---

LangMonitor watches and controls your LangGraph agents from the outside. It does three things:

1. Streams every node start/end, LLM call, and state diff in real time over WebSocket.
2. Controls any live run — kill, pause, resume, inject state, swap prompts — over a REST API.
3. Checkpoints each run on top of LangGraph's native checkpointer, so you can roll back to any point.

Wrap your compiled graph in one line. LangMonitor spins up an interactive dashboard on a port you choose — open it in a browser and operate your agent live. (The raw REST API and Swagger docs stay available at /api/v1 and /docs.)

Install

pip install langmonitor

Python 3.10+.

Quick Start

Wrap your compiled graph and pick a port. That's it:

from langmonitor import monitor

monitored = monitor(compiled_graph, port=8000, open_browser=True)
result = monitored.invoke({"input": "hello"})

monitor() returns a drop-in stand-in for your graph — same invoke / ainvoke. The moment you call it, the interactive dashboard goes live at http://localhost:8000. From it you can watch every node, then kill, pause, resume, inject state, roll back, or A/B-swap the running agent. Prefer raw REST? The Swagger UI is at /docs.

No separate server to run. The dashboard lives inside your process.

Full example

A complete, copy-paste script — builds a tiny graph, monitors it, and opens the dashboard:

from typing import TypedDict

from langgraph.graph import StateGraph, END
from langmonitor import monitor


class State(TypedDict):
    n: int


def inc(state: State) -> State:
    return {"n": state["n"] + 1}


def double(state: State) -> State:
    return {"n": state["n"] * 2}


# 1. Build a normal LangGraph graph
graph = StateGraph(State)
graph.add_node("inc", inc)
graph.add_node("double", double)
graph.set_entry_point("inc")
graph.add_edge("inc", "double")
graph.add_edge("double", END)
compiled = graph.compile()

# 2. Wrap it — launches the dashboard at http://localhost:8000
monitored = monitor(compiled, port=8000, open_browser=True)

# 3. Run it like any compiled graph
result = monitored.invoke({"n": 3})
print("result:", result)                 # {'n': 8}  → (3 + 1) * 2
print("dashboard:", monitored.dashboard_url)

# 4. Keep the process alive so you can explore the run in the dashboard.
#    (The dashboard lives in this process, so it stops when the script exits.)
input("Press Enter to quit…")

Run it:

pip install langmonitor langgraph
python example.py

Then open http://localhost:8000 to inspect the trace, replay checkpoints, add guardrails, or kill/pause the next run.

How it works

monitor() has three modes — pick one by what you pass:

# 1. Embedded (default) — launches a dashboard in this process on the given port
monitor(graph, port=8000)

# 2. Remote — connect to a LangMonitor server running elsewhere
monitor(graph, server_url="ws://monitor.internal:8000", api_key="…")

# 3. In-process engine — route events straight to a MainEngine (used in tests)
monitor(graph, in_process_engine=engine)

If the dashboard can't start (or a remote server is down), monitoring fails open — your agent keeps running, just unmonitored. Monitoring never breaks the thing it's monitoring.

Usage

Wrap a graph

from langgraph.graph import StateGraph, END
from langmonitor import monitor

graph = StateGraph(MyState)
# ... add_node / add_edge / set_entry_point ...
compiled = graph.compile()

monitored = monitor(compiled, port=8000)

# Sync
result = monitored.invoke({"input": "hello"})

# Or async
result = await monitored.ainvoke({"input": "hello"})

print("Dashboard:", monitored.dashboard_url)   # http://127.0.0.1:8000

Each run gets a run_id (emitted as the run_started event and listed at GET /api/v1/runs). Use it for every control below.

Operate from the dashboard (or Swagger)

Open http://localhost:8000 for the interactive dashboard, or http://localhost:8000/docs to call the endpoints directly from Swagger — the fastest way to drive a run by hand. Everything there is also a plain REST call you can script.

From Python:

import httpx

base = "http://localhost:8000/api/v1"

httpx.post(f"{base}/runs/{run_id}/pause")
httpx.post(f"{base}/runs/{run_id}/resume", json={"state_patch": {"context": "updated"}})
httpx.post(f"{base}/runs/{run_id}/kill")

Or from the shell:

curl -X POST localhost:8000/api/v1/runs/<run_id>/pause
curl -X POST localhost:8000/api/v1/runs/<run_id>/kill

Kill and pause take effect before the next node — the wrapper checks for them between steps.

Roll back to a checkpoint

# Save a named checkpoint
curl -X POST localhost:8000/api/v1/runs/<run_id>/checkpoints \
  -d '{"label": "before-tool-call"}' -H 'content-type: application/json'

# Restore it — auto-pauses the run so you can inspect before resuming
curl -X POST localhost:8000/api/v1/runs/<run_id>/checkpoints/<checkpoint_id>/rollback

With CHECKPOINT_AUTO_SAVE=true (the default) a checkpoint is also taken after every node end.

Add guardrails

Guardrails run after every node end. When one trips, it fires the configured action (kill, pause, or alert).

curl -X POST localhost:8000/api/v1/guardrails -H 'content-type: application/json' -d '{
  "name": "cost cap",
  "rule_type": "max_cost_usd",
  "config": { "threshold": 2.0 },
  "action": "kill"
}'

Built-in rule types: max_tool_calls, max_node_repeats, max_latency_ms, max_cost_usd, and custom_condition.

