Help

How to connect the MCP servers of this platform to your AI tools.

Authentication

Most MCP servers here require authentication. There are two ways:

OAuth 2.0 (recommended)

Clients like ChatGPT or Claude support OAuth 2.0 natively. When you add an MCP server URL, the client sends you to sign in and authorize access — no token to copy.

Bearer token

For clients without OAuth, put a token in the Authorization header. There are two kinds:

  • Personal access token — access to every server you may use. Create one in Personal Access Tokens.
  • Server token — bound to a single server, ideal for integrations. Create one in Server Tokens.

The header: Authorization: Bearer YOUR_TOKEN

Server URL

Every server's URL is on the MCP Servers page — copy it from the server's detail. It looks like https://dev.ai.punktero.com/mcp/server-path.

Client setup

ChatGPT supports remote MCP servers with OAuth natively.

  1. Enable Developer mode in Settings → Apps → Advanced settings.
  2. Go to Settings → Apps → Create app.
  3. Paste the MCP server URL (e.g. https://dev.ai.punktero.com/mcp/server-path).
  4. ChatGPT handles the authorization — you are redirected to sign in.
  5. After authorizing, the server's tools are available in your conversations.

Remote MCP servers in ChatGPT need a Plus, Team or Enterprise plan with Developer mode on.

Claude.ai supports remote MCP servers with OAuth.

  1. Open claude.ai → Settings → Connectors.
  2. Click Add custom connector and paste the MCP server URL.
  3. Claude redirects you to authorize.
  4. Once connected, the server's tools appear in your conversations.

Remote MCP integrations on Claude.ai need a Pro or Team plan.

Claude Desktop takes remote MCP servers in its configuration file.

  1. Open Claude Desktop → Settings → Developer → Edit Config.
  2. Edit claude_desktop_config.json — on macOS ~/Library/Application Support/Claude/claude_desktop_config.json, on Windows %APPDATA%\Claude\claude_desktop_config.json.
{
  "mcpServers": {
    "server-name": {
      "url": "https://dev.ai.punktero.com/mcp/server-path",
      "headers": {
        "Authorization": "Bearer YOUR_TOKEN"
      }
    }
  }
}

Older versions do not support remote HTTP servers directly. If the config above does not work, use the mcp-remote bridge through npx instead (needs Node.js):

{
  "mcpServers": {
    "server-name-npx": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://dev.ai.punktero.com/mcp/server-path",
        "--header",
        "Authorization: Bearer YOUR_TOKEN"
      ]
    }
  }
}

After saving, quit Claude Desktop completely and start it again — closing the window is not enough, it keeps running in the tray / menu bar. Right-click the tray icon and choose Quit (or Cmd + Q on macOS).

Add a server with claude mcp add:

claude mcp add server-name \
  --transport http \
  https://dev.ai.punktero.com/mcp/server-path \
  --header "Authorization: Bearer YOUR_TOKEN"

Or put it in .mcp.json in your project root:

{
  "mcpServers": {
    "server-name": {
      "type": "http",
      "url": "https://dev.ai.punktero.com/mcp/server-path",
      "headers": {
        "Authorization": "Bearer YOUR_TOKEN"
      }
    }
  }
}
  1. Open Cursor → Settings → MCP.
  2. Click Add new MCP server.
  3. Choose type HTTP and enter the server URL.

Or put it in .cursor/mcp.json in your project:

{
  "mcpServers": {
    "server-name": {
      "url": "https://dev.ai.punktero.com/mcp/server-path",
      "headers": {
        "Authorization": "Bearer YOUR_TOKEN"
      }
    }
  }
}
  1. Open Windsurf → Settings and search for MCP.
  2. Click Add Server, or edit ~/.codeium/windsurf/mcp_config.json:
{
  "mcpServers": {
    "server-name": {
      "serverUrl": "https://dev.ai.punktero.com/mcp/server-path",
      "headers": {
        "Authorization": "Bearer YOUR_TOKEN"
      }
    }
  }
}

VS Code uses MCP servers in GitHub Copilot's agent mode.

  1. Install the GitHub Copilot extension.
  2. Create .vscode/mcp.json in your project:
{
  "servers": {
    "server-name": {
      "type": "http",
      "url": "https://dev.ai.punktero.com/mcp/server-path",
      "headers": {
        "Authorization": "Bearer YOUR_TOKEN"
      }
    }
  }
}

Or add it to your VS Code settings.json under the "mcp" key.

