Skip to main content

Connecting MCP clients

Kandra ships a dynamic, meta-tools MCP (Model Context Protocol) server — one server that discovers every Document/Dictionary/Report/Data Processor/Constant/Enum in your configuration at runtime, instead of shipping one hand-written tool per entity type. Point an AI desktop client at it and the assistant can list what entities exist, describe their fields, query, create, update, submit, delete and run reports — under a real user's identity and permissions, exactly as if that user had signed in.

This page covers connecting three clients: Claude Code, VS Code / GitHub Copilot Chat, and Codex. Every example below is a real, verified config — not a guessed shape.

Prerequisites​

  • Your Kandra host running, with the MCP endpoint mapped (AddKandraMcp<>() + app.MapKandraMcp() in Program.cs — see the engine's own setup if you're wiring this into a fresh configuration; it ships out of the box in this template).
  • An API key. The MCP endpoint accepts API-key authentication only — a JWT (a signed-in human session) is rejected with 403 Forbidden, on purpose, so an MCP client is always a distinctly-identifiable non-human caller. Create one for yourself from My profile → API keys (or, for another user, from their user edit page) — see Identity & auth for the full mechanism (its "API keys" section). The plaintext key (kdr_...) is shown once, at creation — copy it before closing the dialog.
Prefer HTTP over HTTPS

Use the plain HTTP port in every example below (http://localhost:5219/mcp for the default dev launch profile), not the HTTPS one. The dev HTTPS listener uses a self-signed certificate most MCP clients won't trust, and you'll get an opaque TLS handshake failure instead of a clear error. Use a real certificate before pointing a client at HTTPS in anything other than local dev.

Claude Code​

claude mcp add --transport http kandra http://localhost:5219/mcp --header "X-Api-Key: kdr_..." -s local
  • --transport http — the MCP server is a stateless HTTP endpoint, not stdio.
  • --header "X-Api-Key: kdr_..." — the same header every other Kandra API-key call uses.
  • -s local/-s user — scope the registration so the key never lands in a committed .mcp.json. -s local (the default) stores it in your own per-project settings, outside version control; -s user stores it once for every project on your machine. Only use the project-shared scope (-s project, which does write .mcp.json into the repo) for something that genuinely has no secret in it — never for this.

Run claude mcp list to confirm it registered, then ask Claude something that needs a live entity (e.g. "what Dictionaries exist in this configuration?").

VS Code / GitHub Copilot Chat​

VS Code's native MCP support (used by both Copilot Chat's agent mode and other MCP-aware chat extensions) reads .vscode/mcp.json. Use the inputs/${input:...} mechanism so the key is prompted for interactively and kept in VS Code's own secret storage, never written into the file itself:

{
  "servers": {
    "kandra": {
      "type": "http",
      "url": "http://localhost:5219/mcp",
      "headers": {
        "X-Api-Key": "${input:kandra-api-key}"
      }
    }
  },
  "inputs": [
    {
      "id": "kandra-api-key",
      "type": "promptString",
      "description": "Kandra MCP API key",
      "password": true
    }
  ]
}

The first time the server starts, VS Code prompts for the key (masked, since "password": true) and reuses it for the session without ever writing it to .vscode/mcp.json — safe to commit the file as-is.

GitHub Copilot Chat in VS Code answering a question about FEFO-costed sales, using the Kandra MCP server configured in .vscode/mcp.json

Codex​

Codex reads its MCP server list from config.toml — either a project-local .codex/config.toml, or ~/.codex/config.toml to make the server available in every project. The one gotcha worth calling out explicitly, since it differs from both clients above: Codex's env_http_headers maps a header name to an environment variable name, not a literal header value — Codex reads the actual key from your shell environment at launch, so it never touches the config file either.

[mcp_servers.kandra]
url = "http://localhost:5219/mcp"

[mcp_servers.kandra.env_http_headers]
X-Api-Key = "KANDRA_API_KEY"

Export the key before starting Codex (export KANDRA_API_KEY=kdr_... on macOS/Linux, $env:KANDRA_API_KEY = "kdr_..." in PowerShell) — Codex substitutes it into the X-Api-Key header on every MCP call.

A real session, asking for a report and then a follow-up that needs the same context:

Codex generating a FEFO write-off report by week, split into food products and medicines, via the Kandra MCP server Codex following up in the same Codex session to show the accounting transactions posted for those same FEFO sales

See also​