The hold read a body at four bytes a token, but a chat model's tokenizer reads a digit, a hex character or an emoji's byte as one, and the decision service writes the body out again before the router counts it (1e15 becomes 1000000000000000.0): a 65 KB state the hold priced at 7 cents settled at 56, and a wallet went negative. A model whose output is priced is now held at a token a byte plus 1,024 for the prompt around it, and an answer is settled and debited at most its hold whoever pays, a plan included; the excess is ours, and logged. Until the service passes the state's bytes through, a number-heavy body can still cost past its hold, about $1.76 a call at worst. A token count is the service's report of what a router relayed, so each is bounded where it is priced: none below zero, output at the 4,096 a completion is asked for, input short of where count times price overflows. A reported 1e9 output tokens billed $7,260 on a 3-cent hold, and 1.3e15 wrapped to a negative debit. An answer whose usage does not read as two whole token counts (and a cost, where stated, a number of 0 or more) is refused 502 and billed nothing, whatever model answered, rather than billed at zero for the tokens it did not state. The service writes usage from integer fields, so only a broken service reaches it, and the loss is the one any refusal already accepts. Kai's and Jev's bills and holds are unchanged for every count: their output is priced at zero and the byte-rate hold and the cap apply to priced output alone.
ai
ai is the AI subsystem of the Hanzo cloud: it serves the /v1 inference surface, routes
each request to a model, meters it, and carries the RAG and model-hub surfaces alongside.
Pure Go.
It used to be the whole hanzoai/cloud repository. When the unified binary took the
cloud name (HIP-0106), this became ai and now mounts as the ai subsystem inside it.
That is why the root of this repository is package ai, a library, and not a main.
Call it
The hosted instance is api.hanzo.ai. Get a key at hanzo.ai:
curl -H "Authorization: Bearer sk-YOUR-API-KEY" \
https://api.hanzo.ai/v1/chat/completions \
-d '{"model":"zen5","messages":[{"role":"user","content":"Hello"}]}'
Three auth modes are accepted: a Hanzo API key (sk-*), a JWT from hanzo.id OAuth, or a
provider key you already hold, in which case the request is routed with your account
rather than metered on ours.
Run it yourself
docker run -p 8000:8000 ghcr.io/hanzoai/ai:vX.Y.Z # pin a release, never :latest
Or build the server binary. It is cmd/aid — building the repository root gives you a
library archive, not something you can run:
CGO_ENABLED=0 go build -o aid ./cmd/aid
./aid -h
compose.yml brings up the server with its Postgres.
Models
The gateway owns the catalog; this service does not keep a second list. Ask it:
curl -H "Authorization: Bearer sk-YOUR-API-KEY" https://api.hanzo.ai/v1/models
https://catalog.hanzo.ai/v1/models answers the same question without a key. Zen is our
own family — enso, enso-ultra and the zen5 line, owned_by: hanzo. Models you
connect with your own provider keys are routed as well, resolved from Hanzo KMS and scoped
to your org.
Auto-routing
Send "model": "auto" (alias "zen-router") and the request is classified into a coarse
task — code, reasoning, math, creative, vision, long-context, cheap-chat, general — and
mapped to the first servable model in the matching preference list, before pricing and
billing resolve. So an auto request is billed and reported as exactly the model that
served it.
Turn it on in conf/models.yaml (the cloud-api-models ConfigMap in production), or with
ROUTER_ENABLED=true / ROUTER_ENDPOINT=<url>:
router:
enabled: true
endpoint: "" # "" = built-in heuristic; a URL = the learned router
cost_ceiling: 0.0
prefer:
code: [zen5-coder]
reasoning: [enso-ultra, enso]
cheap_chat: [zen5-mini, zen5-flash]
default: [zen5]
curl -H "Authorization: Bearer sk-YOUR-API-KEY" \
-H "X-Max-Cost: 2.0" -H "X-Max-Latency-Ms: 800" \
https://api.hanzo.ai/v1/chat/completions \
-d '{"model":"auto","messages":[{"role":"user","content":"refactor this function"}]}'
# response header: X-Routed-Model: <the model that served it>
X-Routed-Model and the model field in the body always report what actually ran.
X-Max-Cost (per 1k, float) and X-Max-Latency-Ms (int) express a per-request budget.
Two strategies sit behind one interface. With endpoint set, ai POSTs
{prompt, tasks?, slo} to {endpoint}/route and takes its {model, task, confidence}
reply — that is the learned router,
zenlm/zen-router. On any error, or with no
endpoint configured, it falls back to a built-in Go heuristic, so auto works with no
extra infrastructure.
Per-org override
Orgs can opt in or out through the admin OrgSettings surface (/v1/*-org-settings,
global-admin-gated like /v1/*-model-route). The org row carries autoRouting: ""
(unset), "enabled" or "disabled", and blends with the global flag:
global router.enabled |
org "" |
org "enabled" |
org "disabled" |
|---|---|---|---|
| true | route | route | off — no rewrite, no header |
| false | off | route¹ | off |
¹ Per-org opt-in routes even when the global flag is off, but only when router config is
present (a prefer table or an endpoint) — otherwise there is nothing to route with and
auto is left unchanged.
Configuration
conf/app.conf, or environment variables:
| Variable | Description |
|---|---|
HANZO_API_KEY |
Service token for internal IAM and KMS calls (billing, usage, auth) |
iamEndpoint |
IAM service URL |
clientId / clientSecret |
IAM OAuth client credentials for ai |
dataSourceName |
Database DSN — inject from KMS, never commit it |
KMS_CLIENT_ID / KMS_CLIENT_SECRET |
KMS Universal Auth credentials |
KMS_PROJECT_ID |
Default KMS project |
KMS_ENVIRONMENT |
KMS environment (default production) |
Secrets come from Hanzo KMS; the deploy workflow resolves
them at run time from a single HANZO_API_KEY GitHub secret, with Universal Auth as the
fallback.
airuns single-replica: its balance ledger is an in-pod invariant. Deploy withstrategy: Recreateandmin=max=1— a boot assertion panics ifCLOUD_REPLICAS > 1.LLM.mdhas the scale-out path.
Development
go build -race ./... # build
go test $(go list ./...) -tags skipCi # test (needs MySQL)
golangci-lint run # lint
cd web && yarn install && yarn start # the admin UI
LLM.md is the deep reference — architecture, the mount seam into cloud, and the
conventions that apply here. ZAP.md covers the transport.
Lineage
NOTICE. The routing, auth, billing and model surfaces are ours; the admin UI
and knowledge-base scaffolding started there.
License
Apache-2.0 — see LICENSE.
Hanzo — the open AI cloud. hanzo.ai · docs.hanzo.ai
SDKs: Python · TypeScript · Go · Rust · C++ · Swift · Kotlin · umbrella