MCP
How I use MCP in the Codex and Claude Code harness
The role of MCP in the tests: write, reread, attack, isolate, migrate, then verify artifacts with Codex and Claude Code.
MCP as a real surface
I never test memory as an abstract database. I test it as a tool that agents actually reach for, which is why MCP sits at the center: it exposes the primitives that Codex, Claude Code or Hermes will really call.
The repo suites cover simple cases, scale cases, migrations, concurrent writes, errors, size limits, malicious paths and symlinks.
Codex and Claude Code as the workbench
Codex reads the repo, runs scripts, inspects results and turns logs into verifiable pages. Claude Code appears in transcripts and methodology reviews, especially to challenge public claims.
None of this is a demo harness. It produces real `.md`, `.json` and `.log` files, expected results, assessments and runner scripts, and those are exactly the files the dashboard exposes.
Why I show failures
Failures are where the value hides. They expose planner limits and self-learning risks, symlink problems, competitor ingestion costs, and the spots where security still needs hardening.
A real tester does not bury those lines. The job is to make them readable.