A custom_condition evaluates a small, sandboxed boolean expression against the current node — no eval, no attribute access or calls:

{
  "name": "slow planner",
  "rule_type": "custom_condition",
  "config": { "expression": "node_name == 'planner' and latency_ms > 5000" },
  "action": "alert"
}

Available names: node_name, latency_ms, tokens_used, sequence_order.

A/B test a node prompt

# Create the test
curl -X POST localhost:8000/api/v1/ab-tests -H 'content-type: application/json' -d '{
  "node_name": "planner",
  "prompt_a": "You are a careful planner.",
  "prompt_b": "You are an aggressive planner."
}'

# Swap the active variant mid-run
curl -X POST localhost:8000/api/v1/ab-tests/<id>/swap

The wrapper picks up the active variant automatically before each node — no code changes needed.

Stream events

Consume the live event stream from any client. In Python:

import asyncio, json, websockets

async def watch():
    # Add ?api_key=YOUR_KEY when an API_KEY is set
    async with websockets.connect("ws://localhost:8000/ws/all") as ws:
        async for raw in ws:
            event = json.loads(raw)
            print(event["type"], event["payload"])

asyncio.run(watch())

Two channels are available:

WS  /ws/runs/{run_id}   — events for one run
WS  /ws/all             — every event across all runs

Every message has the same shape:

{
  "type": "node_end",
  "run_id": "<uuid>",
  "timestamp": "<iso8601>",
  "payload": { "node_name": "planner", "latency_ms": 312, "tokens": 148 }
}

Event	Key payload fields
run_started	graph_name, input
node_start	node_name, sequence, input_state
node_end	node_name, sequence, output_state, latency_ms, tokens
llm_call	node_name, prompt, response, model, tokens, latency_ms
state_updated	sequence, state, diff
guardrail_alert	rule_name, rule_type, action
agent_paused	reason, node_name
agent_killed	reason
checkpoint_saved	checkpoint_id, label, sequence
run_ended	status, total_tokens, total_cost_usd, duration_ms

Run a shared server (teams)

Embedded is perfect for one developer and one process. To watch agents running across many processes or machines, run one standalone server and point the SDK at it:

langmonitor                 # or: python -m langmonitor.main  → http://0.0.0.0:8000

monitor(graph, server_url="ws://monitor.internal:8000", api_key="YOUR_KEY")

Configuration

The embedded dashboard picks up the same settings as the standalone server. Set them in the environment or a .env file (see .env.example).

Variable	Default	Description
DATABASE_URL	sqlite+aiosqlite:///./langmonitor.db	SQLAlchemy async URL. Use postgresql+asyncpg://... for Postgres.
SERVER_HOST	0.0.0.0	Bind host for the standalone server.
SERVER_PORT	8000	Bind port for the standalone server.
LOG_LEVEL	INFO	Python log level.
API_KEY	""	Shared secret required on every REST/WS request. Empty = unauthenticated (dev only).
CORS_ALLOW_CREDENTIALS	false	Send Access-Control-Allow-Credentials. Force-disabled with a * origin.
ENABLE_DOCS	true	Expose /docs, /redoc, /openapi.json. Set false in production.
CHECKPOINT_AUTO_SAVE	true	Auto-save a checkpoint after every node end.
GUARDRAIL_EVAL_ENABLED	true	Set false to bypass all guardrail evaluation.
MAX_WS_CONNECTIONS_PER_RUN	50	Cap on WebSocket connections per run channel.
MAX_WS_CONNECTIONS_GLOBAL	200	Cap on connections to /ws/all.
MAX_ACTIVE_GUARDRAIL_RULES	500	Cap on active rules (each is evaluated on every node end).
MAX_REQUEST_BYTES	1048576	Max REST request body size.
MAX_STATE_PATCH_BYTES / MAX_STATE_PATCH_DEPTH	262144 / 32	Bounds on injected state patches.
MAX_AB_PROMPT_CHARS	20000	Max length of an A/B prompt variant.
CORS_ORIGINS	["http://localhost:3000"]	JSON list, comma-separated string, or single origin.
LANGGRAPH_CHECKPOINT_DB	./langgraph_checkpoints.db	LangGraph SqliteSaver path.

Security

LangMonitor is a control plane — anyone who can reach it can kill, pause, or inject state into your agents. The embedded dashboard binds to 127.0.0.1 by default, so it's local-only. Before exposing it on a network:

• Set API_KEY. Once set, every /api/v1/* route and WebSocket requires it (X-API-Key header for REST and the SDK; ?api_key= query for browser clients). Pass the same value to the SDK via monitor(graph, api_key="...") or the LANGMONITOR_API_KEY env var. When empty the server runs open and warns at startup — fine for local dev only.
• Lock down CORS to origins you trust; the * + credentials combination is force-disabled.
• Disable docs with ENABLE_DOCS=false in production.
• Connection caps, rule-count limits, and payload-size bounds (see the table) blunt trivial DoS vectors, and custom_condition guardrails are AST-sandboxed.

Tests

pip install -e ".[dev]"
pytest

Covers every sub-engine (trace, state, guardrails, checkpoints, control), the main engine event routing, every REST endpoint, the WebSocket broadcast, the SDK modes, the security hardening, and an end-to-end run with a real LangGraph StateGraph.

---

License

MIT. Simple and permissive — no surprises.