Connect

MCP clients

Kognite is a standard MCP server over the Streamable HTTP transport. Below is the exact config for each popular client. You don’t need an account before you start: add the server, and your client opens a browser tab where you sign in and approve it.

One URL, no key. https://mcp.kognite.dev/mcp uses the MCP authorization flow (OAuth 2.1 with PKCE and dynamic client registration). Your client gets a short-lived token and refreshes it on its own. Revoke it later under API Keys → Connected applications.

Client can’t do the browser handshake? Use https://mcp.kognite.dev/t/YOUR_KEY instead, pasting a token from your dashboard’s Connect page in place of YOUR_KEY. That URL carries the token in the path, so it both identifies and authenticates you — keep it secret and rotate it if it leaks. A kgn_ key also works as an Authorization: Bearer header or a ?token= query parameter on https://mcp.kognite.dev/mcp.

Cursor

.cursor/mcp.json

{
  "mcpServers": {
    "Kognite": { "url": "https://mcp.kognite.dev/mcp", "type": "streamable-http" }
  }
}

Settings → MCP → “Add new global MCP server”, or drop the file at .cursor/mcp.json in your project.
Toggle the Kognite server ON — you should see a green dot and its tools listed.
Open Composer in Agent mode; the memory_* tools are now available.

Heads-up. Cursor uses tools automatically in Agent mode. To force memory use, add a rule at .cursor/rules/kognite.md telling it to persist durable facts with memory_add and recall with memory_search.

Claude Code

terminal

claude mcp add --transport http Kognite https://mcp.kognite.dev/mcp

Run the command above in any project (add --scope user to make it global across projects).
Run `claude mcp list` to confirm Kognite is connected.
Start Claude Code — the memory_* tools are available to the agent immediately.

Heads-up. Add a line to CLAUDE.md so recall is proactive: “Before answering anything that depends on my history, call memory_search first; persist durable facts with memory_add.”

Claude Desktop

Settings → Developer → Edit Config

claude_desktop_config.json

{
  "mcpServers": {
    "Kognite": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.kognite.dev/mcp"]
    }
  }
}

Easiest (paid plans): Settings → Connectors → “Add custom connector”, paste the connection URL, enable it.
Config-file path: Settings → Developer → Edit Config, add the block above at the top level, then fully quit and reopen.
The mcp-remote bridge requires Node.js installed (it runs via npx).

Heads-up. Claude Desktop has its own built-in memory that can intercept “remember …”. Turn it off under Settings → Capabilities, or name the tool explicitly: “use the Kognite memory_add tool”. A bare {"url": …} is rejected — remote servers must use the mcp-remote bridge shown above.

VS Code (GitHub Copilot)

.vscode/mcp.json

{
  "servers": {
    "kognite": { "type": "http", "url": "https://mcp.kognite.dev/mcp" }
  }
}

Add the file, then Ctrl/Cmd+Shift+P → “MCP: List Servers” → kognite → Start Server.
Open Copilot Chat → Agent mode → click the 🛠 tools icon and make sure Kognite's memory_* tools are checked.
If you run many servers, uncheck unused ones — Copilot caps around 128 tools and trims the rest.

Heads-up. Copilot won't always pick the tool itself. Name it, or add .github/copilot-instructions.md telling it to call memory_add on durable facts.

ChatGPT

Settings → Connectors

server URL

https://mcp.kognite.dev/mcp

ChatGPT connects to remote MCP servers as a Connector (paid plans; you may need to enable Connectors / developer mode first).
Settings → Connectors → “Add custom connector”, paste the URL as the server URL, then approve Kognite in the browser tab it opens.
Enable the Kognite connector in a chat from the tools / Connectors menu.

Heads-up. Ask ChatGPT to “use the Kognite connector to remember / recall …” so it calls the tools rather than its own memory.

Windsurf

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "Kognite": { "serverUrl": "https://mcp.kognite.dev/mcp" }
  }
}

Settings → Cascade → MCP Servers → Add server, or edit the config file directly.
Paste the block, save, then click Refresh in the MCP panel.
Open Cascade in Agent mode — Kognite's tools appear in the list.

Heads-up. Add a .windsurfrules line telling Cascade to use Kognite's memory tools instead of writing local files.

Zed

~/.config/zed/settings.json

settings.json

{
  "context_servers": {
    "Kognite": { "source": "custom", "url": "https://mcp.kognite.dev/mcp" }
  }
}

Open Zed settings (Cmd/Ctrl+, → open settings.json).
Add the context_servers block and save.
Reopen the Assistant panel — Kognite's tools are listed (HTTP MCP needs a recent Zed build).

JetBrains AI

Settings → Tools → AI Assistant → MCP

server URL

https://mcp.kognite.dev/mcp

Settings → Tools → AI Assistant → Model Context Protocol (MCP).
Add → choose the HTTP/SSE type → paste the URL → Apply.
In the AI Assistant chat, enable the Kognite tools for the conversation.

Any other MCP client

Kognite needs no CLI or npm package. Point any Streamable-HTTP MCP client at the connection URL; most accept {"mcpServers":{"Kognite":{"url":"<url>"}}}. Clients that support the MCP authorization flow will discover it automatically — an unauthenticated request answers with theWWW-Authenticate challenge and the protected-resource metadata. The full tool surface is in the MCP tools reference.