Quickstart

1. Start the dashboard

pip install 'lanetrace[langchain,server]'
lanetrace demo    # optional: three sample projects with their runs
lanetrace ui

lanetrace ui prints where everything is, then serves the dashboard:

    __                   ______
   / /   ____ _____  ___/_  __/________ _________
  / /   / __ `/ __ \/ _ \/ / / ___/ __ `/ ___/ _ \
 / /___/ /_/ / / / /  __/ / / /  / /_/ / /__/  __/
/_____/\__,_/_/ /_/\___/_/ /_/   \__,_/\___/\___/

  LaneTrace 0.3.0
  observability for AI agents

  Dashboard   http://127.0.0.1:3000
  Auth        none (no login, no API key)
  Database    sqlite:///./lanetrace.db  (/home/you/my-agent/lanetrace.db)
  Tracking    LANETRACE_TRACKING_URI=http://127.0.0.1:3000
              set it where your agents run to send their traces here
  Docs        https://coulibaly-b.github.io/lanetrace/

Open the dashboard URL. On localhost there is no login: everything belongs to one local user. The Database line shows the file the dashboard reads. The SDK writes to ./lanetrace.db too, relative to the directory your agent runs in: run the agent from the same directory, or send its runs over HTTP (step 4).

2. Trace a LangChain or deepagents agent

import lanetrace
import lanetrace.langchain

lanetrace.set_project("support-bot")   # like mlflow.set_experiment()
lanetrace.langchain.autolog()   # like mlflow.langchain.autolog()

agent.invoke({"messages": [{"role": "user", "content": "Where's my order?"}]})

Each top-level invocation becomes one run of the support-bot project. Open it in the dashboard while the agent runs and watch the arrows arrive. The callback handlers page covers what is recorded, and the middleware is the explicit alternative.

3. Trace anything else

import lanetrace

lanetrace.set_project("support-bot")


@lanetrace.trace   # a tool_call, then its tool_result (or an error)
def web_search(query: str) -> str:
    ...


with lanetrace.start_run("triage"):
    lanetrace.log_event(
        source="User", target="Orchestrator", type="handoff",
        label="Where is my order?",
    )
    lanetrace.log_event(
        source="Orchestrator", target="LLM", type="llm_call", duration_ms=480
    )
    web_search("order 4521")
    lanetrace.log_event(
        source="Orchestrator", target="User", type="final_answer",
        label="It shipped yesterday.",
    )

Outside a start_run() block, log_event() and @lanetrace.trace use one implicit run per process, ended by lanetrace.end_run() or when the process exits. Tracing never raises into your code, except a ValueError for an unknown event type: events are written from a background thread, and a storage error turns tracing off for that run, with one logged warning.

4. Send runs from elsewhere

When the agent runs on another machine, in a container or from another directory, send its runs to the dashboard over HTTP:

# the banner's Tracking line; the key only for a server with --auth password
export LANETRACE_TRACKING_URI=http://127.0.0.1:3000
export LANETRACE_KEY=atr_...

lanetrace ui uses password mode by default when it listens on an address other than localhost. Sign up in the browser; on the Integration page, Regenerate key creates a project key and shows it once.

Notebooks

Two notebooks run all of the above offline, with scripted models and a temporary SQLite file: python-server/examples/notebooks/.