Anyone using both Claude Code and Codex? How do you manage context?
I've been using Claude Code and Codex for product design and prototyping work, and lately I've been paying more attention to how quickly tokens get used up in both.
I often go back and forth on UI changes within the same conversation. As the work continues, the context gets longer, and I'm trying to find a more efficient way to manage it.
I've been looking into whether MCP tools or different workflows could help reduce token usage, but I'm not sure what actually makes the biggest difference.
I'm also curious if anyone here uses Claude Code and Codex together.
If you do, how do you keep project context and progress consistent between them? Do you find the two tools work well together, or does switching between them create more overhead than it's worth?
Would love to hear what's actually working for you.
Replies
I only use Claude Code, but the pattern that keeps my context small would work for both tools, because it lives in plain files, not in either tool. The project's main rules file holds only what applies to every change, and a table at the top maps each area of the app to its own notes file: store work, payments, the music engines, one screen, and so on. The rule is to read the one matching file before touching that area and to record new findings there. So a session about the paywall loads the payments notes and nothing about the music engines. Since the files are just Markdown in the repo, a second tool could read the same table, and progress lives in commits and notes rather than in either chat.
we run several claude code sessions on one repo and what carries over between them is a handoff file that only ever gets appended to.. top 20 lines are the current state, and every step ends in a commit, so a fresh session reads that plus the git log instead of me dragging a long chat along. any tool that reads files can pick it up the same way. when you switch tools do you start a new conversation each time or keep the old one going?
one thing worth trying is keeping a CLAUDE.md and AGENTS.md with shared project rules. That way both tools have a consistent reference instead of relying on chat history.
When passing work between them, Codex dropped the edge cases Claude flagged and reported the task completely finished. A broken function is easy to catch. Getting assured everything is fine after context silently truncates is what actually burns you.