Solutions

Does Codex read CLAUDE.md?

Updated

No. Codex reads AGENTS.md, the cross-tool convention, rather than Claude Code's CLAUDE.md - so the file does not carry by itself, though copying or symlinking it across is a minute's work and the two formats are close enough that it mostly just works. What does not carry, and what no rename fixes, is everything that was never in a file: the findings, the abandoned approaches, the reasons behind the decisions. Those live in whichever agent's session history they happened in, and they are lost when you leave it.

You have a CLAUDE.md that took months to get right, and you want to try Codex, or run both. The obvious question is whether the file comes with you. It turns out to be the least interesting part of the move.

This is the shape of every agent migration. Instructions are portable because somebody standardised them, so moving between agents is mostly a rename and an afternoon. Memory is not portable, because each vendor keeps it in its own store and none of them reads another's. The cost of switching is therefore not the configuration; it is that the new agent starts knowing nothing about work the old one had already done. In Forkbench that material sits in neither agent. A Thread holds one job's plan, its notes and its keys, and whichever agent you open in that Thread reads them - so switching agent is changing the program running in the pane, not migrating anything.

How it works

  1. 1Copy CLAUDE.md to AGENTS.md, or symlink one to the other, so both agents read the same instructions and you maintain one file.
  2. 2Do not paste your findings into it. That is what makes these files unmaintainable, and it is the material most worth keeping somewhere better.
  3. 3Put the job in a Thread and file the findings there as notes, so they belong to the work rather than to whichever agent happened to be open.
  4. 4Run both for a while. They share the Thread's notes and its board, so a Codex session picks up where a Claude Code one stopped.
  5. 5Judge the switch on the work rather than on the setup. The setup is the part that was already solved.

Straight about the guarantee: Which files an agent reads is its vendor's to change, and this was last checked on 6 September 2026 - confirm against the tool's own documentation before relying on it. Forkbench does not convert, sync or maintain either file. It runs both agents in real login shells and gives them a shared place to write that belongs to neither vendor.

Frequently asked

Sources