Quickstart: LangChain / LangGraph

LangChain provides an MCP adapter (langchain-mcp-adapters) that converts MCP tools into LangChain BaseTool instances. Wire Pipeworx in once and any LangChain or LangGraph agent can use it.

Install

pip install langchain langchain-mcp-adapters

Wire it up

from langchain_mcp_adapters.client import MultiServerMCPClient
from langgraph.prebuilt import create_react_agent
from langchain_anthropic import ChatAnthropic

client = MultiServerMCPClient({
    "pipeworx": {
        "transport": "streamable_http",
        "url": "https://gateway.pipeworx.io/mcp",
    }
})

tools = await client.get_tools()
agent = create_react_agent(
    ChatAnthropic(model="claude-sonnet-4-6"),
    tools,
)

response = await agent.ainvoke({
    "messages": [{"role": "user", "content": "Get Apple's latest 10-K"}]
})

The agent now has access to all Pipeworx tools as LangChain tools, including the meta-tools (ask_pipeworx, discover_tools, resolve_entity, compare_entities).

Scope to a vertical

client = MultiServerMCPClient({
    "pipeworx-pharma": {
        "transport": "streamable_http",
        "url": "https://gateway.pipeworx.io/mcp?vertical=pharma",
    }
})

Only pharma-relevant tools load — see context tax.

Multi-server setup

client = MultiServerMCPClient({
    "pipeworx_finance":  {"transport": "streamable_http", "url": "https://gateway.pipeworx.io/mcp?vertical=fintech"},
    "pipeworx_pharma":   {"transport": "streamable_http", "url": "https://gateway.pipeworx.io/mcp?vertical=pharma"},
    "pipeworx_research": {"transport": "streamable_http", "url": "https://gateway.pipeworx.io/mcp?task=academic+papers"},
})

tools = await client.get_tools()  # all servers' tools, namespaced

Auth

{
  "transport": "streamable_http",
  "url": "https://gateway.pipeworx.io/mcp",
  "headers": {"Authorization": f"Bearer {pipeworx_token}"}
}

Reading Pipeworx response metadata

Every Pipeworx tool response embeds _meta (cost, freshness, retry hints, examples, alternatives) inside the text content. With langchain-mcp-adapters the raw response is preserved as the tool’s output:

import json

async for event in agent.astream({"messages": [...]}, stream_mode="events"):
    if event["event"] == "on_tool_end":
        content = event["data"]["output"].content
        # ToolMessage.content is either a plain string or a list of
        # {"type": "text", "text": ...} dicts (langchain_core: content items
        # in the list case are dicts, accessed with ["text"] — not `.text`).
        text = content[0]["text"] if isinstance(content, list) else content
        parsed = json.loads(text) if isinstance(text, str) and text.startswith("{") else {}
        # The wrapper carries _meta when it's a Pipeworx tool
        meta = parsed.get("_meta") or {}
        if meta.get("feedback_hint"):
            print(f"feedback nudge: {meta['feedback_hint']}")

Useful for production agents that need cost transparency, freshness gates, or auto-file feedback on errors.

Memory across runs

Pipeworx’s remember / recall / forget tools persist agent state across LangGraph runs when authenticated as the same account:

# Day 1 — pin a research target
await agent.ainvoke({"messages": [{"role": "user", "content": "Save 'focus_ticker' as AAPL."}]})

# Day 2 — pick up
await agent.ainvoke({"messages": [{"role": "user", "content": "What's our focus ticker, and pull its latest 10-K."}]})
# The agent calls recall({key: "focus_ticker"}) → "AAPL", then edgar_company_filings

See memory.

Auto-passing serverInstructions

LangChain’s tool router doesn’t natively forward MCP serverInstructions — the gateway sends them on initialize but create_react_agent only sees tool definitions. For best results, pass them into the system prompt manually:

async with client.session("pipeworx") as session:
    # `session()` auto-initializes; `instructions` lives at the TOP LEVEL of
    # the InitializeResult, not under serverInfo (serverInfo only has name/version).
    instructions = session.initialize_result.instructions

agent = create_react_agent(
    ChatAnthropic(model="claude-sonnet-4-6"),
    tools,
    prompt=f"You are a research assistant.\n\n{instructions}",  # `prompt=`, not `state_modifier=` — renamed in current langgraph.prebuilt
)

This briefs the model on ask_pipeworx, discover_tools, compound _intel tools, and the prompt playbooks — all upfront context that improves first-call accuracy.

Caveats

• langchain-mcp-adapters lazy-loads tool definitions on first call. First invocation has a small latency overhead while it fetches tools/list.
• LangGraph’s checkpointer (memory between agent steps) is separate from Pipeworx’s remember/recall (memory across user sessions). Use checkpointer for in-flight state, Pipeworx memory for cross-session.
• For long-running agents, scope to a vertical to keep tool count low — see reducing context.