Hardening 1.x, and an HTTP transport

The first release worked, but it was built for one person on one laptop. The three releases that followed in the same week were about two things: making the read-only promise harder to get around, and letting something other than a desktop app connect.

A validator that reads the SQL

Version 1.0 checked statements with regular expressions. That catches the obvious cases and misses the creative ones. Version 1.1 parses each statement into a syntax tree and walks it, so a write hidden inside a comment, a string or odd spacing is still found. The regex check stays as a fallback when the parser can’t read a statement.

The same release added:

• Rate limiting, per client, 100 requests per 15 minutes by default.
• Structured logging with Pino, with passwords redacted.
• A test suite of 128 tests, so later changes don’t quietly break the guarantees.

Limits and secrets

Version 1.2 added a configurable cap on how many rows a query returns, so a careless SELECT * on a large table doesn’t flood the assistant. It also added support for Docker secrets, so the password can come from a mounted file instead of an environment variable, and it validates the host name format before connecting.

HTTP transport

Until 1.3, the server only spoke stdio: the client starts it as a child process and talks over standard input and output. That’s right for Claude Desktop or Cursor on your own machine. It doesn’t work for a web app, a shared service or an agent running somewhere else.

Version 1.3 adds an HTTP transport with three authentication modes:

• required: each caller sends their own IBM i credentials to POST /auth and gets a token. Queries run as that user.
• token: one pre-shared token, for a trusted service.
• none: for a trusted network only.

It keeps a connection pool per user and cleans it up when the token expires. It can serve HTTPS itself, and it publishes an OpenAPI 3.1 description at /openapi.json. Set MCP_TRANSPORT=http, or both to keep stdio running too.

Upgrading

npx mcp-server-db2i@latest

Nothing in 1.1 to 1.3 changes existing stdio setups.

What next?

Several people asked about Mapepire as an alternative to JDBC. I’m looking into it. If there is something else you’d want first, open an issue.