Matt Harrison, a UK-based software and machine learning engineer, has released Loupe, a lightweight in-memory tracing dashboard for LLM applications and agent systems. Published as the npm package @mtharrison/loupe, the tool targets the local development phase of agent building — the rapid iteration cycle where developers need structured answers about what payloads were sent, what tool calls were made, and why a run felt slow, without the overhead of spinning up databases or external services. Traces are stored entirely in memory and cleared on restart, a deliberate design choice Harrison frames as a feature rather than a limitation.
The most practically useful capability is the session-first conversation timeline, which renders user messages, assistant responses, tool calls, and tool results in a single view. Developers debugging multi-step agent interactions no longer have to <a href="/news/2026-03-14-rudel-open-source-analytics-dashboard-for-claude-code-sessions">mentally reconstruct what happened</a> from scattered logs. The tool also tracks first-chunk latency on streaming responses and rolls up cost data when usage metadata is available — two figures that are awkward to extract from raw console output but that matter immediately when an agent loop starts misbehaving. The full spec covers additional details including actor, stage, and tag context metadata, with sensitive headers like Authorization and x-api-key automatically redacted before storage.
Integration requires minimal code changes. Developers using the official OpenAI Node.js client can wrap it with a wrapOpenAIClient() helper that intercepts calls transparently, while wrapChatModel() targets apps with custom invoke() and stream() interface shapes. A lower-level lifecycle API — recordInvokeStart, recordInvokeFinish, recordError — allows direct instrumentation of existing abstractions. Setting LLM_TRACE_ENABLED=1 activates the tool; omitting it leaves production and CI builds unaffected without any code changes.
The production observability platforms — Langfuse, LangSmith, Braintrust, Arize Phoenix — all assume retained data, shared team dashboards, and persistent infrastructure, which makes them the wrong tool for a solo debugging session at a laptop. Arize Phoenix is the closest architectural peer, running locally on localhost:6006, but it persists data to SQLite or PostgreSQL and is Python-centric. Loupe is currently the only npm-native entrant in ephemeral local tracing, which Harrison sees as deliberate scope: the goal is not to replace production platforms but, as he puts it, to make the gap between console.log and a full observability stack less painful to cross.