The execution replay engine for AI agents.

Record every LLM call, tool invocation, and error your agents make — replay it step-by-step like a video, fork from any decision point, and prove the fix.

Retrace turns every AI-agent run into a recorded, replayable, forkable artifact — so engineering teams can see exactly what happened in production, reproduce it on demand, and branch from any step to test a fix.

Add one wrapper to your agent. From then on every model call, tool call, retrieval, and error is captured as a span inside a trace. Open the dashboard and that run becomes an interactive timeline you can scrub through, fork, and re-execute.

Record → Replay → Fork → Prove — the loop AI engineers actually live in: something broke → find the exact step → change one thing → prove the fix → ship it.

• Record — one wrapper captures the full execution tree; no manual logging.
• Replay — step through a past run, or re-run it live to see what changed.
• Fork & cascade-replay — branch at any span, change the input, and the whole agent re-executes down the new path.
• Prove the fix — re-run a failed trace and get a verdict: improved, regressed, or unchanged.
• Detect failures — hallucinations, loops, goal drift, context loss, and schema violations, flagged automatically.
• Control spend — circuit breakers and budget ceilings stop a runaway agent before the next expensive call.
• Govern — PII redaction, prompt-injection scanning, and audit logs.
• Share — publish any run as an interactive, public "tape."

# Python
pip install retrace

# TypeScript
npm install retrace-sdk

# CLI (macOS, Linux, WSL)
brew tap retraceai-tech/homebrew-tap
brew trust retraceai-tech/tap
brew install retrace-cli

import retrace

retrace.configure(api_key="rt_...")

@retrace.record(name="research-agent", resumable=True)
def run_agent(prompt: str):
    plan = call_planner(prompt)      # captured automatically
    results = call_tools(plan)       # captured automatically
    return summarize(results)