Agent Loop Replay
Free online agent loop replay tool. Paste an OpenAI or Anthropic messages array and the flat JSON is rebuilt into a turn-by-turn timeline - see the exact context the model received on each turn, which tool it called, and where the loop stalled.
When an agent calls the wrong tool on turn three, all you have is a flat messages array - and the question you need to answer is what the model was looking at on that turn. This tool rebuilds the array into loop structure: each turn frames the input the model received, the decision it made, and the results you fed back, with tool_call_id pairing shown instead of hunted for. The per-turn input snapshot is reconstructed from the final array by the standard send rules, not captured from the wire, and the tool says so on every snapshot. Everything is parsed locally in your browser.
Agent Loop Replay
Free online agent loop replay tool. Paste an OpenAI or Anthropic messages array and the flat JSON is rebuilt into a turn-by-turn timeline - see the exact context the model received on each turn, which tool it called, and where the loop stalled.
About this tool
When an agent calls the wrong tool on turn three, all you have is a flat messages array - and the question you need to answer is what the model was looking at on that turn. This tool rebuilds the array into loop structure: each turn frames the input the model received, the decision it made, and the results you fed back, with tool_call_id pairing shown instead of hunted for. The per-turn input snapshot is reconstructed from the final array by the standard send rules, not captured from the wire, and the tool says so on every snapshot. Everything is parsed locally in your browser.
FAQ
Common questions and answers about this topic.
How is a turn defined here?
Every assistant message opens a turn. What came before it is that turn's input, the assistant message itself is the decision, and the tool results that follow before the next assistant message are that turn's results. This matches how the loop actually runs: you send the history, the model decides, you execute and feed back, and the cycle repeats. Anthropic conversations carry tool results inside user messages instead of separate tool messages, so switching the API profile changes which messages count as results.
I already use LangSmith or Langfuse - what does this add?
Those platforms are built for continuous monitoring: you wire an SDK in, create an account, and traces flow to their servers. This tool takes the opposite side - it does not touch your runtime at all. When someone pastes an array into a bug report, when you pull a conversation out of a production log, or when you are looking at a trace from a service you do not own, there is nothing to instrument. Paste it here and the loop structure appears, with nothing leaving your browser.
Is the per-turn input the real payload that was sent?
No, and the tool says so on every snapshot. It is reconstructed from the final array by the standard rule that each request carries the history up to that point. Frameworks routinely break that rule - trimming old messages, injecting a system prefix, rewriting tool descriptions, summarising history. So treat the snapshot as the shape of the input, not a byte-exact copy. If your framework rewrites aggressively, the reconstruction shows what a plain client would have sent, which is still the right baseline for spotting where the model's view stopped changing.
Is my conversation data safe to paste?
Yes. Parsing, turn segmentation and pairing all run in your browser - nothing is uploaded, stored, or sent to analytics. Agent traces routinely carry customer data, internal tool arguments and API keys in system prompts, which is exactly why this tool has no server side at all. You can verify it by opening your network tab while pasting.