FutureSearch
FutureSearch is an AI forecasting platform that answers questions about the future with a probability, number, date or category, plus its reasoning. It also runs web research agents across spreadsheets, and publishes a third-party-scored accuracy record.
What is FutureSearch?
FutureSearch is an AI forecasting platform published by Varuna AI, Inc., a United States company founded in August 2023 by Dan Schwarz and Lawrence Phillips, with a founding team drawn from Metaculus and Google. Ask it a question about the future and it returns a probability, a number, a date or a category, together with the reasoning that produced it. The homepage frames six question shapes: probability, numeric, date, categorical, conditional and decision. Under the hood the product exposes eight documented operations. forecast handles the prediction itself, in binary, numeric, date, categorical or thresholded form, and any mode can be made conditional on a stated scenario. decision forecasts the outcome under each alternative of a choice you control, researching all options jointly so the differences are causal rather than correlational. multi_agent puts three, four or six web research agents on one question, each taking a different angle, then synthesises their findings. agent_map runs one agent per row of a table, in parallel, to populate new columns. rank, classify, merge and dedupe apply the same agents to scoring, classification, semantic joins and duplicate detection. Since July 2026 every forecast is reconciled against a shared world model assembled from thousands of past forecasts, a pass the company measured at 0.0003 to 0.0031 Brier improvement across nine base agents. What sets the tool apart is that it argues from a public scoreboard rather than from claims: the homepage shows a ranking of #3 out of 268 in the Metaculus FutureEval tournament, #18 out of 427 on ForecastBench, and a score of 0.122 on its own 1,907-question BTF-3 benchmark against 0.130 for the best frontier model in July 2026. Real positions on Kalshi, Polymarket and the S&P 500 are published, losses included. Access is through a web app, a Python SDK, an API and an MCP server, with the platform open to the public since early 2026 and more than 3,500 users claimed.
What it does
- Forecast an event as a probability, a numeric distribution, a date, a category or a threshold
- Forecast conditionally on a stated scenario, or compare the outcome of each option in a decision you control
- Put a team of web research agents on a single question and get one structured, synthesised answer
- Run a research agent on every row of a table to add columns you do not have yet
- Score and rank rows against a criterion written in plain language
- Classify, filter, semantically merge or deduplicate rows using live web research
- Plug forecasting into an AI assistant through an MCP server, or into Python through the SDK
When to use FutureSearch / When not to
A quick filter to help you decide if FutureSearch is the right fit.
When to use FutureSearch
- Analysts and researchers who need a numerical probability on a future event, together with the reasoning that produced it
- Investment, market and risk professionals modelling revenues, valuations, IPO timing or commodity scenarios
- Founders and product leaders weighing a decision they control, where each option needs its own forecast outcome
- Data and operations teams enriching, scoring, classifying, merging or deduplicating tables through live web research
- Developers and AI power users who want forecasting inside a Python pipeline or plugged into an assistant through MCP
When not to use FutureSearch
- Anyone seeking investment, financial, legal or tax advice: the terms present the service as educational and informational only
- Users who need a mobile app or a browser extension, since the tool ships only as a web app, an API and plugins
- Teams that require a non-English interface, as no supported language is declared anywhere on the site
- Organisations bound to EU-only data residency, a signed DPA or a formal certification, none of which is offered
- People under 18, who are excluded by both the privacy policy and the terms of use
How to use FutureSearch
A typical end-to-end flow, from setup to results.
