FutureSearch - AI Forecasting with a Verifiable Track Record

TL;DR

TL;DR: FutureSearch is an AI forecasting platform with a public track record — ranked #1 of 226 on Metaculus FutureEval and above the superforecaster median on ForecastBench — accessible via MCP server, Python SDK, or web app.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: futuresearch.ai ← verified Aug 2026
• Source repository: github.com/futuresearch ← verified Aug 2026
• License: not confirmed (no LICENSE file in main repos as of Aug 2026)
• HN launch thread: news.ycombinator.com/item?id=42979867
• Pricing page: futuresearch.ai/pricing — free tier available; $20 free credits on signup
• MCP server endpoint: https://mcp.futuresearch.ai/mcp

What Is FutureSearch?

FutureSearch is an AI forecasting platform that answers questions about the future — probabilities, dates, and numbers — with a publicly verifiable track record. Rather than a black-box prediction API, it publishes its accuracy on real unresolved questions and benchmarks itself against human superforecasters.

From the site’s own description:

“AI forecasting with a public track record. First in Metaculus’s Summer 2026 FutureEval tournament, above the superforecaster median on ForecastBench, and every forecast draws on a shared world model that compounds with use.”

Public track record highlights (as of Aug 2026):

• #1 of 226 on Metaculus FutureEval live tournament
• #20 of 335 on ForecastBench (contaminated-free benchmark by Forecasting Research Institute)
• Live positions on Kalshi, Polymarket, and S&P 500 — full research published including losing trades
• 0.122 on BTF-3 benchmark (1,907-question pastcasting benchmark; best frontier model scores 0.130)

Integration options:

• Claude.ai / Claude Desktop (MCP)
• Claude Code (MCP or plugin)
• Python SDK (pip install futuresearch)
• Codex with Skills or MCP
• Gemini with Skills or MCP
• Cursor with MCP
• Web app at app.futuresearch.ai

Setup with Claude Code (MCP)

The fastest way to use FutureSearch locally is via the Claude Code MCP integration.

Prerequisites

• Claude Code installed
• No API key required for MCP transport

Step 1: Add the MCP Server

claude mcp add futuresearch --scope project --transport http https://mcp.futuresearch.ai/mcp

Step 2: Authenticate

Launch Claude Code and run:

/mcp

Select FutureSearch → Authenticate → authenticate with Google. No API key is needed.

Step 3: Ask a Forecast

Once authenticated, try asking:

How many tech layoffs will occur by the end of 2026?

FutureSearch will return a probability distribution with supporting research.

Setup with Python SDK

Installation

pip install futuresearch

or with uv:

uv pip install futuresearch

Basic Usage

from futuresearch.ops import forecast

result = forecast(input="How many tech layoffs will occur by the end of 2026?")
print(result)

For batch forecasting across a dataset:

import pandas as pd
from futuresearch.ops import forecast

questions_df = pd.read_csv("questions.csv")
result = await forecast(input=questions_df)

Deeper Analysis

What Makes FutureSearch Different?

Most AI forecasting tools are black boxes — you cannot verify how accurate they are over time. FutureSearch publishes every forecast it makes, including losing positions on real-money markets (Kalshi, Polymarket, S&P 500). This creates a public accountability record that compounds with each forecast, since the system draws on a shared world model.

The platform’s #1 ranking on Metaculus — a live tournament where 226 bots compete on unresolved real-world questions — is the strongest public signal. Metaculus tournaments are particularly rigorous because questions are resolved against actual outcomes (e.g., “Will X happen by date Y?”), not self-reported confidence.

Use Cases

• Research pipelines — batch forecast entire datasets (e.g., forecast(input=questions_df))
• Decision support — probability-weighted inputs for risk models
• Competitive intelligence — track market-implied probabilities across prediction markets
• AI agent augmentation — give coding agents access to calibrated probability estimates

Pricing

From the pricing page at futuresearch.ai/pricing:

• Free tier available
• $20 free credits on signup
• Usage-based after that

No API key is required when using the MCP transport — authentication is handled through Google OAuth at the MCP connection step.

Practical Evaluation Checklist

• [ ] Added FutureSearch MCP server to Claude Code (claude mcp add futuresearch ...)
• [ ] Successfully authenticated via /mcp in Claude Code
• [ ] Asked a real forecasting question and received a probability + research
• [ ] Explored the live tournament standings at evals.futuresearch.ai
• [ ] Reviewed published trading results at markets.futuresearch.ai
• [ ] Tested Python SDK batch forecast on a dataset

FAQ

Q: Does FutureSearch require an API key? A: No — when using the MCP transport, authentication is handled via Google OAuth in Claude Code. The Python SDK may require an API key for direct API access.

Q: How is the track record verified? A: FutureSearch publishes full research for every forecast, including losing positions on real-money prediction markets. Tournament standings on Metaculus and ForecastBench are updated live as questions resolve.

Q: What question formats does it support? A: Probabilities (“What is the chance X happens by date Y?”), dates (“When will Z occur?”), and numbers (“How many X will there be?”).

Q: Can it batch-process thousands of questions? A: Yes — the Python SDK accepts a DataFrame via forecast(input=questions_df) for bulk operations.

Q: Is there a self-hosted option? A: No — FutureSearch is a hosted service. There is no self-hosted deployment option documented as of Aug 2026.

Conclusion

FutureSearch solves the “black-box AI prediction” problem by publishing everything — every forecast, every result, every trade. Its live #1 ranking on Metaculus FutureEval is a concrete signal of forecasting ability, not a marketing claim.

For developers building AI-powered decision tools, the MCP integration is the fastest path to production — add it to Claude Code in under a minute and start asking calibrated forecasting questions. For data teams, the Python SDK handles batch processing across entire datasets.

The public track record (live trading results, tournament rankings, pastcasting benchmarks) makes FutureSearch one of the few AI forecasting platforms where the accuracy claim is independently verifiable.