OpenCode takes remote MCP servers in its JSON configuration.

  1. Open opencode.json in your project root (or ~/.config/opencode/opencode.json for every project).
  2. Add the server under the "mcp" key:
{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "server-name": {
      "type": "remote",
      "url": "https://dev.ai.punktero.com/mcp/server-path?separator=_",
      "enabled": true,
      "headers": {
        "Authorization": "Bearer YOUR_TOKEN"
      }
    }
  }
}

?separator=_ makes every tool name use underscores — some models used by OpenCode reject hyphens. Use ?separator=- for hyphens, or leave it out to keep the original names.

Needs a recent OpenCode — versions before July 2025 only spoke SSE to remote servers, which these servers do not provide. If the connection fails, update OpenCode or bridge through mcp-remote (see Claude Desktop).

OpenWebUI takes MCP servers for everyone (admin) or per user.

For all users (admin)

  1. Open Admin Panel → Settings → External Tools.
  2. Click Add Connection, then OpenAPI (it switches to MCP Streamable HTTP).
  3. Fill in URL, ID and Name.
  4. For a token, enter Bearer YOUR_TOKEN in API Key.
  5. For OAuth, switch from Bearer to OAuth 2.1 and click Register Client.

For yourself

  1. Click your profile icon (bottom left) → Settings → External Tools.
  2. Click Add Connection.
  3. Fill in API Base URL with the OpenAPI URL of the server.
  4. Fill in API Key with a server token (preferred) or a personal access token.

Admin tools are available to every user of the instance, personal ones only to you. Personal connections work only after the admin enables User MCP Connections in Admin Panel → Settings → External Tools.

Any MCP client can connect. The servers use the Streamable HTTP transport. The client needs:

  1. Server URL — e.g. https://dev.ai.punktero.com/mcp/server-path
  2. Transport — Streamable HTTP (sometimes listed as http or streamable-http)
  3. Authentication — the Authorization: Bearer YOUR_TOKEN header, or OAuth 2.0 if the client supports it

If the client (or its model) rejects hyphens or underscores in tool names, append ?separator=_ or ?separator=- to the server URL — every tool name then uses only that separator.

Known MCP clients and what they support: the official MCP client list.

Memory

Memory is a persistent memory for your AI assistants, shared across every client and machine. Add the server to all your clients and the assistant stores facts worth keeping (preferences, decisions, gotchas, project facts) and recalls them next time. Everything it remembers is visible and editable under Memory, with a full change history you can revert. It does not replace your assistant's own memory — it mirrors it, so every client and machine shares the same facts. You can export your memory as JSON and import it elsewhere (e.g. from a dev instance to production).

https://dev.ai.punktero.com/mcp/memory

Use a server token for Claude Code and other clients without OAuth; claude.ai connects with OAuth as usual.

Projects

Each memory can belong to a project — a stable name independent of machine and path. For a git repository the assistant uses the repository name from its remote URL (e.g. mcp-server), so every checkout on every machine shares one memory. If it picks a wrong name, tell it which project to use. Memories without a project (your preferences) are returned everywhere.

Forced labels (optional)

A per-project client configuration can force labels on everything it stores by appending ?labels=project:… to the URL. Claude Code (local scope, not committed):

claude mcp add memory --scope local --transport http \
  "https://dev.ai.punktero.com/mcp/memory?labels=project:mcp-server" \
  --header "Authorization: Bearer srvt_YOUR_TOKEN"

Tell the assistant to use it

Clients rarely call memory tools on their own. Add a short instruction to your CLAUDE.md, claude.ai preferences or Cursor rules:

At the start of a conversation call memory `recall` (no arguments), then `recall` with the project you are working on.
Whenever you learn a preference, decision, gotcha or stable project fact, call `remember` with `project` set — do not wait to be asked.
Keep using your native memory as usual and mirror everything you save there into memory `remember` — this server is the shared copy, not a replacement.

MCP Inspector

MCP Inspector is a debugging tool for trying MCP servers interactively. The MCP Servers list has its command ready — a server's menu → Copy Inspector command.

  1. You need Node.js 18+ (it includes npx) — nodejs.org.
  2. Copy the command from the server's menu and run it in a terminal.
  3. The Inspector opens in your browser.
  4. For a server that needs authentication, use a token from Personal Access Tokens.

The command already carries --server-url and --transport http for the server. Official Inspector documentation.