AI Flight Recorder

AI Flight Recorder is an open-source developer tool by Matt Smith (GitHub: AllThingsSmitty) for recording, replaying, and inspecting interactions in AI applications. A one-line SDK wrapper captures prompts, streamed tokens, tool calls, latency, and cost into a DevTools-style timeline that can be exported as a portable .flight file. It is released under the MIT License.

What It Is

AI Flight Recorder is a debugging and observability toolkit for LLM-powered apps. Instead of piecing together console logs after the fact, developers wrap their provider client and record each session as a structured event stream. The sessions can then be viewed in a Next.js DevTools app, a hosted live demo, or a VS Code extension that opens .flight files in a timeline viewer.

How the Workflow Works

Developers install the SDK and start a session with FlightRecorder. Provider adapters wrap OpenAI, Anthropic, and Google Gemini clients, supporting streaming and non-streaming calls, and record every call automatically. Ended sessions can be saved through transports (in-memory, filesystem, or a custom API) and serialized to .flight files, which another developer can import and replay locally.

Inspecting Sessions

The DevTools app provides a Timeline tab with chronological events, a Waterfall tab showing streaming spans and tool call durations, and a Cost Analysis tab with token usage and estimated spend. Replay mode offers speed controls from 0.25× to 8× and a seek bar, with the token stream assembling in real time. Events can be filtered by type or searched by text.

Extensibility and Export

A plugin system lets custom observers hook into session start, event, and session end lifecycle points. A transport interface allows any storage backend. Sessions can also be converted to OTLP trace payloads with toOtlp for ingestion into OpenTelemetry-compatible backends such as Jaeger, Grafana Tempo, and Honeycomb.