tracelense 0.0.1

Local-first agent tracing — opens in your browser in 5 seconds. No Docker, no signup, no cloud.

tracelense

Local-first agent tracing. Open in your browser in 5 seconds. No Docker, no signup, no cloud.

See screenshots, comparison table, and full docs on GitHub →

pipx install tracelense && tracelense

That's it. Browser opens. Traces appear as your agent runs. SQLite at ~/.tracelense/db.sqlite. Nothing leaves your machine.

Why this exists

Every agent tracer either makes you stand up Docker (Phoenix), funnel your prompts through a cloud service (Langfuse, LangSmith, Helicone), or look at a UI that feels like 2014 enterprise software. tracelense is what you reach for when you want to see your traces now, on your laptop, the same way you reach for console.log.

Install

pipx install tracelense          # recommended
uv tool install tracelense
pip install tracelense

Python 3.10 – 3.13.

Trace your agent in one line

OpenAI

from openai import OpenAI
from tracelense import trace

client = OpenAI()
trace(client)

Anthropic

from anthropic import Anthropic
from tracelense import trace

client = Anthropic()
trace(client)

LangChain / LangGraph

from tracelense.adapters.langchain import TracelenseCallbackHandler

agent.invoke({"input": "..."}, config={"callbacks": [TracelenseCallbackHandler()]})

LlamaIndex

from tracelense.adapters.llamaindex import install
install()

Anything OpenTelemetry-instrumented

export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318
python your_agent.py

Killer features

1. A trace detail view that looks like 2026, not 2014. Span tree, prompts, tool calls, costs. One-click Copy as curl / Copy as Python.
2. Trace diff — the moat. Pick two runs. tracelense diffs the output text, tool-call sequence, request params, and cost/latency/token deltas.
3. Replay with overrides. Click Replay on any LLM span. Edit the model or prompt. tracelense re-issues against the same provider and renders the diff inline. API keys live in your OS keychain.
4. Your AI coding assistant can read your traces. A read-only MCP server (tracelense mcp) lets Claude Code / Cursor / Windsurf call find_failing_spans, get_trace, diff_traces while debugging the agent it's helping you write.
5. Cost dashboard. Today vs yesterday. Top models. Top agents. One page.

How tracelense compares

	tracelense	Phoenix	Langfuse	Helicone
Time to first trace	5 sec	~3 min	~10 min	signup
Runs without containers	yes	no (Docker)	no (Docker)	no (proxy)
Trace diff	yes	no	partial	no
Replay with overrides	yes	no	playground only	no
AI-assistant readable (MCP)	yes	no	no	no
Data leaves your machine	never	never	depends on deploy	always
Export to bigger tools	yes (OTLP)	n/a	n/a	n/a

This is an honest comparison written by an honest competitor. Phoenix and Langfuse are great when your team grows past one laptop. tracelense is the thing you reach for first. When you outgrow it, tracelense export phoenix ships your data over. No lock-in.

Wire your AI assistant (MCP)

Claude Code (~/.claude.json or project .claude/settings.json):

{
  "mcpServers": {
    "tracelense": { "command": "tracelense", "args": ["mcp"] }
  }
}

Cursor (~/.cursor/mcp.json): same shape.

Then ask your AI: "Find the latest failing trace and tell me which tool call broke it." It will. Read-only — it can't burn money behind your back.

License

MIT. See LICENSE on GitHub.