Connecting Claude Code, Claude Desktop, Cursor and others

1. Create a token#

In the app, Profile → API tokens → New, or:

kelta token create --name agent --expires-in 90

For an agent that should only read, mint the token for a service user whose profile grants API_ACCESS and canRead on the relevant collections (recipe).

2. Add the server#

Claude Code

claude mcp add --transport http kelta-admin \
  https://api.example.com/acme/mcp/admin \
  --header "Authorization: Bearer klt_..."

Claude Desktop, Cursor and other JSON-configured clients

{
  "mcpServers": {
    "kelta-admin": {
      "type": "http",
      "url": "https://api.example.com/acme/mcp/admin",
      "headers": { "Authorization": "Bearer klt_..." }
    },
    "kelta-user": {
      "type": "http",
      "url": "https://api.example.com/acme/mcp/user",
      "headers": { "Authorization": "Bearer klt_..." }
    }
  }
}

Through the CLI (no token in the client config) — kelta mcp install claude-code|claude-desktop|cursor|generic prints a stdio configuration that uses your saved profile and adds the local-only tools (Local MCP bridge).

3. Choose admin, user or both#

Add …/mcp/admin to build and change metadata; …/mcp/user to work with records. Many agents want both.

4. Verify#

Ask the agent to call ping, then list_collections. A 401 means the token is not a klt_ PAT or has expired; a 403 on an admin tool means the token’s owner lacks the permission.

Tips for agents#

• Read kelta://docs/page-layouts, kelta://docs/list-views, kelta://docs/ui-pages, kelta://docs/ui-menus and kelta://docs/dashboards before authoring those objects by hand — they are the same documents as the authoring reference.
• Prefer the apply_* tools; re-running them is safe.
• execute_flow input is double-wrapped: { "input": { "input": { … } } } (why).
• Validation-rule formulas are error conditions — TRUE rejects.