AgentDeck 2.0: Claude Code, Codex and Antigravity in One View
Back in June I wrote about AgentDeck: a small dashboard that runs on your own machine and shows you the conversations Claude Code and Codex quietly write to disk. Nothing leaves your computer.
I’ve used it every single day since, and the annoyances piled up. With three or four projects open, looking at one agent at a time got old fast. Finding a conversation from last week took too many clicks. And the embarrassing one: a CLI vendor quietly changed its file format, and I only noticed days later when a screen broke.
So 2.0 is pretty much a rewrite. Here are the things I actually feel every day.
(The video and a one-page overview are at jimmyblog.site/AgentDeck.)
Switch agents like browser tabs
The biggest change: the UI is now Chrome-style tabs. Every conversation is a tab, whether it’s Claude Code, Codex or Antigravity, whatever folder it lives in. Switching agents or projects feels like switching web pages.
Hit ⌘K (Ctrl+K on Windows) to search any conversation. Can’t remember the name? It shows the first thing you said in each one, and that’s usually enough.
The same repo often has both Claude Code and Codex conversations, so you can now group them into a workspace, and pin whatever you’re busy with. The sidebar finally stopped being a mess.
Google Antigravity is in
Google’s agy CLI is the third provider now, with exactly the same screens as the other two. Its file format isn’t documented anywhere, so I spent a while matching fields one by one until its conversations, sub-agents and token usage all showed up.
Insights: how you work with agents
This page isn’t about tokens. It’s about you: how many of the last 30 days you used an agent, what time of day, how long a conversation usually runs, which projects have gone quiet. Mine told me I do almost all of it in the morning, and Tuesdays are the busiest.
Don’t know how to configure it? Let the agent do it
Adding a hook or an MCP server to Claude Code or Codex used to start with reading docs. Now you just type what you want on the settings page, in plain words, like “set this project up for me”. AgentDeck opens that CLI’s own terminal and hands it your request, the current config file and the official docs.
I made it ask before it acts: it explains what can be configured, asks whether you want each piece, even checks the skills ecosystem for something ready-made, and only edits once you say yes. AgentDeck itself never calls a model or touches your files.
Know the day a vendor changes its format
This is the one I wanted most. Every CLI keeps its records in its own folder, in its own format, and nobody is in charge. AgentDeck now has a written-down spec of what a provider looks like, and on every start it checks the newest records against it. A field or a type it hasn’t seen before puts an exclamation mark on that folder’s chip, and one click shows what changed.
This week it already caught a few new record types Claude Code and Codex had slipped in.
Also
- A new dark theme called Midnight. Colours, how much of a path to show, how long the home lists are — all adjustable.
- Every demo screenshot comes from fake data. No real conversation of anyone’s is in the repo.
- There’s a one-page overview now: jimmyblog.site/AgentDeck.
Same as last time: if you work with AI coding agents every day, come try it and drop a star. PRs for new providers are very welcome.