# Retrace > Retrace records every LLM call and error as a span inside a trace for AI developers. Users can replay a past run step by step and fork it to change inputs, comparing the original and new execution paths side by side. - Type: AI tool - Page: https://completeaitraining.com/ai-tools/retrace/ - Published: 2026-07-01 - Updated: 2026-07-03 - Author: Jeroen Erne - Jobs: IT and development - Topics: Developer Tools, Analytics - Reading time: 3 min ### About Retrace Retrace is a debugging tool that records, replays, and forks AI agent executions. It captures every LLM call, tool invocation, and error as a span inside a trace, then lets you play back the run step by step. The tool is free for up to 1,000 traces per month. ### Review Debugging an agent that fails at step eight of a twelve-step chain usually means re-running the whole thing just to test a single fix. Retrace records the full trace and replays it like a scrubbable video, so you can find the exact break point, fork from there, and let only the downstream LLM calls re-execute live. It treats forking more like a git branch than a prompt re-run - pre-fork steps stay on tape, and tool outputs are replayed instead of re-fired. #### Key Features - **Record and replay traces.** Every LLM call, tool call, and error appears as a span in a nested, scrubbable timeline. You replay the run step by step to see what happened. - **Fork with live re-execution.** Pick any span as a fork point. Steps before the fork replay from the recording; downstream LLM calls run live, while tool calls replay their recorded outputs to avoid side effects. - **Output override and mock tagging.** Edit the recorded output at any overridden span. The edited value gets tagged as "mocked" and flows into the live downstream steps. A per-step map (span id → output) supports scripted sweeps of different "what-if" values. - **First-divergence diff and verdict.** After a fork, the tool highlights the first step where the new path diverges from the original and classifies the result as improved, regressed, or unchanged. - **Multi-agent topology view.** Spans carry agent IDs and roles. An agent-topology graph shows how sub-agents hand off work, with inter-agent detectors flagging reasoning/action mismatches. #### Pricing and Value Retrace currently operates on a free tier that includes 1,000 traces per month. No other pricing plans have been published yet. #### Pros - Tool calls are not re-executed during a fork, so debugging production runs with payments, emails, or checkout flows doesn't trigger duplicate side effects. - The fork-and-replay model allows testing a fix at a specific step without re-running the entire agent chain from scratch. - Override maps let you script "what if" scenarios by supplying alternative tool outputs without touching live APIs. - Traces display original and forked paths side by side, which helps isolate exactly where a change altered behavior. - The replay timeline groups nested calls, so multi-step workflows stay readable instead of becoming a flat list of events. #### Cons - Replay matches tool calls positionally rather than by name or arguments. When a forked model calls the same tool with different parameters, it gets the old taped answer, and a new tool call has no tape entry at all. This makes the tool less authoritative for agents that frequently change tool call arguments. - Handling very large traces with dozens of tool calls and multiple sub-agents is still being actively hardened, so performance and clarity may drop on unusually complex runs. - Teams whose debugging workflow requires actual live re-execution of side-effecting tool calls with fresh arguments will find the tape replay model insufficient. The tool is not a fit when real-world API side effects must fire during a fork. Retrace works best for developers building agents that need to inspect prompt chains and non-deterministic model behavior without constantly re-running expensive or side-effecting workflows. It's particularly useful for catching looping tool calls, testing prompt or model changes mid-run, and performing regression checks with override values. Teams that can work within the positional replay limitation and don't need live side-effect re-execution will get the most out of its current design. --- This is the Markdown form of https://completeaitraining.com/ai-tools/retrace/. The whole site is mapped at https://completeaitraining.com/llms.txt.