Skip to content

Examples

tracesage ships a gallery of runnable examples in the examples/ directory, in three tiers — from a 30-second first taste to a 30-app before/after gallery.

Tier Folder Needs What it is
Getting started examples/getting_started/ no API key 3 standalone demos driven by FakeListChatModel — run instantly
MCP tools examples/mcp/ tracesage[mcp] (+ LLM key for the real-world demos) intro scenarios + two real-world demos (Trip Planner, Gmail+YouTube) — tools attributed by source
Showcase examples/showcase/ an LLM API key 30 real before/after apps across popular use cases

Getting started (zero setup)

pip install "tracesage[langchain]"
python examples/getting_started/01_smart_search_agent.py   # then open the URL it prints (default http://localhost:7842/ui)

01_smart_search_agent (one agent, four tools), 02_research_supervisor (multi-agent supervisor), 03_rag_with_tools (LCEL chain + retriever + tools). These use FakeListChatModel, so they run with no API key.

MCP tools

pip install "tracesage[mcp]"
python examples/mcp/main.py            # then open the URL it prints (default http://localhost:7842/ui)

Tools attributed by source in the topology and the "Tools by source" panel:

  • Intro scenarios (main.py, mcp_only.py, local_only.py, single_agent_multi_mcp.py) — two local stdio MCP servers (weather, math) plus hardcoded tools; no API key.
  • Real-world demos (need an LLM key):
    • Trip Planner (trip_demo/demo.py) — one agent across three bundled MCP servers (flights, weather, hotels; 7 tools each) plus a local tool. No external installs.
    • Gmail + YouTube (gmail_youtube_demo/) — a ReAct agent reading Gmail + YouTube transcripts, with a before.py / after.py pair showing the exact diff.

See MCP support for how attribution works.

Showcase — 30 before/after apps

pip install -r examples/showcase/requirements.txt
export OPENAI_API_KEY=...               # or LLM_PROVIDER=anthropic + ANTHROPIC_API_KEY
python examples/showcase/01_support_faq_router/before.py    # plain app
python examples/showcase/01_support_faq_router/after.py     # same app + live trace

The flagship gallery: each app ships a before.py (plain LangChain/LangGraph, real LLMs) and an after.py (the same app + tracesage), so diff before.py after.py shows exactly how little it takes to add observability.

The LLM_PROVIDER / LLM_MODEL env vars are read by the example apps (via LangChain's init_chat_model) to pick a provider — they are not tracesage settings. tracesage itself is provider-agnostic and has no provider config.

The 30 apps span five themes — see the full index in the showcase README:

  • Foundational patterns — router, ReAct agent, text-to-SQL, sequential chain, parallel fan-out
  • RAG & knowledge — docs Q&A, multi-query, agentic RAG, reranker, conversational (memory)
  • Multi-agent systems — supervisor, hierarchical, support triage, competitive intel, code migration, sales, debate
  • Tools & MCP — MCP personal assistant, GitHub triage, multi-MCP travel, DevOps responder, e-commerce concierge
  • Reasoning loops & evaluation — reflexion writer, plan-and-execute, self-correcting codegen, LLM-as-judge, map-reduce
  • Domain verticals — invoice extraction, contract clause risk, insurance claim intake

Each app folder has its own README explaining what the trace reveals.

Each example is its own application (isolated data dir)

The topology map and the "Tools by source" panel aggregate every run in a data dir, so two applications sharing one dir would merge into a single graph. To keep them separate, every example writes to its own data dir under ~/.tracesage/:

  • getting_started/ → ~/.tracesage/smart-search, …/research-supervisor, …/rag-tools
  • mcp/ → ~/.tracesage/mcp-mixed, …/mcp-only, …/local-tools, …/multi-mcp
  • showcase/<NN>_<name>/after.py → ~/.tracesage/<NN>_<name> (its folder name)

Each script prints its Data dir: and tracesage runs -d <dir> on startup. Inspect one example's runs without others bleeding in:

tracesage runs  -d ~/.tracesage/smart-search
tracesage serve -d ~/.tracesage/smart-search

Use the same pattern for your own apps — one data_dir per application. See Configuration → Isolating multiple applications.