README - Kollektiv - Your private LLM knowledgebase by alexander-zuev

Kollektiv MCP

🧠 Your personal LLM knowledgebase

Kollektiv MCP enables you to build personal LLM knowledge base in seconds and use it from your favorite editor / client. No more infrastructure setup, chunking, syncing - just upload your data and start chatting. Supports all major MCP clients out of the box - Cursor, Windsurf, Claude Desktop, etc.

🧪 Kollektiv is in early beta. If you experience any issues connecting to the MCP client, try going over these steps first. If still unsuccessful please raise an issue here

Why Kollektiv?

No need to re-upload data between chat sessions
Access from any client - Cursor, Windsurf, Claude Desktop, VSCode, PyCharm, etc.
No need for infrastructure setup - just upload your data and start chatting

💿 Connection

The simplest way to connect to Kollektiv MCP is to copy & paste the following configuration into your editor's mcp.json file. All clients (Cursor, Windsurf, Claude Desktop, VSCode, PyCharm) support this json format

{
  "mcpServers": {
    "kollektiv": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://mcp.thekollektiv.ai/mcp"
      ]
    }
  }
}

name:
- kollektiv- you can you give the server any descriptive name
command:
- npx - ensure you have node.js install before running this command
args:
- -y - this enables your shell to install mcp-remote which is currently required to connect to remote servers
- mcp-remote - this enables your client to connect to a remote MCP server (in this case Kollektiv)
- https://mcp.thekollektiv.ai/mcp - is the endpoint you are connecting to

Check out a short demo below or read client-specific instructions on how to connect.

Connection Demo

Cursor

Open Cursor and go to Cursor Settings > MCP > Add new global MCP Server. Paste the configuration above and save (ctrl/cmd+s).

Cursor Configuration

If configuration is successful and you haven't authenticated before, a browser window should open guiding you to the login page.

💡After saving the json it might take a while for Cursor to connect to the MCP. You might need to restart Cursor or give it a bit of time. If you see 'Client is closed' or other errors, taking these troubleshooting steps might help.

If the connection is successful, you should see Kollektiv MCP go green in the settings page:

Successful Cursor connection

Windsurf

Open Windsurf and go to Settings -> Windsurf Settings > MCP Servers > View raw config. Paste the configuration above and save (ctrl/cmd+s).

Windsurf MCP configuration

If configuration is successful and you haven't authenticated before, a browser window should open guiding you to the login page.

💡Windsurf, in contrast to other clients, in my experience requires a restart of the app to properly connect. If the server doesn't go 'green' after a while, try going over the troubleshooting steps below.

If connection is successful you should see Kollektiv MCP go green in the settings page:

Successful Windsurf configuration

Claude for Desktop

Open Claude Desktop and go to Settings -> Developer > Edit config. Open json file in any text / code editor, paste the configuration above and save (ctrl/cmd+s).

Claude Desktop Configuration

If configuration is successful and you haven't authenticated before, a browser window should open guiding you to the login page.

💡Claude for Desktop requires a restart of the app to properly connect. If the server doesn't go 'green' after a while, try going over the troubleshooting steps below.

If connection is successful you should see Kollektiv MCP go green in the settings page:

Successful Claude For Desktop

VS Code

Open VS Code and go to Settings -> MCP: Add server > Command (stdio):

command:
- npx -y mcp-remote https://mcp.thekollektiv.ai/mcp
name:
- give your server a descriptive name such as kollektiv

Your configuration settings.json should look similar to this:

{
  "chat.mcp.discovery.enabled": true,
  "chat.mcp.enabled": true,
  "mcp": {
    "servers": {
      "kollektiv": {
        "type": "stdio",
        "command": "npx",
        "args": [
          "-y",
          "mcp-remote",
          "https://mcp.thekollektiv.ai/mcp"
        ]
      }
    }
  }
}

VS Code Configuration

Next steps:

Click Start to connect to the MCP Server
- if you are not authenticated - you will be taken to the authentication page
Remember to add "chat.mcp.enabled": true, in your settings.json
Switch to Agent mode

💡VS Code requires you to manually start your server, add chat.mcp.enabled and switch to Agent mode to use MCP. If you do not see MCP tools in Agent mode, try going over the troubleshooting steps below.

If connection is successful you should see the tools exposed by Kollektiv MCP.

Successful VS Code Connection

Cline

Open Cline, click on MCP Servers > Edit Configuration and add the following configuration to your cline_mcp_settings.json:

{
  "mcpServers": {
    "kollektiv": {
      "timeout": 60,
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://mcp.thekollektiv.ai/mcp"
      ],
      "transportType": "stdio",
      "disabled": false
    }
  }
}

Note: direct connections to remote servers that support Authorization are not supported yet by Cline.

If connection is successful, you will be taken through to the authentication flow. After logging in, you should see Kollektiv MCP enabled in Cline.

