HydrAI documentation
A small, curated vocabulary that makes a Hydra/JSON-LD API legible and safe for LLM agents. A conservative superset of Hydra core — everything Hydra can express, plus the handful of things an agent needs that Hydra does not.
HydrAI is deliberately opinionated. Where the standards leave a choice, it makes one, states why, and gives you a working reference implementation that proves it. If you want the raw, unopinionated stack, it is all still there underneath — HydrAI only ever adds, and every borrowed term points back to its source.
In one line: Hydra was designed for machines that navigate. Agents reason in examples and plan and act. HydrAI is Hydra for the agent generation.
Where to go next
- The vision — why semantically grounded data is the mechanism for safe, accurate AI, and how HydrAI extends that semantic layer to the surface agents act through.
- Getting started — connect the hosted agent to a Hydra API, or wire up the MCP server, in a few minutes.
- The nine opinions — the stances baked into the vocabulary, and the reasons for them.
- Vocabulary design note — the full design rationale: the earned-term test, the namespace architecture, and the wire posture.
- Browse the namespace — the published
agent#vocabulary, in HTML, Turtle, and JSON-LD.
The three pieces of this repository
This site is served from the HydrAI monorepo, which holds three things that stay in step with each other:
| What it is | |
|---|---|
| The vocabulary | vocab/ — the agent# ontology, published at hydrai.org/ns/agent. |
| The generic agent | examples/hydra-client/ — a generic Hydra/JSON-LD client with an agent loop, hosted here and consumable as an MCP server. |
| This site | site/ — the landing pages and these docs. |
The reference implementation is the conformance proof: the vocabulary is extracted from working code, not designed in advance and hoped into use.
Status
HydrAI is 0.1 and built by harvest, not by decree. Terms are minted one at a time, each shipping with client code that consumes it. Term IRIs are stable across versions; the vocabulary version is not — pin the context version you build against.