Ask Claude Code for a second opinion from your Codex thread, and the answer comes back into the same thread

I'm the author of postbag, a small local MCP server. Free and open source, MIT. I use Codex and Claude Code on the same repository, and I wanted Codex to be able to ask Claude Code something without me carrying text between windows. Both sessions join a named bag. From your Codex thread you ask for Claude Code's opinion, and postbag submits the letter to Claude Code's native inbox. The answer comes back through codex queue into the Codex thread you were already in, with that thread's context. Accepted letters become user turns. Queueing and the receiving app's policy can delay, hold or refuse them. It works in the other direction too. I use it with Codex in the CLI and in the desktop app. I used it on postbag itself. For the review that became 1.4.0, Claude Code handled the core changes and tests, Codex handled the MCP layer, and each reviewed the other's work through postbag. Codex reproduced an injected fsync failure where the core said a letter was not recorded even though its record was readable, so a blind resend could duplicate it. It also found a test fixture that left HOME pointing at my real home, so an inventory test could read my personal bags. Claude Code caught a flaw in the MCP refusal handler that could turn a definite refusal into an unknown outcome. These findings led to fixes and regression tests. Two agents can agree and both be wrong, so their findings still need checking against the code and tests. Setup: pipx install 'postbag[mcp]', then claude mcp add and codex mcp add as in the README. Start or reconnect the sessions and check that postbag's five tools are available. Then two prompts. In Claude Code: "Join postbag bag review as claude." In Codex: "Join postbag bag review as codex, then ask claude for a second opinion on my last commit." Requirements and limits. Use a recent Codex CLI with codex queue, and check that postbag can join from both sessions. The tested host versions are listed in the repository's native compatibility notes. Both sessions run on one machine. Native delivery has been tested on macOS. Fixture tests pass on Linux, but native Linux delivery has not been verified. Native Windows is unsupported. The Claude Code side uses a messaging socket that is not a documented API, so an update there can break delivery. A delivered letter becomes a user turn, so connect only sessions you trust. If you want to review each send, configure the host to ask for that tool and verify the setting. Not affiliated with OpenAI or Anthropic. Repo in the first comment.