Futuresearch Python

---
name: futuresearch-python
description: Use when the user wants forecasts (probabilities, dates, numbers, odds) or dataset research at scale via FutureSearch.
---

# FutureSearch Python SDK

FutureSearch turns questions about the future into forecasts, with dataset research tools around it. Use this skill when writing Python code that needs to:

> **Documentation**: For detailed guides, case studies, and API reference, see:
> - Docs site: [futuresearch.ai/docs](https://futuresearch.ai/docs)
> - GitHub: [github.com/futuresearch/futuresearch-python](https://github.com/futuresearch/futuresearch-python)

**Operations:**
- Forecast probabilities, numbers, dates, and categories for questions about the future
- Forecast the outcome under each alternative of a decision you control
- Research one question with a team of parallel agents
- Run AI agents over dataframe rows
- Rank/score rows based on qualitative criteria
- Deduplicate data using semantic understanding
- Merge tables using AI-powered matching
- Classify rows into predefined categories

## Installation

### Python SDK

```bash
pip install futuresearch
```

### MCP Server (for Claude Code, Claude Desktop, Cursor, etc.)

If an MCP server is available (`futuresearch_classify`, `futuresearch_rank`, etc. tools), you can use it directly without writing Python code. The MCP server operates on uploaded data (via artifact IDs or inline JSON).

To install the MCP server, add to your MCP config:

```json
{
  "mcpServers": {
    "futuresearch": {
      "type": "http",
      "url": "https://mcp.futuresearch.ai/mcp"
    }
  }
}
```

Config file locations:
- **Claude Code**: `~/.claude.json` (user) or `.mcp.json` (project)
- **Claude Desktop**: `~/Library/Application Support/Claude/claude_desktop_config.json` (macOS)
- **Cursor**: `~/.cursor/mcp.json`

## When to Use SDK vs MCP

**Use MCP tools** when:
- Quick one-off operations on CSV files
- User wants direct results without writing code
- Simple lookups and enrichments

**Use Python SDK** when:
- Complex multi-step workflows (dedupe → merge → research)
- Custom data transformations
- Integration with existing Python scripts
- Full control over execution and intermediate results

---

# MCP Server Tools

If you have the FutureSearch MCP server configured, these 18 tools are available. All data processing tools accept input via `artifact_id` (from upload_data or request_upload_url) or `data` (inline JSON rows). Provide exactly one.

## Core Operations

### futuresearch_forecast
Forecast questions about the future. Five outcome types: binary probabilities,
numeric percentiles, date percentiles, categorical (one probability per listed
outcome), and thresholded (one probability per listed threshold condition). Any
of them can be made conditional on a stated intervention by supplying a condition
(see "Conditional forecasting" below).
```
Parameters:
- artifact_id: Artifact ID (UUID) from upload_data or request_upload_url
- data: Inline data as a list of row objects (must include "question" column)
- forecast_type: "binary", "numeric", "date", "categorical", or "thresholded" (always the OUTCOME type)
- context: (optional) Batch-level context for all questions
- effort_level: (optional) "low" or "high" (default; required for categorical/thresholded and for any conditional forecast)
- output_field: Name of the forecast quantity (required for numeric/date)
- units: Units of the forecast quantity (required for numeric)
- categories_field: Column with each row's outcomes as a JSON array of strings (required for categorical)
- thresholds_field: Column with each row's threshold conditions as a JSON array (required for thresholded)
- condition: (optional) Single shared condition string mapped over every row, making the forecast conditional on it
- condition_field: (optional) Name of a per-row column holding each row's own condition (mutually exclusive with condition)
- session_id / session_name: (optional)
```

**Conditional forecasting.** Conditionality is a modifier on any forecast type,
not a type of its own. Whenever the user frames a question with a condition or
intervention ("if A…", "given/assuming A", "conditional on A", "in the world
where A"), pick `forecast_type` from the outcome as usual (`date` for "when will
X", `numeric` for "what will the return be", `binary` for "will X happen", etc.)
and add the condition. Conditional forecasts are HIGH effort only. Each output
keeps the type's normal columns and produces them a second time, suffixed
`_given_condition` (the world where the condition holds) and `_given_not_condition`
(where it does not), so the two branches reflect one coherent view of how the
condition bears on the outcome, plus a `rationale`.

Two mutually exclusive ways to supply the condition:

- **Shared condition** (`condition`): a single condition string applied to every
  row. Use it for a one-off question, or for the "ask the same conditional about
  each of these" case, where the user brings a list of entities and asks how one
  shared intervention moves each entity's outcome (e.g. *"for each company, what
  will its Q3 stock return be if Claude Fable launches worldwide before August?"*).
- **Per-row column** (`condition_field`): the name of an input column holding each
  row's own condition, for a sheet where rows carry distinct conditions.

### futuresearch_decision
Forecast the outcome under each alternative of a choice the user controls
("if I fund this at $0 / $300k / $2M, will it ship by 2027?"). The outcome under
each alternative can be a probability, a number, or a date (`forecast_type`
`binary` / `numeric` / `date`, as in futuresearch_forecast). Use this, whenever
the "if" is the user's own decision: a conditional forecast is correlational,
a decision is causal.
```
Parameters:
- artifact_id: Artifact ID (UUID) from upload_data or request_upload_url
- data: Inline data as a list of row objects (must include "question" column)
- alternatives_field: (required) Column holding each row's mutually exclusive
  alternatives as a JSON array of 2-50 numbers or strings
- forecast_type: (optional) "binary" (default), "numeric", or "date" (the OUTCOME
  type forecast under each alternative)
- output_field: Name of the forecast quantity (required for numeric/date)
- units: Units of the forecast quantity (required for numeric)
- context: (optional) Table-level context applied to every row
- intervention: (optional) What executing an alternative means (publicity, timing,
  how the world responds). Replaces the default assumptions wholesale, so state
  the full set. Tell the user which assumptions were active when presenting results.
- session_id / session_name: (optional)
```
The output always contains a `rationale`. Additionally, for a binary outcome,
there is a `probabilities` field, which is a JSON object mapping each alternative to
the outcome's probability. For a numeric or date outcome, the output contains a
`percentiles` field, which is a JSON object mapping each alternative to its
`{p10, p25, p50, p75, p90}` record. The values across alternatives need not sum to
100 and need not be monotonic.

### futuresearch_multi_agent
Answer one question with a team of agents that each take a different angle, then
synthesize one structured result. Use it for a single deep question; use
futuresearch_agent when you have a table and want one agent per row.
```
Parameters:
- task: (required) Instructions for the multi-agent parallel research
- artifact_id / data: (optional) Omit or pass empty data for a standalone question
- directions: (optional) Up to 6 explicit research directions. Each must be a
  detailed, self-contained brief, not a short title. Auto-generated if omitted.
- response_schema: (optional) JSON schema for the synthesized response per row
- effort_level: (optional) "low" (3 agents) and "medium" (4, default) run fast;
  "high" (2 frontier agents) is deeper but slower
- return_table: (optional) MUST be true when the task asks for a list of items
  (e.g. "find 15 startups"). Pair with response_schema describing a single item.
- session_id / session_name: (optional)
```

### futuresearch_agent
Run web research agents on each row.
```
Parameters:
- task: (required) Natural language description of research task
- artifact_id: Artifact ID (UUID) from upload_data or request_upload_url
- data: Inline data as a list of row objects
- response_schema: (optional) JSON schema for per-row agent response
- session_id: (optional) Session UUID to resume
- session_name: (optional) Name for a new session
```

### futuresearch_rank
Score and sort rows based on qualitative criteria.
```
Parameters:
- task: (required) Natural language instructions for scoring a single row
- field_name: (required) Name of the score field to add
- artifact_id: Artifact ID (UUID) from upload_data or request_upload_url
- data: Inline data as a list of row objects
- field_type: (optional) "float" (default), "int", "str", or "bool"
- ascending_order: (optional) Sort ascending (default: true)
- response_schema: (optional) JSON schema for the response model
- session_id / session_name: (optional)
```

### futuresearch_dedupe
Remove duplicate rows using semantic equivalence.
```
Parameters:
- equivalence_relation: (required) Natural language description of what makes rows duplicates
- artifact_id: Artifact ID (UUID) from upload_data or request_upload_url
- data: Inline data as a list of row objects
- session_id / session_name: (optional)
```

### futuresearch_merge
Join two tables using intelligent entity matching (LEFT JOIN semantics).
```
Parameters:
- task: (required) Natural language description of how to match rows
- left_artifact_id / left_data: (required, exactly one) Left table — the table being enriched (all rows kept)
- right_artifact_id / right_data: (required, exactly one) Right table — lookup/reference (columns appended to matches)
- merge_on_left: (optional) Only set if you expect exact string matches or want to draw agent attention to a column
- merge_on_right: (optional) Same as merge_on_left for right table
- relationship_type: (optional) "many_to_one" (default), "one_to_one", "one_to_many", "many_to_many"
- use_web_search: (optional) "auto" (default), "yes", or "no"
- session_id / session_name: (optional)
```

### futuresearch_classify
Classify each row into one of the provided categories.
```
Parameters:
- task: (required) Natural language classification instructions
- categories: (required) Allowed categories (minimum 2)
- artifact_id: Artifact ID (UUID) from upload_data or request_upload_url
- data: Inline data as a list of row objects
- classification_field: (optional) Output column name (default: "classification")
- include_reasoning: (optional) Include reasoning column (default: false)
- session_id / session_name: (optional)
```

## Data Management

### futuresearch_browse_lists
Browse available reference lists of well-known entities (S&P 500, FTSE 100, countries, universities, etc.).
```
Parameters:
- search: (optional) Search term to match list names
- category: (optional) Filter by category (e.g. "Finance", "Geography")
```

### futuresearch_use_list
Import a reference list into your session and save it as a CSV.
```
Parameters:
- artifact_id: (required) artifact_id from futuresearch_browse_lists results
```

### futuresearch_upload_data
Upload data from a URL or local file. Returns an artifact_id for use in processing tools.
```
Parameters:
- source: (required) HTTP(S) URL (Google Sheets supported) or local CSV path (stdio mode only)
- session_id / session_name: (optional)
```

### futuresearch_request_upload_url
Request a presigned URL to upload a local CSV file (HTTP mode only).
```
Parameters:
- filename: (required) Name of the file to upload (must end in .csv)
```
Steps: call this tool → execute the returned curl command → use the artifact_id from the response.

## Task Lifecycle

### futuresearch_progress
Check progress of a running task. Blocks briefly to limit polling rate.
```
Parameters:
- task_id: (required) Task ID returned by the operation tool
```
After receiving a status update, immediately call futuresearch_progress again unless the task is completed or failed.

### futuresearch_results
Retrieve results from a completed task.
```
Parameters:
- task_id: (required) Task ID of the completed task
- output_path: (stdio) Full path to output CSV (must end in .csv)
- offset: (http, optional) Row offset for pagination (default: 0)
- page_size: (http, optional) Number of rows to load into context (default: auto threshold based on row count)
```
Only call after futuresearch_progress reports status "completed".

### futuresearch_cancel
Cancel a running task.
```
Parameters:
- task_id: (required) Task ID to cancel
```

### futuresearch_status
Check task status and show a live progress widget that auto-updates by polling.
Only registered for widget-capable clients (HTTP mode), so it is not available
in stdio mode such as Claude Code. After calling it once, do NOT also call
futuresearch_progress: the widget polls on its own.
```
Parameters:
- task_id: (required) Task ID to display
```

### futuresearch_task_cost
Get the billed cost of a completed task, in dollars. Cost settles some time
after the task finishes, so this returns "pending" until it does.
```
Parameters:
- task_id: (required) Task ID to price
```

## Sessions & Account

### futuresearch_list_sessions
List sessions owned by the authenticated user (paginated).
```
Parameters:
- offset: (optional) Number of sessions to skip (default: 0)
- limit: (optional) Max sessions per page (default: 25, max: 1000)
```

### futuresearch_list_session_tasks
List all tasks in a session with their IDs, statuses, and types.
```
Parameters:
- session_id: (required) Session ID (UUID) to list tasks for
```

### futuresearch_balance
Check the current billing balance for the authenticated user.
```
No parameters.
```

---

# Python SDK Reference

## Results

All operations return a result object. The data is available as a pandas DataFrame in `result.data`:

```python
result = await rank(...)
print(result.data.head())  # pandas DataFrame
```

## Operations

For quick one-off operations, sessions are created automatically.

