# Runback

> The system of record for AI agents. Captures every decision an AI agent makes — the full context window, tool calls, and results at every step — then lets you replay any step from the exact captured request, gate a model or prompt change in CI before it ships, and export a signed, tamper-evident audit record.

Runback is not a read-only trace viewer (unlike LangSmith, Langfuse, Helicone, Braintrust, Arize, Traceloop, or Portkey). Its core differentiators: replay-from-step-N (re-issue or edit any captured LLM request and see how the model responds differently), a CI release gate that blocks a regression before deploy, and a hash-chained signed audit export that verifies independently of Runback.

Runback is proprietary software with a free, self-hostable Community edition (core capture/replay/audit/evals, single workspace) and a licensed Enterprise edition (multi-tenant, RBAC, SSO, alerting, long retention). See https://runback.dev/pricing.

## Product

- [How it works](https://runback.dev/how-it-works): the capture -> replay -> gate -> audit pipeline
- [Why Runback](https://runback.dev/why): the case for re-execution over observability
- [Integrations](https://runback.dev/integrations): Vercel AI SDK, OpenTelemetry (LangChain, CrewAI, LlamaIndex, raw SDKs)
- [Pricing](https://runback.dev/pricing): free Community edition through Enterprise
- [Live demo](https://runback.dev/runs): open a real failing run, no signup
- [Security](https://runback.dev/security): self-hosting, data residency, redaction
- [Spec](https://runback.dev/spec): the runback.cassette/v1 audit record format, independently verifiable

## Compare

- [LangSmith vs Runback](https://runback.dev/vs/langsmith): It shows what happened. It can't re-execute the step, gate your CI pipeline, or produce a signed audit record.
- [Langfuse vs Runback](https://runback.dev/vs/langfuse): Traces are read-only — there's no re-execution, policy enforcement, or signed compliance artifact.
- [Helicone vs Runback](https://runback.dev/vs/helicone): Proxy capture only sees inputs and outputs — no tool chains, no re-execution, no compliance artifacts.
- [Braintrust vs Runback](https://runback.dev/vs/braintrust): It's built for evals before deploy, not governance after — no re-execution of production incidents, no policy enforcement, no audit trail.
- [Arize AI vs Runback](https://runback.dev/vs/arize): It's ML observability, not agent governance — no re-execution, no policy enforcement, no regulatory compliance mappings.
- [Traceloop vs Runback](https://runback.dev/vs/traceloop): It's built to get traces out of your stack via standard OTel spans — not to re-execute a step, enforce policy live, or produce a signed audit record.
- [Portkey vs Runback](https://runback.dev/vs/portkey): It's built for production reliability at the request layer, not for reconstructing or re-executing what an agent decided across a multi-step run.

## Engineering blog

- [Three ways to wire an agent into Runback](https://runback.dev/blog/three-ways-to-wire-in-runback): How to pick between the Vercel AI SDK integration, an OpenTelemetry exporter, and the framework-agnostic manual recorder — with the actual commands for each.
- [Why replay-from-step-N is harder than it looks](https://runback.dev/blog/replay-from-step-n): Why capturing the request at the wrapLanguageModel boundary — not at your own call sites — is what makes replay mean something.
- [Redacting secrets before they ever leave your process](https://runback.dev/blog/redact-before-it-leaves): How @runback/redact scrubs API keys, JWTs, SSNs, and card numbers in-process, before anything is sent anywhere.
- [What a hash-chained audit trail actually proves](https://runback.dev/blog/hash-chained-audit-trail): The exact algorithm behind Runback's signed, append-only audit record — and what it means for EU AI Act and APRA CPS 230 compliance.
- [Debugging an AI agent with time-travel replay](https://runback.dev/blog/time-travel-replay-walkthrough): A walkthrough of a real failing agent — from error-first navigation to editing the exact captured request and replaying it.
- [We deterministically reproduced a real data race](https://runback.dev/blog/deterministic-data-race): A stress test of Runback's replay engine: reproducing a genuine, lock-free data race byte-for-byte, three runs in a row, proven in CI on every push.

## Other

- [RSS feed](https://runback.dev/feed.xml)
- [GitHub](https://github.com/letsRunback/runback)