Image: GitHub
Why it mattersAnyone running Claude Code with subagents can now see which agent did what and when, and rewind to any point in a session, which turns "the agent did something odd" into a specific moment to reproduce.
The developer furkankly published zoetrope, a local viewer that renders a Claude Code session as an interactive flow graph, and the repository has gained 766 stars. It reads the JSONL transcripts Claude Code writes into ~/.claude/projects/ and draws the main agent, any spawned subagents, and their tool calls as a graph that updates live while the session runs, or replays afterwards.
What it shows
The graph is a hierarchy: the top-level agent, then each subagent it invoked, then the individual tool calls each one made. A timeline you can drag through runs alongside, indexed by events rather than wall-clock time, so a long-running session is shortened to the moments where something happened. Clicking any node opens a detail panel with that agent's prompt, its reasoning, the model it ran on, and the full tool call record. Going back in time is built in: the state at any earlier point in the session can be reloaded without replaying the whole thing.
How it runs
Zoetrope is MIT-licensed and is distributed as a Rust binary. It runs entirely on the local machine, in a terminal user interface, and there is also a WebAssembly build that renders the same graph in the browser. The dependencies do not include an HTTP client, so no session data leaves the machine. It is available through Homebrew, Cargo, and prebuilt binaries.
What it changes
Anyone running Claude Code past a few short conversations knows the class of question the tool answers: "did this run spawn the subagents I expected, and which one is causing the delay?", "which tool call produced this odd bit of output?", "can I look at the session state just before the model chose to edit that file?". Until now the answer has been to reread the raw JSONL by hand, or to accept that the run could not be understood. Zoetrope turns those into a graph and a timeline, and adds the view of the state at any moment that makes it possible to inspect the middle of a run without rerunning it.
For a team that has moved past single-shot prompts into multi-agent workflows, this is the kind of debugging tool that changes what "review" of an agent run looks like. A run that produced the wrong answer becomes a graph a second engineer can read, rather than a folder of JSONL. Zoetrope is early, MIT-licensed, and running locally with no network access, so trying it costs nothing more than the disk space for the binary, and the data stays on the developer's machine. That combination is what has attracted 766 stars in a short time: the same audience that already runs Claude Code sessions large enough to be confusing has an immediate use for a viewer built for those exact files.
Source
Primary source: furkankly/zoetrope on GitHub.
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.
