DEV Community

Cover image for I'm building the trust layer between humans and AI agents
notenki
notenki

Posted on

I'm building the trust layer between humans and AI agents

The moment I realized something was wrong

I'd been using Claude Code every day for months.

Then one day I asked myself a simple question:
"How much have I actually spent?"

I had no idea.

Not approximately. Not roughly.
Zero visibility into what was happening.

That's when I understood the real problem.
It's not about cost. It's about trust.


AI agents are moving fast. Trust infrastructure isn't.

Claude Code can read files, write code, run commands,
and make decisions on your behalf.

But there's no dashboard. No audit trail.
No way to know what happened, why, or what it cost.

We're giving AI agents more and more autonomy
while building less and less visibility into their actions.

That's a problem that compounds.


What I'm building

I started with the most obvious gap:
token visibility for Claude Code users.

The data was always there — sitting in
~/.claude/projects/ as JSONL files.
Nobody told you it existed.

So I built a dashboard that reads it locally.
Nothing leaves your machine.

What I found in my own data:

  • $266 spent since March
  • One project eating 35% of my budget
  • 95.9% cache read ratio — happening every turn, invisibly

But the dashboard is just the beginning

The real goal is bigger.

Think about how we built trust around cars:

  • License (who's allowed to drive)
  • Inspection (is it safe to operate)
  • Insurance (who's responsible when something goes wrong)

AI agents need the same stack.

That's what I'm working toward:

  • Identity — which AI did what
  • Permissions — what it's allowed to do
  • Audit trails — what actually happened
  • Cost accountability — who pays for what

The dashboard is the audit trail layer.
AgentPass is the identity and permission layer.

Together, they're the foundation of
what I'm calling the AI trust infrastructure.


Why this matters now

We're at the moment right before the problem becomes obvious.

Most companies still say "ChatGPT is convenient."
But AI agents are already making decisions,
sending emails, placing orders.

The question isn't whether AI will act autonomously.
It's whether humans will be able to understand,
verify, and trust those actions.

I'm building the infrastructure for that trust.


Try the dashboard

npm install -g @notenkidev/claude-token-dashboard
claude-token-dashboard
Enter fullscreen mode Exit fullscreen mode

GitHub: https://github.com/notenkitoclient-cpu/claude-token-dashboard

If you're building in this space —
identity, permissions, audit, cost accountability for AI agents —
I'd love to connect.

Top comments (0)