Catch expensive sessions before they run. Find the waste hiding in the ones that succeed. Tie every session to whether the work was worth it.
On developer laptops (Claude Code, Cursor, Codex) and inside your product (one line per LLM call).
Here’s where that 40% goes.
On a developer’s laptop
A task, phrased loosely. Twenty turns of drift-and-redirect. The code works, and nobody looks because it succeeded — at 3x the tokens it needed.
In production
A feature got bumped to the premium model during an incident, and nobody switched it back. Retries succeed at triple the cost; one workflow takes forty steps instead of five for certain customers.
Every monitoring stack is built to watch for failure. This isn’t one.
Questions every VP of Engineering, Head of Platform, and VP of Product needs to answer this week, and can’t:
The five-step loop runs the same way on both surfaces. The install differs. The behavior is identical.
Your Local is free, open source, and yours alone. Your Apps runs on the Enterprise tier — the same loop, extended from one developer to your whole team.
Every coding agent on every developer's machine.
Install: pip install aiwatcher-cli. Zero code changes.
Every AI feature inside your product.
Install: npm install aiwatcher. One line per LLM call.
At Enterprise scale, also:
AIWatcher Local is not a demo. It’s the full loop for one developer. Enterprise is what turns individual practice into team policy.
The detection layer
AIWatcher watches for a small set of moments across the loop, on both surfaces, and gives you a clear move for each one.
Prompt scored as likely to wander the repo. Preflight flags it, offers a scoped brief with the file list narrowed, and shows the savings badge for what the tightened version avoided.
Context filling with junk from wrong explorations. AIWatcher offers a handoff brief: current state, decisions made, next step, so you restart clean instead of grinding through a degraded session.
Bulk delete, migration against live data, force push. Hard gate with Allow, Brief, or Cancel, enforced across Claude Code, claude.ai, Codex, Cursor, and Claude Desktop.
Every session ledgered and linked to git. Which sessions produced commits, which changes survived the week, what the surviving work actually cost.
The loop learns which briefs and gates paid off and adapts scoring per repo and per developer. Loose-prompt patterns that keep getting caught become defaults; brief templates that saved tokens become suggestions.
Every LLM call tagged by feature at the wrap. The summarizer, the extraction pipeline, and the classifier show up as separate lines, not one bucket.
Same wrap, customer_id in the productContext. The four accounts driving the majority of your inference spend rank themselves.
Calls succeeding on the third attempt at triple the cost surface as a gap between cost per request and cost per success, not a healthy line.
Filter spend by model, workflow, and date. The premium model nobody switched back after an incident is the first row.
The loop tracks which caps, model routes, and workflow limits reduced spend without hurting outcomes, and recommends the ones worth keeping. What started as one-off fixes become the policy defaults.
ccusage, vendor consoles, OTel exports. They tell you what happened after the money is spent. They're built to investigate failures, and they have nothing to say about the sessions that succeeded expensively.
Blocks and routes API traffic on the one path routed through it. It never sees the coding agents on your developers’ laptops, where half your AI work happens.
Claude Code asks permission for every command until the prompts get turned off. A human watching the terminal is the actual control plan at most companies, and it defeats the point of agents.
AIWatcher replaces none of your observability and all of your blind spots.
The loose prompt that would have burned 3x the tokens? Caught at preflight and rewritten into a scoped brief before the first token is spent.
The 40% bill jump nobody could explain? A filter, not an investigation: spend by feature, model, customer, and workflow. The premium model nobody switched back is the first row.
“Was any of it actually worth it?” Every session lands in a ledger linked to git. Cost per surviving change on the laptop. Cost per success in production.
PLAN
Score the prompt; rewrite loose ones into scoped briefs before a token is spent.
WATCH
Track context and quota; hand off cleanly when either runs out.
CONTROL
Gate destructive commands and looping sessions, across every tool, not inside one.
PROVE
Ledger every session, linked to git: cost per surviving change, not cost per token.
IMPROVE
Learn which prompts, habits, and controls pay off; preflight gets sharper every session.
The same loop applies to the AI running in your product, via a one-line SDK per call.
The shift, in numbers
of AI-native engineering teams now run 3+ AI models in production.
23% run 6+. Most have no per-model, per-team, or per-feature cost attribution today.
Source: Datadog, State of AI Engineering 2026
of platform engineering and SRE leaders will own production AI systems by 2028.
Most aren't equipped to manage the cost and behavior side yet.
Source: Gartner, Innovation Insight: LLM Observability, July 2025
Why we built this: my own $400 lesson
“We ran five Claude Code agents on a thirty-second loop for a meeting-notes pipeline. It ran unattended for ten days: 31,762 session files, 6,352 dispatch cycles, about $400 in tokens. First signal was an email from Anthropic saying I’d burned through a budget that should have lasted months. The agents were doing exactly what I built them to do. A dashboard would have shown me the damage. A control loop would have stopped it on day one.”
Danny Lo, founder, AIWatcher
We’re working with a small number of design partners in 2026. If your team is running multiple AI tools and feels the cost and visibility gap, we’d like to talk.
We reply within 48 hours. No newsletter, no marketing automation. Just a real reply from one of us.
Pricing isn’t public yet. We’re onboarding design partners first, then publishing plans.