Tokenwatch
The agent stuck in a loop
Problem
An agent hits an error it cannot resolve and tries the same thing four hundred times. Nobody is watching because the whole point was that nobody had to. You find out from the bill, or from the rate limit, or from the support ticket.
Why now
Unattended agent runs became normal in 2026, and the failure mode they share is not being wrong, it is being wrong repeatedly and confidently at machine speed.
Solution
It watches the shape of a run rather than its output: the same tool called with the same arguments, the same error class recurring, cost climbing with no progress. It kills the run and tells you, with the loop it saw.
Who it is for
Anyone whose agents run without a human in the room, which was the entire point of running them. The failure they are buying against is not an agent being wrong, it is an agent being wrong four hundred times at machine speed while nobody watches. They find out from the bill, the rate limit, or a support ticket, and all three arrive after the money is gone.
What you could earn from Tokenwatch
Say you build Tokenwatch and charge your own customers $25 a month, the price the score suggests. The numbers below are the revenue that would come from them. This is money you earn, never a fee you pay UpgradIQ, and we never take a cut. It is arithmetic off the price, not a promise the customers arrive on their own.
See what Tokenwatch produces
It watches the shape of a run rather than its output: the same tool called with the same arguments, an error class recurring, cost climbing with no progress. It kills the run and tells you, with the loop it saw.
Your build kit for Tokenwatch
Start the build and the kit opens two steps in, then drip-feeds the rest one phase at a time. Every phase is a copy-paste prompt for your agent with its own verification before the next unlocks, so Tokenwatch moves from an empty folder to a running MVP without you ever holding the whole plan in your head.
Progress saves across every device, and the workspace always reopens on the exact step you left, so a build that runs over several weeks survives them instead of dying in a browser tab you eventually close.
How you actually build it
Connect your coding agent once. It pulls the current phase straight from your kit over MCP: the exact prompt, the outcome to aim for, and the check to run before it moves on. You review and ship, it does the typing.
No copy-paste juggling, no thread to lose. The kit stays the single source of truth, and your agent works through it one phase at a time, verifying its own work at each step before the next phase is unlocked.
Take the context file, free
This is the AGENTS.md that keeps your coding agent on scope for the entire build: what Tokenwatch is, who it is for, the stack, the non-goals, and the rules it must follow. Copy it into your project now, no account needed, and your agent is briefed before you write a line.
# Tokenwatch The agent stuck in a loop ## What this is It watches the shape of a run rather than its output: the same tool called with the same arguments, the same error class recurring, cost climbing with no progress. It kills the run and tells you, with the loop it saw. ## The problem it solves An agent hits an error it cannot resolve and tries the same thing four hundred times. Nobody is watching because the whole point was that nobody had to. You find out from the bill, or from the rate limit, or from the support ticket. ## Who it is for Anyone whose agents run without a human in the room, which was the entire point of running them. The failure they are buying against is not an agent being wrong, it is an agent being wrong four hundred times at machine speed while nobody watches. They find out from the bill, the rate limit, or a support ticket, and all three arrive after the money is gone. ## Stack Cloudflare Workers + D1 ## Non-goals Do not build features outside the core job above. This product exists to do one thing well. Every extra feature is a reason it does not ship. ## Rules for the agent - Build one phase at a time. Do not run ahead into later phases. - Validate input at every boundary where data enters. - Never put a permission or plan check in client code. - Ask before adding a dependency that a few lines would cover. - If something is ambiguous, say so rather than inventing a requirement. --- Context file from UpgradIQ · idea tokenwatch · prompts kit-v4
Then start free to get the ten-phase kit and run the first prompt.
The path to a running MVP
Foundation
Scaffold the stack, then wire the data model and sign-in. It is the skeleton Tokenwatch stands on, and it goes up in an evening rather than a lost week of config.
The product
Build the core feature end to end, then wire how it charges. This is where Tokenwatch finally turns into something that a stranger will actually pay you for.
Ship it
Test and harden, deploy it live, then open the launch-day toolkit that lands the first ten users who prove it.
Is Tokenwatch for you?
- You want revenue reasonably soon, not a science project
- You can give it evenings and a couple of weekends
- You can reach Teams running agents unattended
- You are not ready to talk to a first real customer
- You want it fully passive, with zero maintenance
- You expect it to find its own customers without you
The 3 sections that decide whether you should actually build it, plus the full kit, are one click away.
3 more sections decide whether you build Tokenwatch
You have the case and the plan. Pro opens how it makes money, the one risk that kills it, and the build notes, plus the 10-phase kit that takes Tokenwatch to a running MVP, on this idea and all 200.
How it makes money
Main risk
Build notes
What is free here, and what Pro unlocks
Frequently asked
Is Tokenwatch actually validated?
It scored UIQ 81 with a GO verdict, weighed across eight sub-scores on the same rubric as every idea. GO means 75 and above, so demand and buildability were judged, not guessed.
What stack does it use?
Tokenwatch is built on Cloudflare Workers + D1, chosen because a coding agent can carry most of it. You are not fighting infrastructure, you are shipping the feature that matters.
How long to a running MVP?
The kit is ten phases, drip-fed one at a time so you always know the single next step. An evening and a weekend for the lighter builds, a few focused weeks for the deeper ones.
What does the $25 a month mean?
It is the price the model suggests Tokenwatch could charge its own customers, not a fee to us. It is here so the money question is answered before you build, not after.
Which agent do I need?
Any capable coding agent: Claude Code, Cursor, Lovable, or Bolt. Every phase is a copy-paste prompt with its own check, so the agent does the typing and you keep the judgement.
What is free versus Pro?
The whole case above is free: the problem, timing, solution, buyer, and the full score. Pro adds how it makes money, the one risk that could kill it, and the full kit that carries Tokenwatch from an empty folder to a running MVP.
More like Tokenwatch
All 25 Dev tools ideasBuild Tokenwatch, not just read about it
You have the problem, the timing, the solution and the buyer. A plan opens how it makes money, the risk that kills it, and the kit that takes it to a running MVP.
Start building Tokenwatch