One file, three coding assistants
On October 1, 2026, we used Claude Code, Codex and Cursor to continue work on the same private file in Stuff. Each assistant read the saved original before editing and read it again afterward. No document copying between chats was needed.
The setup
We used a separate Stuff demo account and Project Lantern, a fictional reading-club project. Each client received its own OAuth connection with read/write access to that space only. None could manage sharing. The original project brief stayed unchanged.
Install and connection instructions: Claude Code, Codex, Cursor. The plugin overview explains what skills add to an MCP connection.
The handoff
| Step | Action | Saved result |
|---|---|---|
| Claude Code | Read the project brief, created Assistant handoff test.md, read it back | Version 1, 100 bytes |
| Codex | Read the same file, appended one line using version 1 as its edit guard, read it back | Version 2, 160 bytes |
| Cursor | Read version 2, appended one line using version 2 as its edit guard, read it back | Version 3, 215 bytes |
The final Markdown was deliberately small so the result was easy to inspect:
# Assistant handoff test
- Claude Code: read the Project Lantern brief and created this checklist.
- Codex: continued from Claude Code on the same saved file.
- Cursor: continued from Codex on the same saved file.
You can repeat this with a real task: ask one assistant for a draft, another to revise it, and a third to check it. Give each the same Stuff file link and grant the required access separately. A private link is a reference, not permission.
Inspect the history
Open the file's Actions → Version history. These are three retained versions of one file, rather than three disconnected copies.
In this test, Claude Code reported Anthropic / claude-opus-4-6, and Cursor reported Cursor / Cursor Grok 4.6. Codex's saved version identified the Codex connection but did not include a model. Agent and model details are client-reported attribution, not an independent certification of which model ran.
Cursor's actual chat also showed its targeted edit and fresh read-back:
Revoke, then try a fresh request
After the handoff, we revoked each demo connection in Stuff → Agents. We asked each client to fetch the same file again, without reconnecting or switching connections.
| Client | Fresh request result |
|---|---|
| Claude Code | Required reauthorization; no file content returned |
| Codex | Authentication required |
| Cursor | Unauthorized |
All demo connections, including unused login attempts, were cleaned up. Reloading the file as its owner still showed version 3. Revocation stops new access; it does not delete files or erase content already in an assistant's chat history. Revocation details.
What this test covers
Claude Code used its plugin's own MCP server in an interactive session. Cursor used the local IDE plugin and its safe-edit skill. Codex loaded the installed project-context and safe-edit skills and used a separately configured direct MCP connection with an interactive approval for its edit. Codex's plugin-specific desktop OAuth flow remains unverified.
This demonstrates remote Markdown reads, creation, guarded edits, cross-assistant continuation and revocation in those configurations. It does not establish support for every client version, headless mode, cloud agent or binary editing workflow. The individual setup guides record the login and approval wrinkles encountered.