# H2O.ai

> H2O.ai is an open-source artificial-intelligence and machine-learning company. Its platform
> spans H2O-3 (a distributed, in-memory ML engine), H2O Driverless AI (automatic machine
> learning), H2O MLOps (model deployment, scoring and monitoring), H2O Wave (a Python/R
> framework for realtime AI apps), H2O LLM Studio, and Enterprise h2oGPTe (a private
> generative-AI, RAG and agent platform). Two machine-readable REST contracts are published:
> Enterprise h2oGPTe (OpenAPI 3.0.1, 422 operations, 24 tags) and H2O MLOps Scoring
> (OpenAPI 3.0.0, 7 operations). H2O.ai also ships a first-party MCP server that projects the
> h2oGPTe REST API onto agent tools.

This file was generated by API Evangelist from the public H2O.ai surface on 2026-08-04.
H2O.ai does not publish an llms.txt of its own (probed h2o.ai/llms.txt 404, docs.h2o.ai/llms.txt 404).

## APIs

- [Enterprise h2oGPTe REST API](https://docs.h2o.ai/enterprise-h2ogpte/rest-api): Collections, document
  ingestion, chat, agents, AI assistants, extractors, GraphRAG, guardrails, memory blocks, models,
  prompt templates, jobs, scheduled tasks, permissions, API keys and secrets. Base URL
  `https://h2ogpte.genai.h2o.ai/api/v1`. Auth: `Authorization: Bearer sk-...`.
- [H2O MLOps Scoring REST API](https://docs.h2o.ai/mlops/model-scoring/mlops-scoring-rest-api): runtime
  scoring for a deployed model — model id, schema, sample request, capabilities, row scoring,
  media scoring and Shapley contributions. Host is per-deployment.
- [h2oGPTe MCP Server](https://github.com/h2oai/h2ogpte-mcp-server): first-party local stdio MCP proxy
  over the h2oGPTe REST API. One tool per REST operationId.

## Specs

- [h2oGPTe OpenAPI 3.0.1](https://h2ogpte.genai.h2o.ai/api-spec.yaml)
- [MLOps Scoring OpenAPI 3.0.0](https://docs.h2o.ai/mlops/model-scoring/mlops-scoring-rest-api)
- [h2oGPTe Swagger UI](https://h2ogpte.genai.h2o.ai/swagger-ui/) — authenticated, live requests

## Docs

- [H2O.ai documentation hub](https://docs.h2o.ai/)
- [Enterprise h2oGPTe docs](https://docs.h2o.ai/enterprise-h2ogpte/)
- [Enterprise h2oGPTe get started](https://docs.h2o.ai/enterprise-h2ogpte/get-started)
- [API keys guide](https://docs.h2o.ai/enterprise-h2ogpte/guide/apis)
- [h2oGPTe changelog](https://docs.h2o.ai/enterprise-h2ogpte/changelog)
- [h2oGPTe major releases](https://docs.h2o.ai/enterprise-h2ogpte/releases)
- [H2O MLOps docs](https://docs.h2o.ai/mlops/)
- [H2O AI Cloud docs](https://docs.h2o.ai/haic-documentation/)
- [H2O Wave docs](https://wave.h2o.ai/)

## SDKs and CLIs

- Python: `h2o`, `h2ogpte`, `h2o-mlops`, `driverlessai`, `h2o-wave`, `h2o-lightwave`, `h2o-authn`,
  `h2o-cloud-discovery`, `h2o-drive`, `h2o-engine-manager`, `h2o-notebook`, `h2o-secure-store`,
  `h2o-featurestore`, `h2osteam`, `h2ogpte-mcp-server` (PyPI)
- JavaScript: `h2ogpte`, `h2o-wave` (npm)
- R: `h2o` (CRAN)
- Java: `ai.h2o:*` — 97 artifacts on Maven Central, including `h2o-core` and `mojo2-runtime-api`
- CLI: `wave` (from `pip install h2o-wave`) — https://wave.h2o.ai/docs/cli

## Operations

- [Status page](https://h2oai.statuspage.io/) — H2O AI Cloud, Atlassian Statuspage
- [Support](https://support.h2o.ai/)
- [Security](https://h2o.ai/security/) and [report a vulnerability](https://h2o.ai/security/report-vulnerability)
- [Trust Center](https://trust.h2o.ai/) — SOC 2 Type 2, FedRAMP High, IRAP, HITECH
- [Blog](https://h2o.ai/blog/)
- [GitHub organization](https://github.com/h2oai)
- [Sign up (freemium tier)](https://genai.h2o.ai/appstore)
- [Terms of use](https://h2o.ai/legal/terms-of-use/) · [Privacy](https://h2o.ai/legal/privacy/)

## Known gaps

An agent integrating with these APIs should know what is NOT published:

- No idempotency key — retrying a POST creates a duplicate.
- Errors use a vendor `{code, message}` envelope, not RFC 9457 problem details; no error-code registry.
- Rate limiting exists (h2oGPTe 1.7.0) but no limit values, header names or Retry-After contract are documented.
- No deprecation or version-support policy; no Sunset/Deprecation headers.
- No webhooks and no AsyncAPI — long-running work is job-polling (28 job operations).
- No /.well-known/security.txt, no api-catalog, no A2A agent card on any host.
- No test-mode key or sandbox host; the interactive Swagger UI hits the live tenant.