zoeynine/local-codex-bridge

Developer tools

This page was translated automatically. Technical details may differ from the original. View the original

43 stars 9 forks Quality 63 Trend 33MIT

Repository Install

Allow ChatGPT to call native Codex sessions via MCP using a lightweight Windows bridge.

Overview

A thin supervisory MCP bridge between external AI supervisors and native Codex. Local Codex Bridge is a lightweight MCP stdio adapter for Windows and macOS: It doesn't solve the problem of "creating another Codex," but rather enabling AI skilled in dialogue, planning, and continuous supervision to directly supervise native Codex to complete real engineering tasks. - Do not create a second job/task system; - Do not duplicate Codex conversation history; - Do not maintain parallel thread databases; - Do not cache the "current model" state; - Do not replace Codex's own session/thread/turn semantics. V2.1.3 continues to tighten the boundaries of Bridge as a supervisory adapter and adds: - On-demand discovery of native model/list; - Optional model/reasoning-effort override of codex_turn; - stable permission approval response； - The UNKNOWN semantics of mutating acknowledgement timeouts; - Consistency between public tool descriptions and runtime constraints; - Unified version anchors and upgrade assumption checks. Windows and macOS share the same core bridge, with differences only retained in the platform's native path, launcher, checkpoint default directory, process start, and termination system boundaries. - Understanding user goals; - Breaking down tasks; - Determining the scope of work and risk boundaries; - Deciding when to continue observing, correcting, approving, or interrupting; - Determining whether the results meet acceptance criteria; - Providing supervision when Codex cannot make secure decisions on its own.

Overview (machine translated)

English · Chinese Simplified

A thin supervisory MCP bridge between external AI supervisors and native Codex.

Current release: V2.3.4

Local Codex Bridge is a lightweight MCP stdio adapter for Windows and macOS:

It lets an AI that is good at conversation, planning, and sustained supervision oversee real engineering work performed by local native Codex. It does not recreate Codex.

The supervisor owns the goal, resources, boundaries, risks, approvals, and acceptance. Codex retains its native autonomy for coding and execution.

The Bridge stays thin:

It creates no second job or task system.
It does not copy Codex conversation history.
It keeps no parallel thread database.
It does not cache the “current model.”
It does not rebuild Goal, Queue, or Search.
It does not replace Codex session, thread, or turn semantics.

The native Codex thread/session remains the source of truth for execution.

Changelog · Protocol and compatibility assumptions

What V2.3.0 introduced

V2.3.0 expanded the public supervisory surface to 12 MCP tools while retaining the boundary of exposing only the native surface needed for supervision.

The release added or refined:

This page was translated automatically. Technical details may differ from the original.

Install

This server does not publish a one-line install command.

Open the repository installation guide New to MCP? Start with the mcpskill.org guide →