tracelet is local DevTools for AI agents: point any OpenTelemetry exporter at localhost:4318 and watch your agent's execution tree stream in live — LLM calls, tool calls, prompts, tokens, latency, errors — in a clean local UI.

npx @jnmetacode/tracelet --demo   # two sample runs preloaded — press Compare

Or, from a clone, send them over real OTLP/HTTP:

# terminal 1 — start tracelet (UI opens at http://localhost:4321)
npx @jnmetacode/tracelet

# terminal 2 — send a synthetic multi-span agent trace
node examples/demo.js
#   ✓ Sent a demo agent trace (5 spans) to http://localhost:4318/v1/traces

# …or two runs of the same agent (before / after a fix) to try Compare
node examples/demo.js --compare

Now look at http://localhost:4321: a waterfall with an agent.run root, two LLM calls, two tool calls (one errored), prompts/responses and token counts in the inspector. With --compare, select run A, press Compare, click run B: the steps line up side by side, the fixed tool, the model swap and the prompt diff are flagged, with Δ latency / tokens / cost. examples/demo.js is ~170 lines of zero-dep OTLP JSON you can read.

tracelet is just an OTLP endpoint, so anything that emits OpenTelemetry works. Point your exporter at the ingest URL — both protobuf (the default) and JSON are accepted, so usually no config beyond:

export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318

• Vercel AI SDK v7+ — one import, zero OpenTelemetry packages: import '@jnmetacode/tracelet/ai-sdk/register'. v5/v6 via an OTel exporter. Full walkthrough: vercel-ai-sdk.md.
• LangChain.js / LangGraph.js — { callbacks: [tracelet()] } from @jnmetacode/tracelet/langchain, zero extra packages: langchain.md.
• Mastra — @mastra/otel-exporter with a custom endpoint: mastra.md.
• Python (LangChain, CrewAI, OpenAI Agents SDK, hand-rolled) — the standard OTel Python SDK works as-is, protobuf and all: python-opentelemetry.md.
• OpenInference (LangChain, LlamaIndex, CrewAI…) — any OpenInference instrumentor exporting OTLP renders with full semantic enrichment.
• Raw OpenTelemetry — plain gen_ai.* spans render too.

• Execution tree / waterfall with per-span latency bars
• Per-span inspector: model, prompt/input, completion/output, tokens, errors
• Compare two runs: aligned steps, changed-step flags, prompt/completion diffs, Δ latency / tokens / cost; pin a baseline to auto-compare every new run
• Search across prompts, completions, tool payloads, models and tool names
• Live tail over Server-Sent Events — new traces appear as they arrive
• 100% local: an in-memory ring buffer (last 500 traces); nothing is sent anywhere (opt-in --persist <file> keeps history on your own disk)

See the main README for the full reference.