Receipts #056 — WORKS WITH CAVEATS: @modelcontextprotocol/server-redis (#056)
- Package: @modelcontextprotocol/server-redis, version 2025.4.25 (npm), run exactly as its README instructs:
npx -y @modelcontextprotocol/server-redis <redis-url>. Tested 2026-10-31. No real Redis and no real credentials: decoy URL plus a sandboxed fake-RESP server as the positive control. - Task:
setthengetthe key "receipts-test". - Expected (ground truth): the raw node-redis driver against the same decoy URLs, and the fake-RESP round-trip. The driver fails fast (0.03 s) on the unreachable host and its errors contain NO password.
- Observed: the round-trip itself works — initialize 1.80 s against the control,
setreturns "Successfully set key: receipts-test",getreturns "hello", 0.10 s per call. BUT the server logs the full connection URL, password in clear text, to stderr on every connect, error, retry, and success — the decoy password appears 40+ times across one test session (19 times in the wrong-password run alone), including on the happy path:
[Redis Connected] Successfully connected to redis://decoyuser:<decoy-password>@127.0.0.1:6399
Further caveats: it connects before the MCP handshake, so an unreachable server means initialize is never answered (~35–60 s of retries, then exit); it logs "Successfully connected" on a bare TCP accept, before AUTH; a wrong password is retried identically 5 times over ~34 s.
- Verdict: Works With Caveats. The function works; the logging behavior is a credential spill with any real password, and it is caveat #1.
- Evidence refs: record 056; write-up tests/056-redis.md; transcript tests/056-redis-transcript.txt (sections A–C: no-arg, decoy URL, wrong-password control, driver ground truth).
Corrections and retest requests: open an issue on the Lantern Works repo naming the record id. Paid re-tests are accepted under the Receipts firewall: money can buy being measured sooner, never a changed verdict.
Need GPU prices in your own agent? Try a free, cited comparison. Back to the Receipts index.