### forecast - Predict probabilities

Produce probability estimates for binary questions:

```python
from futuresearch.ops import forecast

result = await forecast(
    input=DataFrame([
        {"question": "Will the US Federal Reserve cut rates by at least 25bp before July 1, 2027?",
         "resolution_criteria": "Resolves YES if the Fed announces at least one rate cut of 25bp or more."},
    ]),
    forecast_type="binary",
)
print(result.data[["question", "probability", "rationale"]])
```

Parameters: `input`, `forecast_type` (`"binary"` | `"numeric"` | `"date"` | `"categorical"` | `"thresholded"`), `effort_level`, `context`, `output_field` (required for numeric/date), `units` (required for numeric), `categories_field` (required for categorical), `thresholds_field` (required for thresholded), `condition` *or* `condition_field` (makes any type conditional), `session`

For **conditional** forecasts (P(B|A) and P(B|not A) for a condition A and outcome B), see "Conditional forecasting" under `futuresearch_forecast` above. Conditionality is a modifier on any `forecast_type`, not a type of its own: keep `forecast_type` describing the outcome B (taken from each row's `question`) and supply the condition A via `condition` (a single shared condition mapped over every row) or `condition_field` (a per-row condition column). The two are mutually exclusive.

Recommended input columns beyond `question`: `resolution_criteria`, `resolution_date`, `background`. For questions tied to a prediction market or forecasting platform (Polymarket, Kalshi, Metaculus, ...), also pass `market_creation_date` and `market_price` (with its as-of date), and copy resolution criteria verbatim from the platform — including fine print. Self-contained questions (e.g. "When will Anthropic IPO?") need none of these.

### decision - Outcome under each alternative

Forecast the outcome under each alternative of a choice you control. Use this
whenever the "if" clause is the user's own decision: a conditional forecast is
correlational, a decision is causal. The outcome under each alternative can be a
probability, a number, or a date, set by `forecast_type`, exactly as in `forecast`.

```python
import json
from pandas import DataFrame
from futuresearch.ops import decision

decisions = DataFrame([
    {
        "question": "Will the study be published in a peer-reviewed journal before 2028-01-01?",
        "grant_size": json.dumps(["$0 (no grant)", "$300k", "$2M"]),
        "resolution_criteria": "Resolves YES if it appears in a peer-reviewed journal before 2028-01-01.",
    },
])

result = await decision(input=decisions, alternatives_field="grant_size")
print(result.data[["question", "probabilities", "rationale"]])
```

The output contains `rationale` plus a per-alternative outcome column:
- For binary forecasts, a `probabilities` field containing a JSON object mapping
  each alternative to the outcome's probability.
- For numeric and date forecasts, a `percentiles` field containing a JSON object
  mapping each alternative to its `{p10, p25, p50, p75, p90}` record).

Parameters: `input`, `alternatives_field` (required, keyword-only), `forecast_type`
(`"binary"` | `"numeric"` | `"date"`), `output_field` (required for numeric/date),
`units` (required for numeric), `context`, `intervention`, `session`

### multi_agent - A team of agents on one question

Several agents each take a different angle on one question, then synthesize a single
structured answer. Use it for a deep single question; use `agent_map` when you have a
table and want one agent per row.

```python
import pandas as pd
from futuresearch.ops import multi_agent

result = await multi_agent(
    task="Research the current state of formal verification for AI systems",
    input=pd.DataFrame(),  # empty input for a standalone question
)
print(result.data.head())
```

Set `return_list=True` when the task asks for a list of items ("find 15 startups"),
pairing it with a `response_schema` describing one item.

Parameters: `task`, `input`, `session`, `directions` (up to 6 detailed, self-contained
briefs), `effort_level` (`low` 3 agents, `medium` 4 agents default, `high` 2 frontier
agents), `response_schema`, `join_with_input`, `return_list`

### agent_map - Batch processing

Run an AI agent across multiple rows:

```python
from futuresearch.ops import agent_map
from pandas import DataFrame

result = await agent_map(
    task="Find this company's latest funding round and lead investors",
    input=DataFrame([
        {"company": "Anthropic"},
        {"company": "OpenAI"},
        {"company": "Mistral"},
    ]),
)
print(result.data.head())
```

