Skip to main content

​
What it is

The Hiveku plugin turns Claude Code (the Claude desktop app or the claude terminal command) into a full operating console for your Hiveku accounts: 18 department skills (SEO, PPC, sales, helpdesk, books, commerce, communications, and more), 9 dispatchable analyst agents, and about 70 slash commands, on top of Hiveku’s complete MCP tool surface. Its core idea is one folder = one account. You bind each working folder to one Hiveku account, and every tool call in that folder goes to that account and no other. An agency managing many clients works out of many folders, and a folder with no binding is inert — the plugin stays out of the way in projects that have nothing to do with Hiveku.
This is the recommended way to use Hiveku from Claude. The generic MCP connector (see LLM Connectors) still works for claude.ai, ChatGPT, and Claude Desktop, but the plugin adds the skills, agents, commands, and per-folder account safety that the raw connector does not have.

​
Install

1

Open Plugins

Open Settings (the gear, or Cmd+, / Ctrl+,), then under Customize click Plugins.
2

Add the marketplace

Click Add at the top right, choose Add from a repository, paste this and confirm:
3

Install the plugin

Adding the repository only registers the catalog — it installs nothing by itself. Find the hiveku plugin in the list, click it, then click Install. If you are asked for a scope, pick the one covering all projects (user scope), so Hiveku follows you into every folder.
4

Turn on auto-update — do not skip this

Still in Plugins, open the hiveku marketplace entry and choose Enable auto-update. Claude Code auto-updates Anthropic’s own marketplaces but starts third-party ones like Hiveku’s switched off — leave it off and you keep the exact version you installed forever, with nothing telling you that you are behind.
You will see a red warning saying the plugin can access everything on your computer. Choose Continue. Claude Code shows that banner for every plugin that ships a local MCP server, including Anthropic’s own, because a local server runs as a normal process with your permissions. The server listed as hk is the Hiveku one: no third-party dependencies, connects only to core.hiveku.com, and reads a single account’s key based on the folder you are working in.

​
Connect your accounts

In any Claude session:
Your browser opens the Hiveku consent page. Sign in, tick the accounts you want to work with, and mark any of them read-only if you want a look-but-do-not-touch connection. No key is ever pasted or typed: keys come back over a one-time local callback and are stored encrypted on your machine. Then create a folder for each account:
That makes one folder per connected account under ~/Hiveku-Accounts, each already bound. Prefer your own folder layout? Skip setup, open a folder you already use, and run /hiveku:bind there instead.
On a machine with no browser (an SSH session, for example), create a key at app.hiveku.com under Settings → LLM Connectors and run hiveku accounts add <key>.
Do this once, 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 — around 1,566 tools. By default a client sends every one of those schemas to the model on every single message, before you have typed anything. Tool search fixes it. Only tool names go into context up front, and Claude fetches a tool’s full schema the first time it actually needs one. Measured on the same task, same model, with and without it: That is a 94% cut on the heaviest request. The cache figures matter as much as the headline: cached input is billed at a fraction of fresh input, so a turn that cost dollars costs cents.
Running Kimi or GLM through the Hiveku Gateway instead of Claude models? Tool search matters even more there, and gateway access is by request — email support@hiveku.com.
1

Open the settings

Open Settings, then the third-party inference configuration.
2

Find the toggle

Type tool in the settings search box and turn on Enable tool search.
3

Apply, then start a new conversation

Click Apply Changes. The tool list is fixed when a session connects, so an existing conversation will not pick it up — start a new one.
On the Hiveku Gateway, this toggle is not enough on its own. Claude Code treats any host other than api.anthropic.com as non-first-party and turns tool search off, logging [ToolSearch:optimistic] disabled: … is not a first-party Anthropic host. Measured on a gateway session with the toggle on: 447,400 tokens of tool definitions, 224% of the context window, before the first message.Plugin 0.8.0 and later fixes this for you — the plugin ships the full catalogue on disk and advertises a core set plus a hiveku_find_tools search tool, so a session starts at a few thousand tokens instead of hundreds of thousands. Every tool stays findable and callable. Update the plugin and start a new conversation.
Tool search defers a tool’s SCHEMA, not its description. Names and descriptions still ship for every tool on every message, and Hiveku’s descriptions average around 200 characters. That is why the toggle alone does not get a 1,500-tool catalogue down to nothing, and why the plugin’s local index exists.
No prompt or CLAUDE.md can fix this. The 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. Advertising fewer tools is the only lever, and this is the supported way to do it.
The trade-off is one extra round trip the first time Claude uses a given tool, in exchange for not paying for the whole catalogue on every message. The Model Context Protocol docs recommend switching to on-demand loading once tool definitions pass 1–5% of the context window; Hiveku’s catalogue is well beyond that.

