An npm-installable wrapper for agent-trace, the deterministic record/replay tool for LLM agents. This package does not reimplement agent-trace. It puts the real agent-trace command on your PATH through npm install/npx, then forwards every argument straight through to the actual Python CLI.
agent-trace records every HTTP request and response your agent makes, verbatim, to a local SQLite fixture, then replays those exact bytes later with zero network calls. It patches httpx.Client, httpx.AsyncClient, and requests.Session at the transport layer, so it works with any Python HTTP client, including the clients used internally by the OpenAI Python SDK and the Anthropic SDK. The project's own benchmark suite (benchmarks/test_overhead.py, benchmarks/test_replay_vs_live.py, benchmarks/test_fidelity.py in the main repo) reports 0.011% recording overhead, 0.93ms mean replay latency, and byte-for-byte replay fidelity, at $0 API cost per replay.
agent-trace is a Python tool. If your team, CI pipeline, or agent harness already reaches for npx to run CLIs, this wrapper skips the "how do I invoke a Python tool from a Node-first workflow" step. It is a thin exec-and-forward shim, nothing more: the actual record/replay engine, interceptors, and CLI logic all live in the agent_trace Python package.
- Python 3.10 or newer
- The
agent-observability-trace-cliPyPI package installed on the same machine (its console command isagent-trace):
pip install agent-observability-trace-cli
# or
uv add agent-observability-trace-cli
# or, for an isolated global install
pipx install agent-observability-trace-cliWithout that package installed, this wrapper prints install instructions and exits non-zero. It never installs anything on your behalf.
npm install -g agent-observability-trace-cli
# or run without installing
npx agent-observability-trace-cli versionOnce installed, agent-trace on your PATH is this wrapper, and every subcommand is forwarded unchanged to the real CLI:
agent-trace version
agent-trace list
agent-trace show run_abc123def456
agent-trace show run_abc123def456 --errors-only
agent-trace replay run_abc123def456
agent-trace inspect run_abc123def456
agent-trace diff run_a run_b
agent-trace run -- langgraph devA recording/replay round trip in Python, using the underlying library directly:
from agent_trace import tracer
import httpx
@tracer.instrument(record=True)
def fetch_data(query: str) -> dict:
with tracer.span("http-call") as span:
resp = httpx.get("https://httpbin.org/get", params={"q": query})
span.set_attribute("http.status_code", resp.status_code)
return resp.json()
result = fetch_data("hello")
# Trace and fixture saved to ~/.agent-trace/runs/run_<id>/Full command reference — every subcommand's flags, defaults, and --json support — is in the main repository's CLI reference. The record/replay model and framework integrations (LangGraph, OpenAI Agents SDK, CrewAI, and more) are also documented in the main repository.
The agent-trace bin script this package installs does two things, in order:
- Looks for a real
agent-traceexecutable on your PATH (the console scriptpip/uv/pipxinstalls) and execs it with your arguments. - If that isn't found, falls back to invoking the
agent_tracePython module directly throughpython3/python.
If neither is available, it prints the install instructions above and exits with a non-zero status.
Most observability tools for LLM agents, including LangSmith, Langfuse, Helicone, and OpenLLMetry, are observe-only: they show you a trace of what happened, but reproducing a failure still means re-running the full agent against live APIs. agent-trace additionally lets you reproduce that exact run offline, deterministically, without touching the live API.
The closest built-in comparison is LangSmith's LANGSMITH_TEST_CACHE (VCR-style cassettes via langsmith[vcr]). It's Python plus LangChain only, captures HTTP calls to api.openai.com specifically rather than any HTTP client, doesn't record full wire-level bytes, and requires a LangSmith account. agent-trace works with any Python HTTP client, records full request and response bytes locally, and needs no account or hosted service.
The full capability table against LangSmith, Langfuse, Helicone, and OpenLLMetry lives in the main repository's README.
This wrapper only forwards arguments; the limitations below belong to agent-trace itself (documented in full in the main repo's README):
- Recording and replay happen inside the Python process you import
agent_traceinto. It cannot observe or replay calls made by a third-party hosted service you don't run yourself, only your own process's outbound calls. - gRPC coverage is partial: unary-unary and sync unary-stream calls (used by Gemini/Vertex AI) are captured and replayed, but client-streaming, bidirectional-streaming, and any
grpc.aiostreaming call are not. - Capture starts once a fully-constructed HTTP request object reaches the interceptor. Exceptions raised earlier, while an SDK is still serializing a tool schema or building headers, produce zero fixture rows unless a wired-in framework integration's own error callback catches them first.
Does this package reimplement agent-trace in JavaScript?
No. It is a thin wrapper. The bin/agent-trace.js script it installs execs the real agent-trace console script if it finds one on your PATH, or falls back to invoking the agent_trace Python module through python3/python. All record/replay logic, HTTP interceptors, and CLI commands live in the Python package.
Why would I install a Node wrapper for a Python tool?
If your team already standardizes on npx/npm to run CLIs, whether in local scripts, CI steps, or agent tooling, this lets agent-trace slot into that same invocation pattern without a separate "activate a Python environment first" step. You still need the Python package installed; this wrapper does not remove that dependency.
What do I need installed before this works?
Python 3.10 or newer, plus the agent-observability-trace-cli PyPI package (pip install agent-observability-trace-cli, uv add agent-observability-trace-cli, or pipx install agent-observability-trace-cli). This npm package alone does nothing useful without it.
What happens if I run agent-trace through this wrapper without the Python package installed?
The wrapper tries the real console script first, then falls back to a direct Python module import. If both fail, it prints the exact pip/uv/pipx install commands above and exits with a non-zero status. It never installs anything automatically.
Which commands does this wrapper support?
All of them, unmodified. Every argument you pass to agent-trace through this wrapper is forwarded verbatim to the real CLI, so agent-trace version, agent-trace list, agent-trace show <run-id>, agent-trace replay <run-id>, agent-trace inspect <run-id>, agent-trace diff <run-a> <run-b>, and agent-trace run -- <command> all work exactly as documented in the main repository.
How is this different from LangSmith's tracing/caching?
LangSmith's LANGSMITH_TEST_CACHE needs LangChain and a LangSmith account, and only captures calls to api.openai.com. agent-trace works with any Python HTTP client (httpx, requests, and the transports used inside the OpenAI and Anthropic SDKs), records full request/response bytes locally, and needs no account. See the comparison table in the main repository for the full breakdown against LangSmith, Langfuse, Helicone, and OpenLLMetry.
Does this wrapper work on Windows?
The wrapper itself is plain Node.js and has no OS-specific code. The underlying agent-trace Python package's automated CI passes on Ubuntu, macOS, and Windows, documented in the main repository.
Is it safe to commit recorded fixtures to version control?
Not by default, and this is a property of agent-trace itself, not this wrapper. Fixture files can contain full HTTP request and response bodies, including API keys and prompt contents. See the main repository's README for guidance on .gitignore entries and redacting secrets before committing a fixture.
Is this free to use commercially?
Yes. Both this npm package and the underlying agent-observability-trace-cli PyPI package are Apache 2.0 licensed, which permits commercial use, modification, and redistribution, subject to the license's attribution and notice terms.
Apache-2.0