Using Claude Code, Codex or Cursor in a client's repo? Check the account and data settings before the first prompt
FDE Times walks through the defaults of the three main coding agents and the exceptions most likely to put a forward deployed engineer in breach of a client's data policy.
FDEs spend their days in repositories that belong to someone else, so where an AI coding agent's data goes matters more for them than for engineers working on internal products. A new FDE Times piece compares Claude Code, OpenAI's Codex and Cursor on exactly that question.
Some of the points it draws from each vendor's documentation: Claude Code keeps session transcripts as plaintext on the local machine for 30 days by default, including sessions run on a client's code. Anthropic says it does not train on Claude Code data under commercial terms (Team, Enterprise, API), but zero data retention is enabled per organisation only after the account team confirms eligibility. Codex running locally is confined by an OS-level sandbox with networking off by default, plus a separate approval policy. Cursor turns Privacy Mode on by default for team members, but the article flags that plugging in a personal API key forfeits the zero-retention commitment.
It also warns that a CLAUDE.md file is only an instruction to the agent; what actually limits it is the sandbox and the permission mode.
The practical lesson for FDEs is to ask which account and plan the work runs under, and to set sandbox and retention settings, before pointing any agent at customer code.
Source: fdetimes.net