Open source SDK

Ship predictive modelsfor your businessyour business, fraud detection, churn prediction, pricing, anomaly detection, customer segmentsWith an API call.

Point it at your historical data, get back a prediction API. No training runs, no ML team. We’re in private alpha, working with a few design partners.

claude mcp add zerograd -- zerograd-mcpWorks with Claude Code, Cursor, Codex, Gemini CLI and any other MCP client.
1from zerograd import ZeroGrad
2
3zg = ZeroGrad() # runs on your machine
4
5fraud = zg.create(
6 "fraud-detection",
7 examples="transactions.parquet", # past outcomes
8 target="is_fraud",
9 drop=["transaction_id"],
10)
11
12fraud.predict({"amount": 1840, "country": "BR"})
OutputChampionTabICLv2, best of 4 models on your holdout
{ "prediction": 1, "score": 0.94, "proba": { "0": 0.06, "1": 0.94 } }
Built by
engineers from
Tabular foundation models

What's a tabular foundation model, and why should you care?

The short version

It's like ChatGPT, but for predictive tasks: get predictions over your data in minutes, without training a model.

  1. 01 You already know this

    ChatGPT and Claude learn from your prompt

    You don’t retrain them every time you need a new task done. Show a few examples and they pick it up from context, without changing their weights: no feature engineering, no expensive retraining, no team tuning a model for each use case.

  2. 02 Same idea, on tables

    A tabular foundation model reads rows instead of words

    It was pretrained once, on millions of tables. Give it your history as examples and it predicts the next row: fraud, churn, demand. No training run, no tuning.

    POST /v1/fraud-detection/predict
    amountcountryageis_fraud
    $1,840BR9dfraud
    $62AR3yok
    $410US41dok
    $1,320BR6d?
    200 OK

    p(fraud) = 0.94

  3. 03 2026

    They broke through

    During 2026, tabular foundation models went from research papers to solid baselines. On many real datasets they already hold their own against gradient-boosted trees and deep learning.

  4. 04 Next

    They follow the LLM curve

    Language went from one model per task to one pretrained model that gets better every few months. Tables are on the same path, and the next generation will likely beat what you run today.

  5. 05 You

    So keep them in your stack

    Test them on your data, use them as your baseline, and be ready to adopt the next one the week it ships. That's the problem we solve.

Use cases

Ship machine learning around your business use cases.

Without depending on an expensive team. You ship fraud detection, not an XGBoost container: a handful of verbs cover most of the predictive work products need, and more are on the way. Some of what teams build with them:

CLASSIFY01
  • Block card-not-present fraud at checkoutFintech
  • Score which trials will convert to paidSaaS
  • Flag loans likely to default before approvalLending
  • Find accounts about to cancelSubscriptions
ESTIMATE02
  • What a listing will sell forReal estate
  • Lifetime value of a new customerE-commerce
  • Days until an invoice gets paidFinance
  • Expected loss on a claimInsurance
DETECT03
  • Unusual ledger entries before month-end closeFinance
  • Refund spikes by merchantPayments
  • Sensor readings outside the machine's normIndustrial
  • Logins that don't look like the userSecurity
SEGMENT04
  • Customer groups for each campaignMarketing
  • Stores that behave alikeRetail
  • Merchants with the same risk profilePayments
  • Patients with similar care pathsHealth
FORECASTSoon05
  • Weekly units per SKU and storeRetail
  • Support tickets per hour, to staff the queueOps
  • Cash each ATM will need next weekBanking
  • Next month's usage per customer, for billingInfra
RECOMMENDSoon06
  • What to add to the cart nextE-commerce
  • Which offer to send each merchantMarketplaces
  • The next article a reader opensMedia
  • Which lead a rep should call firstSales
CAUSAL IMPACTSoon07
  • Who buys only because of the discountRetail
  • Did the new onboarding cut churn?SaaS
  • Which customers a retention call actually savesTelco
  • What a price change did to demandPricing
MARKETING MIXSoon08
  • Sales each channel really droveConsumer brands
  • Where the next dollar of budget goesGrowth
  • How long a TV flight keeps paying offMedia
  • Spend at which a channel saturatesPerformance
Open source SDK · MCP

Use it from your coding agent.