**Effort levels** - control research thoroughness:

- `LOW`: Quick lookups, basic web searches
- `MEDIUM` (default): More thorough research, multiple sources
- `HIGH`: Deep research, cross-referencing sources

```python
from futuresearch.ops import agent_map
from futuresearch.task import EffortLevel

result = await agent_map(
    task="Comprehensive competitive analysis",
    input=competitors,
    effort_level=EffortLevel.HIGH,
)
```

Parameters: `task`, `input`, `effort_level`, `response_model`, `session`

### rank - Score and rank rows

Score rows based on criteria you can't put in a database field:

```python
from futuresearch.ops import rank

result = await rank(
    task="Score by likelihood to need data integration solutions",
    input=leads_dataframe,
    field_name="integration_need_score",
    ascending_order=False,  # highest first
)
print(result.data.head())
```

**Structured output** - get more than just a score:

```python
from pydantic import BaseModel, Field

class AcquisitionScore(BaseModel):
    fit_score: float = Field(description="0-100, strategic alignment")
    annual_revenue_usd: int = Field(description="Estimated annual revenue in USD")

result = await rank(
    task="Score acquisition targets by product-market fit",
    input=potential_acquisitions,
    field_name="fit_score",
    response_model=AcquisitionScore,
    ascending_order=False,
)
```

Parameters: `task`, `input`, `field_name`, `field_type` (default: "float"), `response_model`, `ascending_order` (default: True), `session`

### classify - Categorize rows

Assign each row to one of the provided categories:

```python
from futuresearch.ops import classify

result = await classify(
    task="Classify this company by its GICS industry sector",
    categories=["Energy", "Materials", "Industrials", "Consumer Discretionary",
                 "Consumer Staples", "Health Care", "Financials",
                 "Information Technology", "Communication Services",
                 "Utilities", "Real Estate"],
    input=companies,
)
print(result.data[["company", "classification"]])
```

**Binary classification** - for yes/no questions, use two categories:

```python
result = await classify(
    task="Is this company founder-led?",
    categories=["yes", "no"],
    input=companies,
)
```

**With reasoning** - understand why each row was classified:

```python
result = await classify(
    task="Classify each company by its primary industry sector",
    categories=["Technology", "Finance", "Healthcare", "Energy"],
    input=companies,
    classification_field="sector",
    include_reasoning=True,
)
```

Parameters: `task`, `categories`, `input`, `classification_field` (default: "classification"), `include_reasoning` (default: False), `session`

### merge - Merge tables with AI matching

Join two tables when the keys don't match exactly (LEFT JOIN semantics). The AI knows "Photoshop" belongs to "Adobe" and "Genentech" is a Roche subsidiary:

```python
from futuresearch.ops import merge

result = await merge(
    task="Match each software product to its parent company",
    left_table=software_products,   # table being enriched — all rows kept
    right_table=approved_suppliers,  # lookup/reference table — columns appended to matches
    # merge_on_left/merge_on_right: omit unless you expect exact string matches
    # on the chosen columns or want to draw agent attention to them.
)
print(result.data.head())
```

Parameters: `task`, `left_table`, `right_table`, `merge_on_left`, `merge_on_right`, `relationship_type`, `use_web_search`, `session`

### dedupe - Deduplicate data

Remove duplicates using AI-powered semantic matching. The AI understands that "AbbVie Inc", "Abbvie", and "AbbVie Pharmaceutical" are the same company:

```python
from futuresearch.ops import dedupe

result = await dedupe(
    input=crm_data,
    equivalence_relation="Two entries are duplicates if they represent the same legal entity",
)
print(result.data.head())
```

**Strategies** - control what happens after clusters are identified:

- `"select"` (default): Pick the best representative from each cluster
- `"identify"`: Cluster only, no selection (for manual review)
- `"combine"`: Synthesize a single combined row per cluster

```python
result = await dedupe(
    input=crm_data,
    equivalence_relation="Same legal entity",
    strategy="select",
    strategy_prompt="Prefer the record with the most complete contact information",
)
deduped = result.data[result.data["selected"] == True]
```

Results include `equivalence_class_id` (groups duplicates), `equivalence_class_name` (human-readable cluster name), and `selected` (the canonical record when using select/combine strategy).

