◆ Sparround

Transports: stdio and HTTP

MCP supports two transport mechanisms:

stdio — the server runs as a separate process on your machine and communication goes over standard input/output streams. No network overhead, so it is the fastest option. It typically serves a single client.

Streamable HTTP — client-to-server messages go over HTTP POST, and the server can use Server-Sent Events when streaming is needed. It is for remote servers and supports standard HTTP authentication: bearer tokens, API keys, custom headers. The official recommendation is to obtain tokens with OAuth. One server can serve many clients.

CriterionstdioStreamable HTTP
Where the server runsOn your machineRemotely (or on the local network)
SpeedFastest — no networkNetwork latency applies
AuthenticationVia environment variablesOAuth, bearer tokens, headers
Number of clientsTypically oneMany
Typical useLocal tools, filesystem, your own scriptsSaaS services, a central server for the team
bash
# stdio — a local process. Everything after `--` is the server command.
claude mcp add --transport stdio my-tools \
  --env API_TOKEN=$MY_TOKEN \
  -- node ./tools/mcp-server.js

# HTTP — a remote server
claude mcp add --transport http notion https://mcp.notion.com/mcp

# HTTP with an authentication header
claude mcp add --transport http internal-api https://mcp.internal.company/mcp \
  --header "Authorization: Bearer $INTERNAL_TOKEN"

Connection examples. The `--` separator matters: everything before it belongs to Claude, everything after it is the server's launch command.

The SSE transport is deprecated — new servers should use http (Streamable HTTP). Existing --transport sse connections still work but should not be chosen for a new integration.

The practical rule:

  • Choose stdio when the server exposes things only your machine has (local files, adb, a local database), or when you wrote it yourself and are still iterating.
  • Choose HTTP when the server is a central service for the team, when it fronts a SaaS product, or when authentication needs to be centralised.

When writing your own internal server you typically start with stdio and move to HTTP once the team needs it.

📚 Sources and documentation