Can Codex and Claude Code Share Context?
Two terminal agents, one source of truth
You run Codex CLI in one terminal and Claude Code in another. Both are working the same repo, both are genuinely good, and both start every session knowing nothing about how you work, what you decided last week, or why the architecture looks the way it does. So you explain it twice — once to each — and re-explain it tomorrow.
The obvious question: can they just share what you told the other one?
Can Codex CLI see what you told Claude Code?
No — Codex CLI cannot see Claude Code's context, and vice versa. They're built by different companies (OpenAI and Anthropic), store their working state separately, and have no channel between them. Anything you established in one session stays in that tool.
This isn't an oversight. Conversational history and learned context are a retention moat — the longer a tool understands you, the harder it is to leave. Keeping that context locked to the platform is a feature for the vendor, not a bug you can configure away. Major LLM providers keep conversational history inside proprietary silos to create user retention moats, making direct native context sharing between competing vendors impossible.
Both tools do have their own project memory. Claude Code reads a CLAUDE.md at the repo root. Codex CLI has its own equivalent config for standing instructions. Each is real, and each is scoped to that tool in that project — it doesn't cross to the other agent, and it doesn't follow you to a different repo or a different machine. For continuity within one tool, they help. For sharing across the two, they don't touch the problem.
Why do you keep re-explaining the same repo twice?
Because each tool starts every session from zero, and you're running two of them. The cost of missing context isn't a wrong answer — it's a generic one. An agent that doesn't know your conventions writes code that's technically correct and stylistically foreign to your codebase, so you spend the session correcting it toward what you'd have gotten for free if it knew your setup.
Running more than one AI tool is now the norm, not the exception — 65.8% of Claude mobile app users also actively use ChatGPT, and the same multi-tool pattern holds among developers juggling terminal agents. Two capable agents on the same repo should not need the same briefing twice a day. The fix isn't picking one tool. It's giving both the same source of truth.
What actually works — one shared layer both connect to
The fix is a context layer both agents read from and write to: a single structured profile that lives outside either tool, connected to both through MCP. Instead of Codex CLI and Claude Code each maintaining their own private picture of you, they point at the same vault.
MCP — the Model Context Protocol — is the open standard that makes this possible. MCP provides an open JSON-RPC 2.0 communication standard that lets client applications read from and store data back to external context vaults. It's the same protocol both agents already use to connect to external tools; here it connects them to your context.
Two properties make a layer different from copy-pasting a file between the two:
- It's read + write, both directions. Both agents load your context at session start, and both can save new context back to the vault. A decision you make working with Claude Code is there when you open Codex CLI next. Custom agent instructions paired with MCP server integration enable automated bi-directional context synchronization across ChatGPT, Claude, Cursor, Claude Code, Codex CLI, Hermes, and OpenClaw.
- It stays current instead of going stale. Exporting conversational history manually creates a static, one-off snapshot that quickly becomes outdated, whereas protocol-driven context vaults update dynamically across active sessions. A
CLAUDE.mdyou hand-copy into your Codex config drifts the moment either changes. A shared layer updates in one place.
The result: both terminals share one memory. What you tell Codex CLI, Claude Code knows. What you decide in Claude Code, Codex CLI picks up.
How to set it up
Use Unabyss as the shared context layer both agents connect to. It holds your context in one vault you own and serves it to any MCP-compatible tool — which both Codex CLI and Claude Code are.
- Build your context vault. Connect your real sources — GitHub, Notion, email — and Unabyss extracts a structured profile in under 90 seconds. This is your identity, your working style, and your current projects, built from what you've actually done rather than a bio you typed once.
- Connect Claude Code over MCP. Point Claude Code at the Unabyss MCP server. From then on it loads your context at the start of every session, before you type a word.
- Connect Codex CLI over MCP. Same server, same vault. Codex CLI now reads the same context Claude Code does.
- Turn on save-back. With custom instructions, both agents can write new context back to the vault as you work — so a decision made in one shows up in the other. Save-back is automatable via instructions; it's not on by default, so you switch it on deliberately.
- Control what each sees. You decide which parts of your context each tool gets. Both agents need your project and working style; you choose whether either sees anything more.
One setup. Both terminals, same memory — and any other MCP tool you add later reads the same vault.
→ For the full cross-tool walkthrough beyond these two agents: How to Share Memory Across AI Tools
Codex CLI and Claude Code: what shares, what doesn't
| Native (no layer) | With a shared context layer | |
|---|---|---|
| Claude Code reads Codex CLI's context | No | Both read one shared vault |
| Codex CLI reads Claude Code's context | No | Both read one shared vault |
| Context follows you to a new repo | No (per-project files) | Yes (vault is portable) |
| Stays current automatically | No (manual file edits) | Yes (updates in one place) |
| You own and control it | Scattered per tool | One vault, you own it |
Native project files keep each tool consistent with itself. A shared layer keeps both tools consistent with you.
The point
Codex CLI and Claude Code will not share context on their own — that's how the vendors designed it. But you don't need them to talk to each other. You need one place that holds who you are and what you're working on, that both connect to. Own your context, connect both agents to it over MCP, and the two terminals stop starting from zero.
→ The full picture of why cross-tool memory works this way: What Is Multiagentic Memory?
