Codex Claude Bridge delegates work from Codex to Claude Code via local MCP

What the bridge does

A project published on dev.to on October 12, 2026, named Codex Claude Bridge, connects two coding agents that people usually run side by side in separate windows: Codex and Claude Code. The plugin exposes Claude Code to Codex through a local MCP (Model Context Protocol) bridge, so Codex can hand off work, collect the results, and inspect the actual changes rather than trusting a summary.

According to the post, the workflow starts by handing Codex a goal, a workspace, and acceptance criteria. From there Codex can spin up new Claude Code sessions or reconnect to an old one using its UUID, which the author presents as a way to keep long-running work alive across handoffs. Delegated implementation runs inside its own Git worktree, so changes stay separated from the main branch until Codex has reviewed them and chosen to merge.

A supervision loop, not fire-and-forget

The post describes a checkpointed delegation cycle. Claude Code reports back with finished batches, blockers, and points where it needs a decision, while Codex keeps the planning and review role and Claude Code handles the scoped implementation. Two parameters can be tuned separately: interaction frequency and reasoning effort.

The bridge also caches validation. Stored check results can be reused as long as the source snapshot, the environment, and the scope of the check have not changed, which avoids re-running the same checks against unchanged code.

Setup and compatibility

The requirements are modest: the Codex CLI, a working Claude Code installation, Node.js 20 or newer, and Git. Installation runs through Codex's plugin system in two steps: register the marketplace from the project's GitHub repository, then install the claude-code plugin from it. The suggested first task in the post is deliberately small — ask Claude to polish a repository's README in its own worktree, show the diff, and merge only after review.

The bridge preserves an existing Claude Code model configuration, third-party providers included, so custom routing carries over untouched. Both Windows and macOS work, and the post notes that Orca is optional and ordinary terminals are fine too, though it does not elaborate on what Orca adds.

Why it matters

Anyone running multiple coding agents today usually manages them by hand: copying tasks between windows, polling for progress, and reconciling diffs themselves. Codex Claude Bridge formalizes that into a supervisor-worker pattern on a local machine, with MCP as the connective tissue instead of a hosted orchestration service. A human still sees the diff before anything merges, but the busywork of checking in on the worker moves to the coordinating agent.

The project is early. The author explicitly invites feedback on session recovery, handoffs, and coordination overhead — a hint that those are the least settled parts — and points to the GitHub repository for the full setup guide, examples, and current limitations. If the session-recovery mechanics hold up in practice, the tool offers a template for how competing agent CLIs might cooperate rather than compete for a single terminal.