The SDK ships an MCP server, so Claude Code, Cursor, Codex or any other MCP client can create capabilities and predict when you ask in plain language. It runs on your machine with no API key, and your data stays there: the agent passes file paths, never rows.

$pip install "zerograd[mcp]"$claude mcp add zerograd -- zerograd-mcp

Other clients take the same command, zerograd-mcp, over stdio.

Set it up
Zero Grad Platform · managed

Everything around the model, handled.

The SDK is open source and runs on your machine. The platform is the managed layer on top, for teams running predictions in production: you call one endpoint, and we run the evaluation, deployment, monitoring and upgrades.

The Zero Grad platform demo: a fraud detection capability with its production version, challenger and how to call it
The Zero Grad platform, on a demo workspace with simulated data. Explore the demo →
Open source & platform

Run it yourself, or let us run it.

Open source, self-hostedManaged platform
Where it runsYour laptop, notebook or your own cloudManaged GPU inference with SLAs
Who operates itYou, with full controlWe do: scaling and upgrades
ChampionshipWhenever you run itAlso scheduled, on fresh history
In productionPlugs into your own stackVersions, shadow, promotion, monitoring
PriceFree, Apache 2.0Pay as you go, or a platform plan

The SDK is complete on its own: stay open source and self-hosted for as long as you like. If you ever want us to run it, it is the same code: ZeroGrad() runs on your side, ZeroGrad(remote=True) on the platform.

Monitoring

Always the best model.

We watch for drift and re-run a championship on your latest data. The winner waits in shadow, ready to promote.

  1. TabPFN · new release.958shadow
  2. TabPFN · current.947prod
  3. TabFM.944−.003
Deployments

Ships like app code.

Every version runs in shadow first, then canary or A/B behind your SLOs, with one-click rollback.

ShadowCanaryA/BBlue-green

v8 10% of traffic

Upgrades

Better models, zero code.

New foundation models keep landing. We prove them on your workload and swap them in behind the same endpoint.

Agents

Your agent gets a probability, not a guess.

Every capability is an MCP tool with a Skill that says when to call it. Claude, Cursor, Copilot, LangGraph or your own loop.

  1. agent

    → fraud_detection.predict

  2. zerograd

    probability 0.94 · model v7

  3. agent

    Blocking it: 94% is over the 0.90 threshold.

Where it runs

Runs where your data lives.

  • Zero Grad CloudAvailable
  • AWS VPCEnterprise
  • GCP VPCEnterprise
  • Azure VNet
  • On-premise
  • Edge
How it works

From objective to production — and it keeps getting better.

You do two things: describe the prediction and call the endpoint. Everything in between is ours, and every number comes with the model, the data window and the measured error behind it.

See it in action
  1. 01

    Objective + examples

    Point at past outcomes and say what you want predicted. Nothing gets trained.

  2. 02

    Evaluation

    Validation, candidate approaches and temporal backtests — on your history, not a benchmark.

  3. 03

    Deployyou · one call

    An autoscaled endpoint, with monitoring and drift checks on from the first request.

  4. 04

    One API

    classify · estimate · detect · segment. Your application consumes an API; you don't manage models.

  5. 05

    Observability

    Quality, latency, data health and the business metric you actually care about.

  6. 06

    A better model appears

    A challenger beats production on your data, under your constraints.

  7. 07

    Shadow · A/B · canary · blue-green

    Proven on live traffic, behind the SLOs and guardrails you set.

  8. 08

    Safe improvementyou · zero code

    Promoted with instant rollback. You change zero code.

Private alpha · design partners

What ML feature are you not building because it's too expensive to operate?

We're working with a few design partners who have real predictive workloads. Bring a historical dataset; leave with a benchmark.

Read the docs