- Open an account at futuresearch.ai/app; the free tier comes with $20 of credit and needs no credit card
- Type a question about the future and pick the shape of the answer you want: probability, number, date, category, threshold or conditional scenario
- Set the effort level, which drives how many research agents work on the question and therefore what it costs
- Read the result with its reasoning; while it runs, a line tells you how many of your past forecasts and public forecasts it is drawing on
- Check the closing note explaining whether the shared world model upheld or adjusted the raw forecast
- Publish a forecast with the globe icon if you want a permanent public page and want it to feed the shared world model; otherwise it stays private
- To use it from an assistant, add the connector https://mcp.futuresearch.ai/mcp in Claude.ai, then enable code execution, file creation and network egress
- For Claude Code, run claude mcp add futuresearch --scope project --transport http https://mcp.futuresearch.ai/mcp and authenticate; no API key is needed
- For Python, install with pip install futuresearch on Python 3.12 or later, get a key from the app and export FUTURESEARCH_API_KEY
- Wrap operations in asyncio, group related calls in a session, and use the _async variants for long-running jobs
Pros & Cons
Pros
- An accuracy record that is public, dated and scored by third parties rather than self-reported
- Real-money trading positions published in full, losing ones included
- Very low unit prices, from a few cents per table row to a handful of dollars for a high-effort forecast
- $20 of free credit at sign-up with no credit card required
- Usable without writing code through the web app, and from an existing AI assistant with a single command
- Forecasts private by default, with contribution to the shared world model left as a deliberate choice
- Team and research pedigree from Metaculus, Google and published academic work, including open benchmark datasets
Cons
- No postal address and no legal notice page anywhere: only the corporate name appears, in the terms and privacy policy
- No Data Processing Agreement, no formal subprocessor list and no security certification such as SOC 2 or ISO 27001
- No Article 27 GDPR representative and no named data protection officer
- Data hosted and processed in the United States only, with no European hosting option
- Anonymised data derived from personal information may train the publisher's models, with no documented opt-out
- No pro-rata refund on cancellation, and the terms describe quarterly renewal while the pricing page shows monthly amounts
- No mobile app, no browser extension, no declared interface language other than English, and demo videos still marked as coming soon
Pricing & Plans
A permanent free plan is available at $0 and includes $20 of credit, all operations and API access with no credit card required. The lowest paid entry point is the Analyst Starter plan at $20 per month (USD), which grants a 15% discount on usage rates. Paid plans do not buy a quota: usage is billed on the tokens actually consumed, so the final cost varies with the effort level chosen.
- Free — $0 — $20 free credit
- 10 researchers at a time
- all operations included
- API access
- no credit card required
- Analyst Starter — $20 per month — 15% off usage rates
- 20 researchers at a time
- all operations included
- API access
- priority support
- Research — $99 per month — 30% off usage rates
- 40 researchers at a time
- all operations included
- API access
- priority support (flagged as most popular)
- Expert — $500 per month — 30% off usage rates
- 80 researchers at a time
- all operations included
- API access
- priority support
- Top-up credit packs — any amount
- valid for one year and not reset monthly like a plan allowance
- Custom enterprise pricing and volume discounts on request via hello@futuresearch.ai
Data, GDPR & hosting
A consolidated view of how FutureSearch handles your data.
GDPR overview
The privacy policy, effective 29 August 2025, contains a dedicated section for users in the European Economic Area, the United Kingdom and Switzerland. It names Varuna AI, Inc. as the data controller, gives privacy@futuresearch.ai as the contact point, and lists four legal bases: contract performance, legitimate interests, consent and legal obligations. The rights of access, rectification, erasure, restriction, portability, objection, withdrawal of consent and complaint to a supervisory authority are all set out, with a stated 30-day response time. Transfers to the United States rely on Standard Contractual Clauses approved by the European Commission. Three things are missing: no Article 27 representative is designated, no data protection officer is named, and no Data Processing Agreement or security certification is published anywhere on the site.
Who owns the data?
The publisher, Varuna AI, Inc. trading as FutureSearch, is named as the data controller and, under the terms of use, owns all rights in the Service itself, including its software, algorithms and forecasting models. The terms grant the user no explicit licence over the outputs and contain no assignment clause, so ownership of generated forecasts is simply not addressed. Anything submitted as feedback becomes a perpetual, irrevocable, worldwide and sublicensable licence for the company. Spreadsheet data sent through the integrations is limited to the cells the user selects and is not retained once processing ends. Forecasts stay private by default and only become public if the user publishes them deliberately.
Reuse rights
Nothing in the terms restricts what a user may do with the forecasts and research results they obtain: there is no repurposing clause, no attribution requirement and no permission step before reusing an output commercially. The counterweight sits in the disclaimers rather than in a licence, since the company insists the results are educational and informational, may be imperfect or wrong, and must not be relied upon for investment, financial, legal or tax decisions. What the user may not reuse is the Service itself, which cannot be resold, sublicensed, framed, reverse-engineered or commercially exploited. On the publisher's side, personal data may be aggregated, de-identified or anonymised and then used to train its own AI models and shared with third parties, with no documented way to opt out. Prompt content may also be accessed by the AI providers behind the service, named as Anthropic, OpenAI, Google and X.