Parameters: `input`, `equivalence_relation`, `strategy`, `strategy_prompt`, `session`

### single_agent - Single input task (DEPRECATED)

> `single_agent` is deprecated and will be removed in a future release; calling it
> emits a `DeprecationWarning`. Use `agent_map` (one task per row, optionally with
> `return_table=True`) or `multi_agent` (parallel agents synthesized per row).
> Both accept an empty input for a standalone task.

Run an AI agent on a single input:

```python
from futuresearch.ops import single_agent
from pydantic import BaseModel

class CompanyInput(BaseModel):
    company: str

result = await single_agent(
    task="Find the company's most recent annual revenue and employee count",
    input=CompanyInput(company="Stripe"),
)
print(result.data.head())
```

**No input required** - agents can work without input data:

```python
result = await single_agent(
    task="What company has reported the greatest cost reduction due to internal AI usage?",
)
```

**Return a table** - generate datasets from scratch:

```python
from pydantic import BaseModel, Field

class CompanyInfo(BaseModel):
    company: str = Field(description="Company name")
    market_cap: int = Field(description="Market cap in USD")

result = await single_agent(
    task="Find the three largest US healthcare companies by market cap",
    response_model=CompanyInfo,
    return_table=True,
)
```

Parameters: `task`, `input`, `effort_level` (LOW/MEDIUM/HIGH), `response_model`, `return_table`, `session`

## Explicit Sessions

For multiple operations that should be grouped together, use an explicit session:

```python
from futuresearch import create_session

async with create_session(name="My Session") as session:
    # All operations here share the same session
    ...
```

## Async Operations

All operations have `_async` variants for background processing. These need an explicit session since the task persists beyond the function call:

```python
from futuresearch import create_session
from futuresearch.ops import rank_async

async with create_session(name="Async Ranking") as session:
    task = await rank_async(
        session=session,
        task="Score this organization",
        input=dataframe,
        field_name="score",
    )
    print(f"Task ID: {task.task_id}")  # Print this! Useful if your script crashes.

    # Continue with other work...
    result = await task.await_result()
```

**Tip:** Print the task ID after submitting. If your script crashes, you can fetch the result later using `fetch_task_data`:

```python
from futuresearch import fetch_task_data

# Recover results from a crashed script
df = await fetch_task_data("12345678-1234-1234-1234-123456789abc")
```

## Long-Running Operations (MCP)

FutureSearch operations (classify, rank, dedupe, merge, forecast, agent) take 1-10+ minutes.
All MCP tools use an async pattern:

1. Call the operation tool (e.g., `futuresearch_agent(...)`) to get a task_id
2. Call futuresearch_progress(task_id) — the tool handles pacing internally
3. After each status update, immediately call futuresearch_progress again
4. When status is "completed" or "failed", call futuresearch_results(task_id)

## Chaining Operations

Operations can be chained to build complete workflows. Each step's output feeds the next:

```python
from futuresearch import create_session
from futuresearch.ops import classify, dedupe, rank

async with create_session(name="Lead Pipeline") as session:
    # 1. Filter to qualified leads
    classified = await classify(
        session=session,
        task="Does this lead have a company email domain (not gmail, yahoo, etc.)?",
        categories=["qualified", "unqualified"],
        input=leads,
    )

    # 2. Dedupe across sources
    deduped = await dedupe(
        session=session,
        input=classified.data[classified.data["classification"] == "qualified"],
        equivalence_relation="Same company, accounting for Inc/LLC variations",
    )

    # 3. Prioritize for outreach
    ranked = await rank(
        session=session,
        task="Score by likelihood to convert",
        input=deduped.data[deduped.data["selected"] == True],
        field_name="conversion_score",
    )
```

## Best Practices

FutureSearch operations have associated costs. To avoid re-running them unnecessarily:

- **Separate data processing from analysis**: Save FutureSearch results to a file (CSV, Parquet, etc.), then do analysis in a separate script. This way, if analysis code has bugs, you don't re-trigger the FutureSearch step.
- **Use intermediate checkpoints**: For multi-step pipelines, consider saving results after each FutureSearch operation.
    - You are able to chain multiple operations together without needing to download and re-upload intermediate results via the SDK. However for most control, implement each step as a dedicated job, possibly orchestrated by tools such as Apache Airflow or Prefect.
- **Test on a slice first**: Before running a large job, pass a few rows (e.g. `input=df.head(5)`) to check the task wording and output shape, then run the full table.