Junctionary Pulse — LangChain Toolkit

Give your LangChain agents finished stock-market briefings: one tool call per ticker, every weekday morning. No raw-news plumbing, no 50k-token web-research loops — just structured briefings with sources.

• Product: junctionarypulse.com — $5/mo agent plan, any 5 tickers
• Auth: X-API-Key header (get a key at /agents)
• Rate limit: 120 requests/min

Installation

pip install junctionary-pulse-langchain

Usage

import os
from langchain.agents import create_agent
from junctionary_langchain import JunctionaryPulseToolkit

toolkit = JunctionaryPulseToolkit(
    api_key=os.environ["JUNCTIONARY_API_KEY"],
)

# Works with any LangChain agent — tools are just BaseTools
agent = create_agent(
    model="claude-haiku-4-5",
    tools=toolkit.get_tools(),
    system_prompt="You are a market research assistant. Use Junctionary Pulse "
                  "tools for stock news; always cite sources from the briefing.",
)

result = agent.invoke({
    "messages": [{"role": "user",
                  "content": "What's the latest on Nvidia, and is sentiment improving?"}]
})
print(result["messages"][-1].content)

With a lower-level chain, call tools directly:

tools = {t.name: t for t in toolkit.get_tools()}

briefing = tools["get_briefing"].invoke({"pulse_id": "nvidia"})
history  = tools["get_briefing_history"].invoke({"pulse_id": "nvidia", "days": 14})

# async works too
briefing = await tools["get_briefing"].ainvoke({"pulse_id": "nvidia"})

Tools

Reading the room: which tool when

• "What's happening today?" / morning catch-up → get_watchlist_digest (1 call, everything)
• "Latest news on TSLA" → get_briefing
• "Catch me up on Apple this month" → get_briefing_history, then get_briefing_for_date for the interesting days
• "Is NVDA sentiment improving?" → get_sentiment_trends
• "NVDA vs AMD — which looks better?" → compare_tickers
• "When does Apple report earnings?" → get_earnings_date
• "What's the ticker for ServiceNow?" → resolve_ticker
• "What happened with MSFT last Tuesday?" → get_briefing_for_date
• "Did anything move today?" → get_alerts
• "What do you cover?" → list_pulses

Token efficiency

get_watchlist_digest is the cheapest way to orient an agent: one HTTP call returns a compact row per ticker (headline, TL;DR, previous close, story count) instead of one full get_briefing call per ticker. For a 5-ticker subscription, that's roughly 1/10th the tokens of pulling five full briefings. Scan the digest first, then get_briefing only for the tickers that matter.

Error handling

Tools never raise into the agent loop. Failures come back as plain-text messages the agent can act on:

Configuration

JunctionaryPulseToolkit(
    api_key="jp_...",                              # or JUNCTIONARY_API_KEY env var
    api_base="https://api.junctionarypulse.com",   # override for staging/tests
)

Tests

python3 -m pytest test_toolkit.py -v

All HTTP is mocked — no live API calls, no API key needed.