> ## Documentation Index
> Fetch the complete documentation index at: https://docs.hiveku.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Hiveku Gateway
> Run Kimi K3 and GLM-5.3 inside the Claude desktop app through Hiveku's inference gateway
## What it is
The Hiveku Gateway (`gateway.hiveku.com/anthropic`) lets the Claude desktop app run non-Anthropic
models (Kimi K3 and GLM-5.3) while keeping the app you already know. You paste two fields
into the app's third-party inference settings, and the models appear in the picker under
their real names.
Your key authenticates you to Hiveku. The provider credentials live only on the gateway, so
no one hands out a Moonshot or OpenRouter key, and no provider ever sees yours.
**Access is by request.** Gateway keys are issued per person, not self-service. Email
**[support@hiveku.com](mailto:support@hiveku.com)** with the name of each person who needs one. Each key is individually
revocable, and usage is attributed to the person it was issued to.
## Models
| Shown in the picker | What it is | Cost |
| ------------------------ | -------------------------------------------------------- | ------- |
| **Kimi K3 (max) \$\$\$** | Moonshot's flagship. 1M context, reasons deeply. | highest |
| **GLM-5.3 (max) \$\$** | Z.ai's flagship at maximum reasoning effort. 1M context. | middle |
| **GLM-5.3-Flash \$** | Fast and cheap. 1.3M context. | lowest |
The `$` markers are relative, and the spread is wide: roughly 20x on input and 30x on output
between Flash and Kimi K3. Pick deliberately: Flash for routine work, Kimi K3 when a problem is genuinely hard.
Kimi K3 answers more slowly than the others. That is the model, not the gateway. It produces
many more reasoning tokens per reply, at a comparable tokens-per-second rate.
## Set it up
Email **[support@hiveku.com](mailto:support@hiveku.com)** with who needs access. You will receive a key beginning
`gwk_`. It is shown once and cannot be recovered, so store it somewhere safe.
In the Claude desktop app, open **Settings**, then **Configure third-party inference**.
| Field | Value |
| ------------------- | -------------------------------------- |
| Gateway base URL | `https://gateway.hiveku.com/anthropic` |
| Gateway auth scheme | `bearer` |
| Gateway API key | your `gwk_…` key |
| Model discovery | **on** |
| Model list | leave empty |
The base URL ends at `/anthropic` and has **no** `/v1` on the end -- the app adds
that itself, giving `https://gateway.hiveku.com/anthropic/v1/messages`. The
`/anthropic` names the protocol the gateway is speaking, leaving room for other
protocols later. The older base URL without it, `https://gateway.hiveku.com`, still
works and will keep working, so an existing setup does not need to be changed.
Type `tool` in the settings search box and switch on **Enable tool search**. This is not
optional with Hiveku's tool surface; see below.
Click **Apply Changes**, then fully quit and reopen the app. The model list is fetched at
launch, so a running session will not pick up the change.
## Turn on tool search
Do this before real work. Without it, sessions compact after a turn or two, replies take
minutes, and every message costs far more than it should.
Hiveku's MCP surface is large on purpose, at around **1,566 tools**. By default the app sends
every one of those schemas to the model on **every message**, before you have typed anything.
Tool search changes that: only tool *names* go up front, and a full schema is fetched the first time it is needed.
Measured on the same task and model, with and without it:
| | Tool search off | Tool search on |
| --------------- | --------------------- | -------------- |
| Largest request | 440,407 input tokens | **26,110** |
| Compaction | yes, over two minutes | none |
That is a **94% cut** on the heaviest request.
**No prompt or `CLAUDE.md` can substitute for this.** Tool schemas are placed in the request
before the model reads a word, so instructions to "only use the tools you need" have nothing
to act on. Sending fewer tools is the only lever.
The trade-off is one extra round trip the first time a given tool is used, in exchange for not
paying for the whole catalogue on every message.
## Auto mode
**Auto mode** (in the permission selector) stops the app asking for approval on every tool
call. It decides which actions need confirmation by asking a model, and that judgement call
goes to whichever model your top tier is set to.
If you see *"opus is temporarily unavailable, so auto mode cannot determine the safety of…"*,
the model is not down. It is a slow, deeply-reasoning model taking too long on what should be
a quick yes-or-no. Switching your session to a faster model, or turning Auto mode off, both
clear it.
Auto mode reduces prompting on live client accounts. Decide deliberately whether that is
what you want on real client data.
## Troubleshooting
**"Model discovery returned HTTP 401."**
The key is wrong, revoked, or expired. Check it, or email [support@hiveku.com](mailto:support@hiveku.com) for a new one.
**"Gateway /v1/models returned HTTP 404."**
The base URL has `/v1` on the end. It should be exactly
`https://gateway.hiveku.com/anthropic` -- the app appends `/v1/models` and `/v1/messages`
itself, so a base URL ending in `/v1` asks the gateway for `/v1/v1/models`.
A base URL of `https://gateway.hiveku.com` with no `/anthropic` is also fine.
Any other path (`/openai`, `/gemini`, anything else) is not served yet and will 404 with a
message naming the surface that does exist.
**"Rate limited · Retrying (4/10)" that never succeeds.**
The upstream provider account has run out of balance. Moonshot and OpenRouter both report an
exhausted balance with the same status code as a rate limit, so the client treats it as
temporary and retries — which can never help. Gateway builds from 2026-08-28 return a billing
error instead, naming the cause. Switch models in the picker to keep working, and email
**[support@hiveku.com](mailto:support@hiveku.com)**.
**Spend guards.**
Each key has a daily spend ceiling, and there is a second ceiling across all keys together.
Both are denominated in dollars rather than tokens, because the models differ by roughly 20x
in price: the same heavy session costs about $11 on Kimi K3, $5 on GLM-5.3 and \$0.56 on
GLM-5.3-Flash. If you hit a ceiling the message says which one and when it clears. Email
**[support@hiveku.com](mailto:support@hiveku.com)** to raise it.
**Long sessions and "Prompt is too long".**
Every message carries the whole conversation, so a long session eventually fills even a 1M
context. Compact the chat, or start a new one. If it happens within a few turns, the cause is
usually the tool catalogue rather than your conversation — see the
[plugin's tool search notes](/integrations/claude-code-plugin#turn-on-tool-search).
**"Models are still loading" and the picker is empty.**
Discovery runs at launch. Fully quit the app and reopen it.
**Sessions compact every turn or two, and replies take minutes.**
Tool search is off. See above. This is by far the most common cause of a slow, expensive
session.
**"This key has used … tokens today."**
A safety ceiling that only a runaway loop should reach. Check what is running; if it was
genuine work, email [support@hiveku.com](mailto:support@hiveku.com) and the ceiling can be raised.
**A model is missing from the picker.**
Only the three models above are offered, and the list is fixed at launch. If none appear,
discovery failed. Check the base URL and key.