Kiteworks Agent Marketplace
Contact
Quel assistant IA utilisez-vous ?

Installer les agents Kiteworks dans Claude

Choisissez où vous utilisez Claude. Chaque chemin ajoute le même marketplace en lecture seule — seul l'hôte change.

What this is

Every agent in this marketplace talks to your files through the official Kiteworks MCP connector — a small, Kiteworks-provided connector that Claude signs in to. The agents are the plugins you install; the connector is the secure pipe they use to reach Kiteworks. They are two different things.

This page covers how the connector gets connected, and how to add it yourself. Pick the path that matches you:

  • Most people: do nothing here — the agent connects it for you on first use. See *Easiest path*.
  • Prefer to connect it first, in the app: add it from Claude's Connectors directory. See *Add it yourself (Claude web / Desktop)*.
  • Admins / Claude Code / power users: run the connector yourself. See *Add it yourself (terminal and Claude Code)*.

Easiest path — the agent connects it for you

You usually do not set anything up by hand. When you install an agent and first ask it for Kiteworks data, Claude prompts you to sign in through the official Kiteworks MCP connector. After that first sign-in the connection is reused for later sessions. That's it.

Check whether it's already there: open Claude web or Desktop, go to Settings, Connectors, and look under "Your connectors" for Kiteworks. Present and enabled means you're set.

Add it yourself (Claude web / Desktop)

Best when you want the connector in place before installing an agent, using the built-in directory — no downloads, no terminal.

  1. Go to claude.ai/settings/connectors (or Settings, Connectors in Claude Desktop).
  2. Click Browse directory and search "Kiteworks."
  3. Click Add, then sign in with your Kiteworks tenant when prompted.
  4. Confirm Kiteworks now appears under "Your connectors."

That's all — the agents pick it up automatically. The manual setup below is only for Claude Code or self-run installs.

Add it yourself (terminal and Claude Code)

Best when you run Claude Code, are an admin scripting setup, or want to run the connector binary directly.

First, get the kiteworks-mcp connector from the official repository at github.com/kiteworks/mcp (releases). Version 0.9.2 or newer is required. Confirm it is on your PATH:

kiteworks-mcp --version

Sign in once. The connector manages its own login and stores the token in your operating system's keychain — you never paste a token into a config file or into chat:

kiteworks-mcp login https://content.kiteworks.com

Use your own tenant host if it is not content.kiteworks.com.

Wire it into Claude Desktop (manual config)

Open Settings, Developer, Edit Config, and add a kiteworks entry:

{
  "mcpServers": {
    "kiteworks": {
      "command": "kiteworks-mcp",
      "args": ["start", "https://content.kiteworks.com"]
    }
  }
}

Restart Claude Desktop. The Kiteworks tools now appear in the connector list.

Wire it into Claude Code

From a terminal:

claude mcp add kiteworks -- kiteworks-mcp start https://content.kiteworks.com

Verify with claude mcp list — you should see kiteworks connected. The agents find the connector by the word "Kiteworks" in its server name, in any capitalisation: kiteworks, Kiteworks and KITEWORKS-prod all work. In Claude Desktop, where connectors are named by an internal ID rather than by you, the add-ons recognise Kiteworks by its tool set instead. If you named it something else — kw, say — the agents will report that no Kiteworks connector is available. Add it again under a name that contains "kiteworks".

What each connection can do

Not every way of connecting Kiteworks gives the add-ons the same reach. What matters is where the connector runs.

  • The local kiteworks-mcp server (the terminal path on this page, running on your computer over stdio) can download files to and upload files from this computer. That is what lets an add-on read the contents of PDF, Word and Excel files, and save a branded PDF report.
  • The Claude connector-directory entry, and any remote (HTTPS) Kiteworks MCP server, cannot touch files on this computer — by design. Over those connections an add-on reads text files (csv, txt, json, xml, md, log) directly, lists binary files by name, size, owner and sharing only (and marks them "not content-checked"), and saves reports as CSV and text rather than PDF. The add-on tells you at the moment it matters. No admin setting adds file download or upload to a remote server; if you need PDF reports or binary content checks, run the local server.
  • Tools can also be hidden by the OAuth scopes your Kiteworks admin grants the MCP client, by the server's KW_MCP_ENABLED_TOOLS allowlist, or by Claude Desktop's per-tool permissions. An add-on that finds a tool missing says which capability it lost.

Claude Code version. The add-ons rely on plugin hooks that Claude Code runs before each connector call; that behaviour was verified on Claude Code 2.1.258. Use that version or newer.

Tool safety — what the connector can do

The connector exposes a read, browse, search and identity tool set by default. That is what most add-ons use: they browse and search your content and report on it. Some go one step further, always with a clear boundary:

  • Add-ons that save a report into Kiteworks also create a folder and files (create_folder, create_file_from_content, upload_file_from_path). They write only their own report; they do not change your existing files.
  • Preview / apply pairs. Where an add-on can rearrange your content — the duplicate finder, inbox triage, invoice organizer, naming cleanup and offboarding content finder — it ships as two halves. The *preview* half never writes anything. The *apply* half moves or renames files and folders (move_file, move_folder, rename_file, rename_folder) only after you have reviewed the preview and confirmed. It never deletes.
  • No add-on uses the delete tools. Nothing in this marketplace calls delete_file or delete_folder.
  • Destructive tools are off by default. Move and delete tools (move_file, move_folder, delete_file, delete_folder) only turn on if you add the -enable-destructive-tools flag when starting the connector. If an *apply* half reports that a move tool is missing, this flag is why; nothing here needs delete. Do not enable the flag unless you have your own reason to — an AI client with those tools on can delete your files with no undo. If you do enable it, the flag must come before the URL or the connector silently ignores it.
  • Some tools move file content. read_file_contents pulls file text into the model's context and download_file_to_path writes file bytes to your disk. Add-ons that check file contents use them: text files through read_file_contents, binary files (PDF, Word, Excel) through download_file_to_path on the local server, parsed on your machine. Each add-on's listing says whether it reads file bodies or metadata only.

Security notes

  • No pasted secrets. Sign-in is handled by the connector's own login (token in the OS keychain) or by Claude's OAuth prompt. Never put a token or client secret in a config file or paste one into chat.
  • Best-effort. These agents report on Kiteworks content; they are productivity-grade, not the audited Kiteworks compliance runtime, and not a basis for a deletion or compliance decision on their own. Where an add-on moves or renames files, it does so only after you confirm its preview. See What to expect.