Trace replay & time-travel

Trace replay lets you take any production trace and step through its recorded span sequence — every LLM call, tool call, and guardrail verdict, in order — for offline inspection and debugging.

What replay does. Stepping through a recorded trace (tlane replay <traceId> and the MCP replay_trace tool) is live — it reads the stored spans as-is and does not re-execute any model or tool. Shadow-fork replay — re-running a trace against a different model, prompt version, or tool response and diffing the outputs, plus the dashboard “Time-travel” UI — is not available. Inspecting stored evidence is not a fresh execution or a guarantee that the original behavior can be reproduced.

From the CLI (live)

tlane replay <traceId> fetches the trace’s stored spans and renders the step sequence in your terminal — read-only, no re-execution:

# Step through a recorded trace (table view)
tlane replay abc123

# JSON output for piping into your own tooling
tlane replay abc123 --format json

# Point at a self-hosted gateway
tlane replay abc123 --endpoint https://gateway.example.com --token tlane_...

From the MCP server (live)

The replay_trace tool returns the full stored structure of a trace (ordered spans with their LLM/tool attributes) directly in your Claude conversation. Read-only — it does not re-execute any model or tool:

mcp.tracelane.replay_trace({
  trace_id: "abc123",
  include_tool_calls: true,
})

Recorded-trace output

{
  "trace_id": "abc123",
  "spans": [
    {
      "span_id": "span-1",
      "kind": "llm_request",
      "model": "claude-sonnet-4-6",
      "latency_ms": 823,
      "input_tokens": 512,
      "output_tokens": 128,
      "guardrail_verdict": "allow"
    }
  ]
}

Executable replay is not available

Earlier versions of this page described shadow forks and a time-travel scrubber as roadmap capabilities. They are not delivered by tlane replay or replay_trace. Neither accepts model overrides, prompt overrides or injected tool responses, and neither creates a new execution to compare. For a fresh test of a prompt change, use evaluation runs with captured inputs or a curated dataset. Review the case source and reference coverage before treating a passing evaluation as evidence of improvement.

Get started

SDKs

Integrations

Concepts

Features

Tools

Audit & compliance

Operations

Migrations

Comparisons

Trace replay & time-travel

From the CLI (live)

From the MCP server (live)

Recorded-trace output

Executable replay is not available

​From the CLI (live)

​From the MCP server (live)

​Recorded-trace output

​Executable replay is not available

​Related

From the CLI (live)

From the MCP server (live)

Recorded-trace output

Executable replay is not available

Related