crewai-prediction-markets

CrewAI tools for real-time prediction market data. Drop-in tools that give any CrewAI agent or crew world awareness from 30,000+ Kalshi and Polymarket contracts — no auth required.

from crewai import Agent, Crew, Task
from crewai_prediction_markets import prediction_market_tools

analyst = Agent(
    role="Market Analyst",
    goal="Identify the most actionable prediction-market edges right now",
    backstory="You read the world through prediction-market prices.",
    tools=prediction_market_tools(),
)

task = Task(
    description="What is the highest-conviction trade idea right now? Cite the ticker and the catalyst.",
    expected_output="One trade idea with ticker, conviction, current price, and catalyst.",
    agent=analyst,
)

crew = Crew(agents=[analyst], tasks=[task])
print(crew.kickoff())

Install

pip install crewai-prediction-markets

Only depends on requests. CrewAI itself is not a hard dependency — you can use the tools standalone (see the Standalone usage section below).

Tools

All six tools hit the public SimpleFunctions API. No API key, no rate limit, no auth. Every endpoint below is verified live.

| Tool | Endpoint | When to use | |------|----------|-------------| | GetContextTool | /api/public/context | Start here. Single bundle: edges, movers, highlights, traditional-market context. | | GetWorldStateTool | /api/agent/world | ~800-token compressed snapshot of all markets, ideal for system-prompt injection. | | GetWorldChangesTool | /api/agent/world/delta | ~30-50 token incremental delta — cheap polling loops. | | GetMarketEdgesTool | /api/edges | Raw mispricings (thesis price vs market price) with reasoning. | | GetUncertaintyIndexTool | /api/public/index | Single numeric pulse: uncertainty, geopolitical risk, momentum, activity. | | GetIdeasTool | /api/public/ideas | LLM-generated trade ideas with conviction, catalyst, time horizon. |

Multi-agent crew example

from crewai import Agent, Crew, Task, Process
from crewai_prediction_markets import prediction_market_tools

scout = Agent(
    role="Market Scout",
    goal="Surface the top 3 mispriced edges across Kalshi and Polymarket",
    tools=prediction_market_tools(),
)

writer = Agent(
    role="Note Writer",
    goal="Turn raw edges into a concise daily note for a portfolio manager",
    backstory="You write tight, opinionated, 200-word market notes.",
)

scout_task = Task(
    description="Find the 3 highest-conviction edges with catalysts in the next 7 days",
    expected_output="Bullet list of 3 edges with ticker, executableEdge, catalyst, time horizon",
    agent=scout,
)

write_task = Task(
    description="Write a 200-word daily note covering the 3 edges",
    expected_output="A 200-word markdown note",
    agent=writer,
    context=[scout_task],
)

crew = Crew(agents=[scout, writer], tasks=[scout_task, write_task], process=Process.sequential)
print(crew.kickoff())

Standalone (no crewai needed)

Every tool exposes both a _run (returning the parsed payload) and a run (returning a JSON string ready for an LLM tool message). Use whichever fits your loop:

from crewai_prediction_markets import GetUncertaintyIndexTool, GetIdeasTool

idx = GetUncertaintyIndexTool()._run()
print(f"Uncertainty: {idx['uncertainty']}/100")

ideas = GetIdeasTool()._run()
top = ideas["ideas"][0]
print(f"Top idea: {top['headline']} ({top['conviction']} conviction)")

Response shapes

GetContextTool → dict

{
  "edges": [...],            # top mispricings
  "movers": [...],           # 24h price movers
  "highlights": [...],       # recent narrative-shaping events
  "traditionalMarkets": {...}
}

GetUncertaintyIndexTool → dict

{
  "uncertainty": 22,          # 0-100
  "geopolitical": 0,          # 0-100
  "momentum": -0.08,          # -1 to +1
  "activity": 99,             # 0-100
  "components": {...},
  "timestamp": "2026-04-07T07:18:03.451Z",
}

GetMarketEdgesTool → dict

{
  "edges": [
    {
      "ticker": "KXFEDDECISION-25DEC",
      "venue": "kalshi",
      "title": "Will the Fed cut rates in December?",
      "marketPrice": 42,           # cents
      "thesisPrice": 31,           # cents
      "executableEdge": 11,        # cents (after spread)
      "confidence": 0.78,
      "liquidityScore": "high",
      "direction": "no",
      "reasoning": "...",
    },
    ...
  ]
}

GetIdeasTool → dict

{
  "generatedAt": "2026-04-07T01:29:41Z",
  "cached": True,
  "ideas": [
    {
      "headline": "...",
      "pitch": "...",
      "conviction": "high",
      "direction": "buy_yes",
      "markets": [{"url": "...", "ticker": "...", "currentPrice": 14.5, "venue": "polymarket"}],
      "catalyst": "...",
      "timeHorizon": "2w",
      "risk": "...",
    },
    ...
  ]
}

Errors

Tools call requests.Response.raise_for_status() on every response, so any non-2xx surfaces as a requests.HTTPError. Wrap in try/except if your crew should degrade gracefully:

import requests
from crewai_prediction_markets import GetContextTool

try:
    ctx = GetContextTool()._run()
except requests.HTTPError as e:
    ctx = {"error": str(e), "edges": [], "movers": []}

Sister packages

If you're not using CrewAI, use the wrapper for your stack:

| Stack | Package | |-------|---------| | Vercel AI SDK | vercel-ai-prediction-markets | | LangChain / LangGraph | langchain-prediction-markets | | OpenAI Agents SDK / function calling | openai-agents-prediction-markets | | MCP / Claude / Cursor | simplefunctions-cli | | Bare Python SDK | simplefunctions-python |

Testing

pip install -e .[dev]
pytest

11 tests, all requests.get-mocked — no network required.

License

MIT — built by SimpleFunctions.