Image: GitHub
Why it mattersRunning two coding agents at once is now common. Tabbing between terminals to see which one needs you costs time. One page removes that.
Running two coding agents in two terminals, you lose track of which one is waiting for you. whatsmyagentdoing opens one browser tab. It shows the live status of every Claude Code, Codex and Gemini CLI session on your machine. The author is Charonyuu. The repository picked up 141 stars inside a day of first commit.
How it works
whatsmyagentdoing reads the session logs the three tools already write. Claude Code writes one JSONL file per session under ~/.claude/projects/. Codex writes one under ~/.codex/sessions/. Gemini CLI writes one under ~/.gemini/tmp/. The dashboard watches those files for new lines. It turns each one into a status and sends it to the page. The server binds to 127.0.0.1 only, so nothing leaves the machine. The project has no runtime dependencies and needs Node.js 18 or newer.
Each session shows one of six statuses. Thinking means the agent is reasoning or reading a tool result. Running means a tool call is in flight, and names which one. Writing reply means a reply block has landed and the turn is not over. Waiting for you means the agent asked you a question. Done means the turn is finished. An error state covers interrupted, failed and idle.
Two views, both without extra packages
The default view is a hand-drawn 2D isometric office. Each agent sits at a cubicle desk while it works. A thought bubble shows up while it is thinking. Code scrolls on the monitor while a tool runs. The agent stands up and waves when it needs you. Recently finished agents wander through the pantry, meeting room, phone booth or arcade cabinet.
A High setting swaps that for a 3D version with real-time lighting. The 3D version uses Kenney's CC0 Mini Characters and Furniture Kit. The author says the 3D bundle only loads when you pick that view. It is three.js at about 180 KB gzipped, plus 1.3 MB of models. Three extra scenes are 100 to 160 KB each. The scenes are Pirate Cove, Café, Space Station and Castle. They are fetched on demand and pinned by SHA-256.
Limits the author states
The README is clear about what the tool cannot do. Each agent writes to its log once per reply block, so word-level typing never shows up. "Writing reply" means the reply block has landed, with the turn still open. Claude Code leaves permission prompts out of its logs. A tool waiting on your approval shows as "Running" until you answer. Gemini CLI writes its reasoning together with the reply. Between your prompt and the first tool call it only shows "Thinking". The log formats are internal to each vendor. An update by one of the three could degrade the statuses. Google Antigravity keeps its conversations in an undocumented binary store, and it has no support yet.
A single view of running agents is a problem the next layer of developer tools has to solve. Each tool has its own terminal, its own notifications and its own history. A developer running three at once ends up tabbing between them to answer one approval prompt. A read-only reader of the logs the tools already write is the smallest fix that works. It is released under MIT.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.