Cline Configuration

Others (PyCharm, Claude Code)

Most MCP clients follow the same .json format and should work with similar configuration steps as previously mentioned clients:

Copy & paste configuration into your client's json configuration
Restart the app
Authenticate if not already
Kollektiv MCP should go green & be available in chat / agent mode

Success of your connection depends on many factors including but not limited to:

how strong developers of a particular client wanted to support MCP connections
whether the client supports the latest MCP spec with Oauth support

If you are experiencing issues, going through these simple troubleshooting steps might help.

Supported clients

I've validated connection works to the following MCP clients:

Cursor ✅
Windsurf ✅
Claude Desktop ✅
VS Code ✅
Cline ✅

Other MCP clients should be supported in theory, but in practice things might be a little different. If you have a client you really want to connect to - let me know!

🎮 Usage

Available Tools

/query_documents — Submit a question to the documents you’ve uploaded to Kollektiv and receive an answer based on the sources from your documents.
/list_documents — Return a list of your synced documents together with basic metadata.
Pro tip: Include the phrase “use Kollektiv MCP” so the client knows to call these tools.

Usage Tips

Always add "use Kollektiv MCP" — This tells the client which MCP server to use.
Wait for document to be Available — After upload, it takes 1–2 minutes before the document can be queried.
Rephrase queries when needed — If the client generates a poor query, edit or rewrite it yourself.

❓ Troubleshooting & Support

This MCP server uses Cloudflare Agents SDK as well as other libraries to provide the most modern way for users to connect to and use MCP servers. MCP clients on the other hand have yet to implement support for the 2 critical pieces:

remote MCP servers
MCP server authorization

In case you experience connection issues, please go through the following troubleshooting steps which should help you connect to the MCP server.

Support

If you require additional support please open a GitHub issue or reach out at

Connection Troubleshooting

If you are getting Invalid Authorization Request error as below or can not connect for another reason, try going through the steps below which should fix the issue.

Authorization Error

Ensure you're connecting to the correct endpoint:
- Use https://mcp.thekollektiv.ai/mcp as the MCP endpoint.
Clean mcp-remote cache:
- What this does:
  - Removes cache of mcp-remote library that is used to connect to the remote server from a client that doesn't support remote connections
- How:
  - Run the following command in your terminal

# MacOS
rm -rf ~/.mcp-auth  

# Windows
Remove-Item -Recurse -Force "$env:USERPROFILE\.mcp-auth"

Clear your browser data & cookies:
- What this does:
  - Removes browser cookies which are used to store authentication information when logging into Kollektiv.
- How:
  - Open your browser settings and delete browsing data for the last several hours

⚠️ Note: this will sign you out from all active sessions, including Kollektiv. Only do this if you're stuck in a broken login flow.

Restart your MCP client and try to reconnect to the MCP server:
- What this does:
  - MCP clients (Cursor, Windsurf, etc.) often cache connection / configuration settings from previous runs which might interfere with authentication.
- How:
  - Restart your editor / client
  - Try reconnecting to the MCP server

Using MCP Inspector

For debugging purposes, you can use the MCP Inspector to connect to Kollektiv MCP server.

npx @modelcontextprotocol/inspector

Select either SSE or Streamable HTTP transport

SSE: connect to the server at https://mcp.thekollektiv.ai/sse
Streamable HTTP: connect to the server at https://mcp.thekollektiv.ai/mcp

🛠️ Implementation Details (for the 🤓)

If you're just here for Kollektiv - skip this. This section is for devs and builders curious about how it works.

Kollektiv MCP is part of a modular system enabling users to set up RAG over their data in seconds — without the need to manage infrastructure, pipelines, or model configs.

It consists of three independently deployed services:

MCP Server (Cloudflare Worker)
https://mcp.thekollektiv.ai
Acts as a secure gateway for clients to interact with indexed data via the Model Context Protocol. Supports OAuth.
Frontend (React + Vite Worker)
https://thekollektiv.ai
A clean, minimal user interface for uploading and managing their content.
Backend (FastAPI)
https://api.thekollektiv.ai
Handles source ingestion, validation, and orchestration of a RAG pipeline.

🔐 Security

Kollektiv MCP implements several security measures:

Sign-in happens via the standard OAuth 2.1 “Authorization Code” flow powered by Supabase; only short-lived, HttpOnly, Secure cookies are stored—no passwords ever touch this server.
All traffic is served exclusively over HTTPS through Cloudflare’s edge, and every sensitive POST request carries a one-time CSRF/transaction token.
The backend runs inside the Cloudflare Workers sandbox (no local file-system, no long-running processes), drastically reducing the attack surface.

For detailed disclosure guidelines see .

🪪 License

Released under the Apache License 2.0 — commercial support or alternative licensing: