Hanzo OpenAPI

hanzo.yaml is THE published document, and it is OUTPUT. One OpenAPI 3.1 surface for every Hanzo service — the one every SDK, doc site and tool generates from — DERIVED by publish.py from hanzo-inc/cloud's own emitted openapi.yaml at one pinned release. To change the API, change the code in hanzo-inc/cloud. There is nothing here to edit that would make a route exist.

curl -fsSL -H "Authorization: Bearer $GITHUB_TOKEN" \
  https://raw.githubusercontent.com/hanzoai/openapi/main/hanzo.yaml

This repo is private, so that URL needs a token — unauthenticated it is a 404, not a 403, which reads like a missing file rather than a missing header. Any fetch of it without an Authorization header is broken; curl -f under set -e turns that into a build failure with nothing that names the cause.

Nothing generates from anything else, and nothing is pushed to an SDK. The SDK repos PULL this file and regenerate themselves; a client is never hand-written and never receives a patch from here.

Until recently this file was a hand-merged union of 52 authored specs with cloud's document laid on top, and it carried 185 operations cloud does not serve while missing 425 it does. Both are gone: publish.py derives it, and publish.py --check regenerates and diffs, so an operation no code declares is now structurally impossible. See LLM.md for the measurements.

Forward-only. No backwards compatibility, no deprecated paths, no /api/ prefix, no cross-brand references. A single bearer JWT from Hanzo IAM (https://hanzo.id) authenticates every route, and every route is /v1/<service>/<resource> — the same path through the gateway (api.hanzo.ai) and through the service's own hostname.

One exception — AI inference is top-level. hanzoai/ai serves its OpenAI- and Claude-compatible surface at the bare /v1/* paths (/v1/chat/completions, /v1/completions, /v1/embeddings, /v1/models, /v1/rerank, /v1/messages) because every OpenAI/Anthropic SDK hard-codes them: there, the path IS the compatibility contract.

curl -H "Authorization: Bearer ${HANZO_TOKEN}" \
     https://api.hanzo.ai/v1/chat/completions -d '{"model":"zen5","messages":[...]}'

curl -H "Authorization: Bearer ${HANZO_TOKEN}" \
     https://api.hanzo.ai/v1/billing/balance

OIDC discovery: https://hanzo.id/.well-known/openid-configuration. The live document is https://api.hanzo.ai/v1/openapi.json — cloud's emission, unprojected, which is this file's input rather than this file.

The files

File	What it is
hanzo.yaml	the published document. GENERATED by publish.py; never edited.
.spec-lock	which hanzo-inc/cloud release it is a projection of — repo, path, ref, sha256.
CAPABILITIES.md	GENERATED index, one row per served capability. Never hand-edit.
capabilities.yaml	the one editorial file: which DOMAIN each capability is shown under.
flows.yaml	the canonical example set — the same six journeys in every SDK.
publish.py generate.py skills.py	publish · clients · agent skills.

There are no <service>/openapi.yaml specs any more, and nothing to author. A capability appears the release it starts being served and leaves the release it stops.

One command

python3 publish.py            # re-pin to hanzo-inc/cloud's main, derive, write
python3 publish.py --check    # re-derive at the pinned ref and diff
python3 publish.py --current  # has cloud's document moved past the pin?
python3 publish.py --served   # THE GATE: is every published operation one api.hanzo.ai answers for?

The first three read a hanzo-inc/cloud checkout. --served reads nothing but https://api.hanzo.ai/v1/openapi.json — cloud's own emission, unauthenticated, from the binary that is running — which is why it is the one wired into hanzo.yml's test: block and the one that actually runs. It refutes in one direction: a published operation the deployment does not answer for fails the build, and it never fails on an unreachable API.

It reads a hanzo-inc/cloud checkout (../cloud, or --cloud DIR, or CLOUD_DIR) and writes exactly three files: hanzo.yaml, CAPABILITIES.md, .spec-lock. Nothing is ever written back to hanzo-inc/cloud.

Cloud's document is the one description of this API that cannot lie about what is served — cloud weaves it from its own routers and gates the weave by regenerating from source. publish.py adds no operation and no prose; it applies six rules that make that document GENERATABLE, because openapi-generator refuses it as it stands (1012 errors, one per route published with an address and no responses) and writes zero files in every language. Each rule is a candidate to move upstream into cloud's emitter; see LLM.md.

Conventions

• The /v1 path is the immutable contract — there is no v2, ever. Breaking changes ship as new resources.
• info.version: 8.0.0 is the generation of this PUBLICATION, and every SDK's package version is cut from it. The emitting cloud release is info.x-spec.ref; the API's own version is /v1. Three things, three places.
• operationIds are the document's own (get_billing_balance) — no <svc>_ prefix, because a prefix invented here is a second naming authority and the bare id is what the MCP server already uses as a tool name.
• No deprecated: true. No /api/ prefix. No dates in filenames.
• Security: every operation uses cloud's bearer scheme. No service has its own login.
• Errors: 4xx client, 5xx server, JSON body with status and msg.
• Nothing here can declare a route cloud does not serve. skills.py has no liveness filter — an unserved route would ship as a live instruction to an agent to call a dead endpoint — so the guarantee has to be structural rather than a habit. See LLM.md, "Who reads this repo".

Validate

python3 publish.py --served                      # nothing published is unserved
python3 publish.py --check                       # the artifact is its input's projection
python3 test_publish.py                          # the six rules, offline
python3 test_flows.py                            # every example's operationId resolves

java -jar ~/.cache/openapi-generator/openapi-generator-cli-7.14.0.jar validate -i hanzo.yaml
npx -y @redocly/cli preview-docs hanzo.yaml

Clients

generate.py runs openapi-generator (pinned 7.14.0) over hanzo.yaml for every language; sdks.yaml is the per-language data, and each SDK repo carries only a call site. There is no second generator and no unified SDK monorepo.

python3 generate.py --all            # regenerate every client in place
python3 generate.py go               # just one
python3 generate.py --all --check    # fail if a committed client drifted

Lang	Canonical repo	Generator	Ships as
Python	hanzoai/python-sdk	python	hanzoai on PyPI
TypeScript	hanzoai/js-sdk	typescript-axios	hanzoai on npm
Java · Kotlin	hanzoai/java-sdk · hanzoai/kotlin-sdk	java · kotlin (okhttp)	ai.hanzo:hanzo-{java,kotlin}-cloud
Rust	hanzo-rs/sdk	rust (reqwest)	crates/hanzo-client — own scripts/generate.sh (needs a template override)
Go	hanzo-go/sdk	go	package hanzoai at the module root — own scripts/generate.sh (no directory to own)

A language stays in sdks.yaml while its whole invocation is expressible as data, and leaves when it needs something that is not — a template file, or a layout take cannot describe. It then owns the whole invocation in its own repo. A row here AND flags there is two declarations of one contract.

Examples are not a per-SDK decision either: flows.yaml names the six journeys (hello, chat, money, store, agent, tools) and every SDK renders the same six from the same operationIds. python3 test_flows.py is the gate — an id that no longer resolves fails here, once, instead of in six clients separately.

Agent skills

skills.py derives /.well-known/agent-skills/ from hanzo.yaml, cut by each operation's own capability tag, clustering GET endpoints by resource — read-only by construction, no destructive verb ever becomes a skill — and emits a SKILL.md per cluster plus an index.json carrying a sha256 of every file.

python3 skills.py                 # dist/agent-skills/<brand>/…  (hanzo, lux, zoo)
python3 skills.py --no-services   # unified tree only (what the cloud binary embeds)
python3 skills.py --check         # drift gate
python3 test_skills.py            # golden test

Output is white-labeled per brand: a Lux surface reads Lux / api.lux.network / lux.id, a Zoo surface Zoo / api.zoo.ngo / zoo.id — never Hanzo. dist/ is generated at build time and is not committed.

License

MIT OR Apache-2.0, at your option — per HIP-0137. Copyright 2016-2026 Hanzo AI, Inc.