正在加载视频...
视频加载失败
Finally, I can understand what the coding agents are doing 👀 🎞️ zoetrope — Watch a Claude Code session as a live flow graph 💯 See the tool calls, workflows token usage & more in the terminal ⏳ Supports replaying an entire session with time travel 🦀 Written in... show more
21 条评论

try it on the web via WASM: provably offline — no network stack in the dependency tree. transcripts are parsed in your browser: on chromium a file handle keeps the graph live, elsewhere it's a snapshot of dropped files. they never leave your machine.

@ratatui_rs This would be really cool over @atomicdotdev changes and reading the provenances from any agent not just Claude code

Using Claude's own transcript for this is nice. You get the full replay without adding another layer to the agent loop.

Thanks for sharing this, @orhundev. This is one of the best designs I've seen for a terminal UI. Great work. Is this processing native .claude JSON for the realtime data?

🫡🫡

making an agent session legible is underrated. terminal scroll hides what actually happened, which tool ran, what it read, where it looped. a live tui view turns the black box into something you can supervise and interrupt mid-run. ratatui is the perfect tool for exactly this. love it.

Replay with time travel only works if the recorded tool results are in the log, not just the calls. We hit that building agent harnesses: replay the calls and you get a different session. Does it persist the responses too?

This is so cool. Going to make it work for OMP

Seeing the agent flow visually is incredibly useful.

A live flow graph of tool calls is a game changer for debugging agent loops. I've wasted hours on opaque sessions—being able to replay with time travel would've saved me. Does it handle parallel tool calls well?

Been curious about this kind of workflow coding agents get more complex, seeing the steps behind the output becomes really useful.

Watching the flow graph matters more once you've built one that breaks. My own content pipeline ingests structured JSON, pulls images through the Pexels API, and generates slideshow and reel content. I keep a running bug list on it because production is never finished.

时间旅行重放这功能也太适合复盘代理抽风现场了哈哈

the time travel replay is the part I'd use daily. tool call graphs are pretty until you need to know which branch burned the tokens

Visibility is half of coordination. A live flow graph answers ‘what is it doing?’; we’re working on ‘how do agents hand context to each other?’ with persistent terminals and lobbies. These layers would pair nicely.

@grok Is this any better than MLflow

I built a version of this that allows you to view all your previous sessions, resume specific sessions, share sessions, all with an MCP perver that lets your agent interact with the graph, highlighting nodes to answer questions, etc.

This is the missing piece for agent trust. Nobody fully adopts something they can't see working. A live flow graph of tool calls turns 'the agent is doing something' into 'I can watch exactly what, and stop it if it's wrong.'

Amazing

Seeing token usage per tool call live is the sleeper feature, you catch a runaway loop before it quietly burns the whole budget.

A flow graph could reveal where an agent repeatedly calls the same tools or burns tokens without progress. The most useful view may be the gap between its intended plan and actual actions.
