AI writes the code. But who keeps track?

AI coding observability is the missing layer in every modern development workflow. Tools like Claude Code are changing how we build software — but most developers quickly realize: without visibility, documentation, and real numbers, productivity gains turn into chaos.

I've been building software for over 20 years. Over the past months, I've integrated Claude Code deeply into my daily workflow and noticed something that bothered me: I never knew exactly what my AI sessions actually cost, where I needed to intervene, and what went well or poorly.

Most developers face the same AI coding observability gap — and don't notice until they're deep in rework, burned tokens, and sessions they can't explain.

So I built what was missing.

The problem: AI coding without a cockpit

Imagine flying a plane with no instruments. You can see the horizon, but nothing else. That's what AI-assisted programming feels like without observability. Claude Code completes tasks, writes files, runs commands. But: What did it cost? How often did you need to correct it? Which projects burn the most tokens?

"AI is an aggressive autocomplete — not a responsible senior engineer. Without oversight, it produces chaos instead of productivity."

These questions went unanswered for too long. Yet they're critical — not just for your own efficiency, but to honestly assess: where does AI actually help me, and where am I just wasting time and money on rework?

SessionPilot: The cockpit for your AI workflow

SessionPilot is a self-hosted developer dashboard that closes exactly this gap. It captures Claude Code sessions in real time, stores them in a database, and makes them analyzable — by cost, project, tool usage, and quality.

Live Session ViewerTrack Claude Code sessions in real time with Markdown rendering and tool details

Cost DashboardAPI costs by model, project, and time period — Opus, Sonnet, Haiku

Session RatingOK / Needs Fix / Reverted / Partial — with notes and cross-session links

AI TimesheetsAutomatic time tracking based on real session data

Docker & ProjectsContainer status, Gitea integration, auto-discovery of projects

Export & SearchExport sessions as JSON, Markdown, HTML, or XLSX

What this means in practice

Since using SessionPilot, I can see in black and white: some projects cost three times more tokens than others — not because the tasks are harder, but because the CLAUDE.md was incomplete. I can see which sessions I had to revert, and when Claude Code works reliably on its own.

This isn't a toy. It's real data about my own productivity — and the first honest answer to the question: does AI coding actually pay off?

Self-hosted, no SaaS, no dependencies

SessionPilot runs as a Docker container on your own infrastructure. No cloud dependency, no data sharing, no monthly subscription. Just Python, Flask, PostgreSQL — and a clean dark UI that works out of the box.

What good AI coding observability actually looks like

Most developers think observability means logs. But for AI-assisted workflows, it means something different: knowing why a session went wrong, not just that it did.

Good AI coding observability answers four questions:

Cost — What did this session actually cost in tokens and time? Which projects burn the most resources?

Quality — Did the AI output need correction? How often did I revert? What was the rework rate?

Context — Was my CLAUDE.md complete enough? Did the model have the right information to succeed?

Patterns — Which task types benefit most from AI assistance — and which ones consistently fail?

Without answers to these questions, you're not using AI coding tools. You're just hoping they work.

SessionPilot was built to answer exactly these four questions — session by session, project by project, with real data instead of gut feeling.

SessionPilot is open source.

GitHub: github.com/web-werkstatt/session-pilot

Homepage (coming soon): session-pilot.com

Feedback, issues, and real-world experiences are very welcome.