Your Agent Sessions Are Disappearing
I use Claude Code as my primary interface to my computer. Not just for coding—for everything. Research, writing, system administration, media processing, calendar management. If it can be done from a terminal, Claude Code does it for me.
This means that a huge amount of my daily work—the conclusions I’ve reached, the investigations I’ve done, the decisions I’ve made—lives scattered across hundreds of agent sessions. And unless I do something about it, all of that just… disappears. It’s the browser history problem all over again.
The Browser History Problem, Redux
Browser history has never been good. You visit a page, read something useful, close the tab, and it’s gone. You vaguely remember that you saw something relevant three days ago, but good luck finding it.
That’s exactly where we are with coding agent sessions right now. Every Claude Code or Codex session is a rich conversation full of context, decisions, and insights. And when you’re done, it gets serialized into a JSONL file in a hidden directory and effectively ceases to exist.
I have months of these sessions. They contain things I’ve figured out about my own codebase, debugging insights, architectural decisions, research results. It’s a gold mine that I can’t access.
Why Existing Tools Didn’t Fit
There’s no shortage of agent orchestration tools. Sculptor, Conductor, Claude Squad—they’re all good projects with thoughtful design choices. But they all share an assumption that doesn’t match how I work: you start by picking a single git repo.
That’s not how I use coding agents. My Claude Code sessions start at my home directory. I have a global store of memories, skills, and workflows in ~/AGENT that I want applied everywhere, regardless of which repo I’m in. And my work routinely spans multiple repositories—a feature implementation might touch three or four small repos that are all connected. The single-repo model is a straitjacket.
The other mismatch is about what I actually need. These tools focus on orchestration—spinning up isolated environments, managing concurrent agents, containerizing for security. Those are real concerns. But they’re not my concern. What I need is sense-making. I need to look back over my last week of agent sessions and understand what happened. I need a high-level timeline that says “Monday: debugging the auth flow. Tuesday-Wednesday: refactored the scanner. Thursday: three separate research sessions on MCP servers.”
Orchestration is a nice-to-have. Understanding your own history is a must-have.
Agent Kitchen
So I built Agent Kitchen. It’s a local web dashboard that reads your Claude Code and Codex session history, generates LLM-powered summaries, and presents everything as a browsable, searchable timeline grouped by repository.
The core workflow is simple. agent-kitchen index scans your sessions and generates summaries (using Claude Haiku via the Agent SDK, so it runs on your Claude Max subscription—no separate API bill). agent-kitchen web serves the dashboard. That’s it.
What you get is a reverse-chronological view of every folder and repo you’ve worked in, with your sessions grouped and summarized. Each repo group has a timeline at the top—an LLM-generated narrative of what you’ve been doing over the past days and weeks. You can click any session to resume it—Agent Kitchen spins up a terminal emulator right there in the browser. You can start new sessions with one click. There’s a time slider to filter by recency, fuzzy search, and vim-style keyboard navigation.
What This Is Really About
Agent Kitchen is a small tool that scratches a specific itch. But the itch points at something bigger.
As coding agents become our primary work interface, the history of our agent sessions becomes the history of our work. It’s not just conversation logs—it’s an institutional record of decisions, investigations, dead ends, and breakthroughs. We’re generating this record every day, and right now we’re throwing almost all of it away.
Browser history became genuinely useful when someone realized it wasn’t just a list of URLs—it was a map of your information-seeking behavior. Agent session history has the same potential. It’s not just a list of conversations. It’s a searchable, summarizable record of everything you’ve figured out.
The tools we build around agent sessions shouldn’t just help us run more agents in parallel. They should help us understand what we’ve already done. That’s the harder problem, and the more valuable one.
Try It
Agent Kitchen is open source and on PyPI. You can run it without installing anything: uvx agent-kitchen web.
One tip: by default, Claude Code deletes sessions older than 30 days. Set cleanupPeriodDays to something large in your ~/.claude/settings.json to keep them around. Your old sessions have more value than you think.