​
Daily use

Open one of your account folders in Claude (the folder picker in the desktop app, or cd in a terminal) and that client’s account is live — the session banner tells you which one. From there:
  • /hiveku:brief — get oriented on the account before doing any work
  • /hiveku:daily — the morning operating brief: what changed, what needs attention
  • /hiveku:integrations — audit every integration: what is connected, what is broken
  • /hiveku:code / /hiveku:deploy — pull a website project local, edit, verify, ship
  • Or just ask in plain language — the department skills load themselves when the work touches their territory, and Claude can dispatch read-only analyst agents (PPC, SEO, revenue, support, books, commerce, tracking) to investigate while you keep working.
Run /hiveku:status any time to see which account the folder is bound to.

​
Update to a new version

Two commands, in this order. The usual path: click the terminal icon in the Claude app — it opens a shell in whatever folder you have open, and the folder does not matter — then run:
Two commands because they do different jobs: the first refreshes the catalog so your machine knows a newer version exists; the second actually installs it. Then restart the session (or run /reload-plugins inside it). Plugins load at session start, so a session that is already open keeps the version it started with. Newer plugin versions tell you at session start when the session you are in is running an older version than what is installed.
The catalog-refresh step matters more than it looks. The desktop app refreshes marketplaces in the background, so right after a release the Settings → Plugins screen can still show the old version with no Update button. claude plugin marketplace update hiveku forces the refresh and the Update button appears.
If you enabled auto-update at install time, none of this is needed — Claude Code picks up new versions on its own shortly after a session starts. Inside a running Claude session, the same two steps are slash commands: /plugin marketplace update hiveku then /plugin update hiveku@hiveku. The slash form works only inside a session — in a plain terminal, use the claude plugin form above.

​
Permission modes

Auto mode asks a classifier model whether each new action is safe. That is one extra round trip per never-before-approved tool, and on a large catalogue it adds up. Plugin 0.8.0 and later removes most of it: Hiveku’s read-only tools are pre-approved by the plugin itself, so listing, fetching and reporting never wait on a permission check. Writes, deletes and anything outward-facing still go through the normal flow. For a working session where you are watching the output — a migration, an audit, a cleanup like a GTM swap — Bypass permissions is the mode to use. Set it per chat from the mode selector next to the message box, or press 5 in that menu.
Bypass permissions accepts everything, including writes, deletes and sends. Use it while you are actually reading what Claude is doing, on one focused task. It is the wrong default for a team of people working across many accounts — drop back to Auto when the task is done.
If the mode is greyed out, an administrator has to enable Allow bypass permissions mode in the app’s settings first.

​
Multi-agent workflows

Large jobs — auditing every tag in a container, sweeping an account, comparing a mirror against a live site — go much faster when Claude fans the work out across several agents at once instead of doing it in sequence. Turn it on with Dynamic workflow size in /config. You can also trigger it per-message by including the word ultracode, if the keyword trigger is enabled.
Workflows spawn many agents and spend proportionally more. That is usually the right trade for an audit you would otherwise do by hand, and the wrong one for a quick question. The first time you run one, the app asks you to accept a usage warning.