Data retention & training
Hosting summary
The company states that it is headquartered in the United States and that personal data may be transferred to and processed there, where its servers and operations are located. It also warns that it may use service providers operating in other countries, without naming those countries. For users in the European Economic Area, the United Kingdom and Switzerland, transfers rely on Standard Contractual Clauses approved by the European Commission and, where applicable, other valid transfer mechanisms. No European hosting region is offered and no choice of jurisdiction is available. The named third parties with access to data are the AI providers Anthropic, OpenAI, Google and X, plus Stripe for payments; beyond those names, no formal subprocessor list is published and no hosting certification is claimed. Technically, the domain resolves to an address geolocated in Council Bluffs, United States, on Google infrastructure, with Cloudflare as the registrar.
Things to keep in mind
Risks and trade-offs to weigh before adopting FutureSearch.
- A confident probability invites over-trust: the terms warn the outputs are educational, may be imperfect or wrong, and must not be relied upon for investment, financial, legal or tax decisions
- Outsourcing judgement to a forecaster can erode the habit of reasoning through uncertainty yourself, especially when the number arrives already justified
- Publishing a forecast is a one-way door in practice: it creates a permanent public page and feeds the shared world model
- Anonymised data derived from your personal information may train the publisher's models, and no opt-out is documented
- Costs are billed on tokens actually consumed, so a curious afternoon of high-effort questions can spend far more than the headline price suggests
- Cancelling a subscription refunds nothing pro rata, and the billing period stated in the terms does not match the one shown on the pricing page
- The publisher discloses no postal address, no DPA and no certification, which limits what a cautious buyer can verify before committing sensitive work
Setup & Integrations
Technical difficulty
Very low for most users, moderate for developers. The web app needs only an account and no credit card. Connecting an AI assistant means pasting one connector URL in Claude.ai, plus enabling code execution, file creation and network egress, or running a single command in Claude Code with no API key. The Python route is the only demanding one: Python 3.12 or later, an API key in an environment variable, and operations that are coroutines and must run inside asyncio or a notebook. Nothing has to be hosted or configured on a server.
Deployment
Integrations
Behind FutureSearch
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
What does FutureSearch actually do?
How much does it cost?
Do I need to know how to code?
How do I connect it to Claude?
What proof of accuracy is there?
Are my forecasts public?
Is my data used to train AI models?
Where is the data hosted?
Is there a mobile app?
Who publishes the tool and how do I get help?
Should you pick FutureSearch?
FutureSearch occupies a narrow, well-defended niche: quantified forecasting and web research at the scale of a spreadsheet. Its real distinguishing feature is not the model behind it but the willingness to be scored in public. Rankings come from Metaculus and ForecastBench rather than from the company, benchmarks and datasets are published openly, and the real-money positions on Kalshi, Polymarket and the S&P 500 are shown with their losers as well as their winners. For a category where accuracy claims are usually unfalsifiable, that is a meaningful stance. Getting started is unusually cheap and unusually easy. Twenty dollars of credit without a card, a web app for people who do not write code, a single command for anyone using Claude Code, and a Python SDK for those who want forecasting inside a pipeline. Unit costs are measured in cents per table row and dollars per forecast, which makes experimentation genuinely affordable. The reservations are on the governance side rather than the product side. No postal address or legal notice is published, no Data Processing Agreement or security certification exists, no Article 27 representative is designated, and hosting is United States only. The privacy policy also permits anonymised data derived from personal information to train the publisher's own models, with no documented opt-out, and the terms describe quarterly renewal where the pricing page shows monthly amounts. None of this is unusual for a young company, but a European organisation with compliance obligations will notice it. The product opened to the public in early 2026 and still carries the marks of that youth, including demo videos announced but not yet shipped. Treat it as a serious forecasting instrument and an immature vendor relationship at the same time.
- Choosing a selection results in a full page refresh.
- Opens in a new window.