his CTO he couldn't answer. Eleven days, two agents, nobody watching. The dollar figure survives every retelling. The details don't, because none of those teams had a number to point to, just dashboards someone was supposed to remember to check.
That's the actual lesson, not the number. Agent Ops Command Center is the gate every version of the story was missing: cost circuit breakers that block the next tool call before a session, project, or day goes over budget, built from nine months running Wiz, a real 24/7 autonomous Claude Code agent, in production. A watchdog that catches a silently stalled automation. An error registry with a self-fix loop. Kill switches, gated behind an allowlist and a confirm click, for when blocking isn't enough.
Self-hosted. Your infra, your data. $499 once, no subscription.
Free dashboards watch. This one acts.
What you have now
- A read-only chart of tokens and cost, updated after the fact
- A Slack ping once the bill already hurts
- No idea an automation went silent until something downstream breaks
- Errors living in your head, or scattered across old chat logs
- Killing a stuck process means SSHing in and hoping you remember the PID
Agent Ops Command Center
- Cost circuit breakers: block the next tool call before a session, project, or day goes over your limit
- A health watchdog: flags a stalled automation, an orphaned lock, a stuck heartbeat the moment it's stale
- An error registry: a self-fix loop from detect to root cause to resolved, logged where you can actually see it
- Kill and restart controls: allowlisted, confirm-gated, one click from the dashboard
- One JSON file you own: runs on your machine, against your data, forever
What it looks like
Where this sits
Agent Efficiency Kit
Drop-in hooks that tell you what your agent is doing: file-reread detection, a WebFetch fallback, a local dashboard, an optional Haiku audit pass. You read the numbers and decide what to do about them.
- Measurement hooks
- Local read-only dashboard
- No enforcement, no kill switches
Command Center
A control plane that acts on what it sees: breakers that block, a watchdog that flags, kill switches gated behind a confirm click. Built from nine months running a real fleet, not a weekend project.
- Cost breakers, watchdog, self-fix loop
- Kill/restart with allowlist + confirm gate
- Full source, self-hosted, watermarked license
Questions
Is this self-hosted, or do you see my data?
Fully self-hosted. It runs on your own machine as a local FastAPI server, reading your own ~/.claude session data and your own automation registries. Nothing gets sent anywhere. We never see your sessions, your costs, or your fleet.
What platforms does it run on?
macOS and Linux, Python 3.11 or newer. The installer registers a launchd agent on macOS or a systemd user unit on Linux so it survives a reboot. No Windows installer; run it inside WSL2 if that's your setup.
Do I get updates?
Your license covers the version you bought, self-hosted and yours to keep, with no subscription to lapse. Fixes and improvements to this release ship as a re-download through the same license key. There's no forced upgrade path and nothing stops working on a renewal date, because there is no renewal date.
What's the refund policy?
Try the live demo first, it's the real dashboard against real planted incidents, so you know exactly what you're buying before you pay. If it genuinely doesn't do what this page says once installed, email hello@theforge.ventures within 14 days.
What exactly will the breaker catch, and what will it miss?
Short version: it blocks Claude Code tool calls over your configured limit, and it can't see API calls made outside Claude Code (a raw script hitting the Anthropic API directly bypasses it entirely). Full, unsparing detail ships inside the zip as docs/HONESTY.md, a whole page titled "what the breaker will and will not catch." Ask for it before you buy at hello@theforge.ventures and we'll send it over.
What if Anthropic ships something like this natively?
Possible, and it would compress this product's shelf life. It wouldn't undo what you already have: a one-time purchase means you keep the source, the breaker, and the watchdog running on your own infra whether or not Anthropic ships a built-in version later. We'd rather sell something useful today than wait for a platform feature that might not arrive.
How is this different from the $49.99 Agent Efficiency Kit?
The Kit measures. It gives you hooks and a local dashboard so you can see what your agent is doing, and you decide what to do about it. Command Center acts: cost breakers that block a tool call, a watchdog that flags a stalled automation, kill switches for when blocking isn't enough. If you just want visibility, start with the Kit. If you want something that intervenes, this is that.