LangChain install

Install Softprobe Agent QA for LangChain / LangGraph. Prefer the copy/paste prompt from Explorer (Quick start).

Softprobe adopts your conversation and user ids (LangGraph thread_id, chat id, user_id, …). It does not invent Softprobe session UUIDs. Softprobe is additive — it never replaces your existing callbacks.

Best path (zero invoke changes)

1. Install

Python

pip install 'softprobe[langchain]'

TypeScript

pnpm add @softprobe/langchain @softprobe/tracing @langchain/core

2. Credentials

export SOFTPROBE_PUBLIC_KEY="spk_…"
export SOFTPROBE_BASE_URL="https://explorer.softprobe.ai/api/thelake"
export SOFTPROBE_OTLP_ENDPOINT="https://explorer.softprobe.ai/api/thelake/v1/traces"
export SOFTPROBE_ENVIRONMENT="Production"

Variable	Required	Meaning
SOFTPROBE_PUBLIC_KEY	Yes	Agent API key (spk_…) from Explorer Agents → + Connect agent
SOFTPROBE_BASE_URL	Yes	https://explorer.softprobe.ai/api/thelake
SOFTPROBE_OTLP_ENDPOINT	No	Defaults to {SOFTPROBE_BASE_URL}/v1/traces
SOFTPROBE_ENVIRONMENT	No	Label matching the Agent environment in Explorer
SOFTPROBE_LANGCHAIN	No	Set to 0 / false to disable auto-instrumentation
SOFTPROBE_SESSION_ID	No	Fallback only when the run has no thread/session/chat id
SOFTPROBE_USER_ID	No	Fallback only when the run has no user id

3. Enable once at process start

With credentials set, Softprobe auto-instruments when the package loads. Import it before your agent runs:

Python

import softprobe.langchain  # auto-instruments when SOFTPROBE_* credentials are set

TypeScript

import "@softprobe/langchain"; // auto-instruments when SOFTPROBE_* credentials are set

Or call explicitly (same effect, clearer in code review):

from softprobe.langchain import instrument
instrument()

import { instrument } from "@softprobe/langchain";
instrument();

Keep using your existing configurable.thread_id / user_id (or chat id). Softprobe reads them from LangChain metadata — no callbacks edits.

agent.invoke(inputs, config={"configurable": {"thread_id": chat_id, "user_id": user_id}})

Disable auto: export SOFTPROBE_LANGCHAIN=0, or uninstrument().

Escape hatch: append a handler

If you already manage callbacks and want Softprobe only on some invokes:

from softprobe.langchain import CallbackHandler

handler = CallbackHandler()
agent.invoke(
    inputs,
    config={
        "callbacks": [*existing_callbacks, handler],  # append — do not replace
        "configurable": {"thread_id": chat_id, "user_id": user_id},
    },
)
handler.flush()

Verify

1. Run one real agent turn that uses a tool.
2. In Explorer, Agents → + Connect agent → Check connection, or browse Sessions.

Sessions match the Agent via SOFTPROBE_PUBLIC_KEY. Softprobe groups Steps by your thread/session/chat id.

What is traced

Kind	Role
Agent / chain	Orchestration / graph roots
Generation	Model call with input/output and usage
Tool	Tool invocation with arguments and result
Retriever	Retrieval steps when present

Troubleshooting

Symptom	Check
No Sessions	Credentials in the same process; Softprobe imported / instrument() called before the agent runs; SOFTPROBE_LANGCHAIN not 0
Auth errors	Key and base URL match Connect Agent
Only generations	Run a turn that calls a tool
Split Sessions	Reuse the same app thread/chat id across turns

Packages: softprobe (Python), @softprobe/langchain (TypeScript).