​
Seeing what a session costs

The Usage page shows token counts by default. Turning on Show estimated cost prices them at Anthropic list price.
On the Hiveku Gateway that estimate is blank until you configure rates. The gateway reports the real upstream model — kimi-k3, z-ai/glm-5.3, z-ai/glm-5.3-flash — and Claude Code cannot map those to a Claude model, so they are left out of the estimate entirely rather than guessed at.
Give it the real rates with inferenceModelPricing. All four fields are required, in USD per million tokens:
A gateway alias matches that exact ID only, so the spellings above must be used verbatim. cacheWritePerMtok covers both 5-minute and 1-hour cache writes.
Those are list prices at the time of writing and providers change them — treat the estimate as a guide, not a bill. GLM-5.3-Flash is priced here at its post-promotion rate so the estimate errs high rather than low.
The spread is worth knowing before you pick a default model: the same heavy session costs roughly 11onKimiK3,11 on Kimi K3, 5 on GLM-5.3 and $0.56 on GLM-5.3-Flash. The gateway enforces its own per-key and fleet-wide daily ceilings in dollars regardless of this setting.

​
Troubleshooting

“I do not see the plugin after adding the repository.” Adding a marketplace only registers the catalog. Open the hiveku entry and click Install. “Tools are missing in a folder.” That folder is not bound to an account. Run /hiveku:bind there, or /hiveku:status to see what Claude thinks is going on. “I get 401 errors from Hiveku tools.” The key for that account was revoked or rotated. Run /hiveku:connect again to re-mint it. “The commands do not show up.” Run /reload-plugins. If they still do not appear, clear the plugin cache and reinstall: rm -rf ~/.claude/plugins/cache, restart Claude Code, install again. “A new version seems not to exist.” Run claude plugin marketplace update hiveku in a terminal — the catalog on your machine is stale. See the update section above. “Sessions compact every turn or two, and replies take minutes.” Tool search is off. Every message is carrying the whole tool catalogue — around 370,000 tokens before you type anything. Turn on Enable tool search (see Turn on tool search) and start a new conversation. This is by far the most common cause of a slow, expensive-feeling session. “Claude says a tool exists but cannot call it.” With tool search on, Claude loads a schema the first time it needs one, so an unfamiliar tool costs one extra round trip. If a call genuinely fails, check the folder is bound (/hiveku:status) — an unbound folder has no tools at all. “Failed to authenticate. API Error: Blocked” — and signing in again does not help. This is not an authentication problem. Blocked is a value in the desktop app’s Built-in tool policy (allow / ask / ask once per session / blocked), and the client reports any local refusal as an authentication failure. The tell is timing: the error appears in milliseconds with no tokens used, because no request was ever sent. Check Settings → Built-in tool policy for an entry set to Blocked, including any policy arriving from org-managed settings. If nothing is set, start a new conversation. “A shell command cannot reach Hiveku, or a script gets deny network-outbound.” Sandboxed sessions ship with a fixed allowlist — npm, PyPI, GitHub, Ubuntu, crates, Playwright and Anthropic’s own hosts — and no Hiveku entry. MCP tools still work, because that endpoint is permitted separately, but anything shell-level (a script calling the API, or curl) is denied. Add *.hiveku.com to the egress host settings. Entries are host or *.host, with no scheme and no path. “Everything says Rate limited and retrying, over and over.” On the Hiveku Gateway, a provider account with no balance left answers with the same status code as a rate limit, so the client retries a billing problem ten times. Gateway builds from 2026-08-28 return a clear billing error instead. If you see it, switch models in the picker and email support@hiveku.com. “Can I have Claude call every tool to check they all work?” No — and it will break the session if you try. Around 1,500 tools means a permission decision and a transcript entry each, and the transcript is re-sent on every later message, so the run fills the context and dies partway through with “Prompt is too long”. Use the sweep script instead, which does the whole thing in one process and writes a report:
It only calls tools the server declares read-only, so it cannot create, update, delete or send anything. “I also run the Hiveku VS Code extension.” That is fine; both can be live at once. The session-start banner names which account each connection serves so you never act on the wrong tenant.

​
Remove it

To also revoke the account keys server-side (recommended if you are handing the machine on), run hiveku accounts remove <account-id> first — it revokes the key with Hiveku and only then forgets it locally, so a failed revoke never leaves a live key you can no longer see.