New · Connect your AI with MCP

The Prediction OS
for every decision.

Bring your data and name the outcome. predictAI discovers what drives it, trains and races models, and keeps the best one answering — over REST, official SDKs, and live WebSockets.

Get startedAPI Reference
pip install predictai·npm install @predictai/sdk
create a goal · python
from predictai import PredictAI

client = PredictAI(token="pa_live_…")

goal = client.goals.post_goals(json={
    "target_key": "daily_sales",
    "goal_type": "forecast_value",
})

# drivers discovered
# models raced
# champion deployed
wss · goal:b7e9c2d4
goal.status      { "status": "serving" }
alerts.evaluated { "fired": 1 }

One loop, end to end

Data in, predictions out — and every stage is an API you can hold onto or hand over.

Core concepts
01

Ingest

Connect a source, upload a file, or push values over HTTP — data becomes signals.

02

Understand

Correlation, relevance, causality — see what actually moves your numbers.

03

Train

Pipelines race a pool of models on the same data; the best score wins serving.

04

Predict

One stable URL per goal or pipeline — and every step streamed over WebSockets.

Or skip the loop — state a goal

Two fields — the metric and the question. Discovery finds the drivers, a tournament picks the champion, scenarios answer your what-ifs, and alerts watch the result.

{ "target_key": "daily_sales", "goal_type": "anomaly" }

Model Context Protocol

Your AI already knows predictAI.

The official MCP server hands the whole platform to Claude, Cursor, or any MCP-capable agent. Say “upload this CSV and warn me if next week looks bad” — the agent ingests the file, creates a goal, and arms an alert on its own.

Connect your AI
mcp.json · that's the whole setup
{
  "mcpServers": {
    "predictai": { "url": "https://mcp.predict.ai/mcp" }
  }
}
# then, in your agent:
you · forecast daily_sales and alert me on dips
agent · create_goal → wait_for_goal → create_alert ✓

Take it to code

Every endpoint in these docs is a method call in the official SDKs — auth, workspace scoping, errors, and WebSockets handled. Every code sample on the site has cURL, Python, and Node tabs; pick a language once and the whole site follows.

pip install predictai

Python 3.9+ · requests only · optional realtime extra

npm install @predictai/sdk

TypeScript · zero dependencies · Node 18+ & browsers

Explore the platform

Every capability is a small, predictable REST surface. Start anywhere — they compose.

Ingestion

Connect databases, SaaS tools, files, and webhooks; schedule syncs.

Signals

Your data as named, analyzable series — plus the events that move them.

Segments

Model-ready tables: the grid, normalization, engineered features.

Models

One catalog: custom architectures, foundation models, and BYOM.

Pipelines

Bind a segment to a pool of models — every run races them all.

Trainings

Run history, accuracy reports, artifacts you can download.

Inference

Predictions by goal, pipeline, or deployment — one stable URL each.

Deployments

Put winners into serving — shared or dedicated capacity.

Goals

Name an outcome; the platform builds and runs the forecast.

Billing

Credits, quotas, usage, and pricing you can query before you spend.

Workspaces

Tenancy, organizations, members, roles, and invitations.

Realtime

WebSocket channels for goals, alerts, trainings, and live progress.

Notifications

The activity feed, Slack / PagerDuty / webhook delivery, preferences.

Account

API tokens, scopes, and your profile.

Ready to make your first call?

Grab an API token and get a prediction in about five minutes.

Open the quickstartBrowse examples