OpenAI closed the MCP door to Codex. This reopens it — bounded, pinned, and honest about what it is.

What changed

For a while, codex mcp-server let any MCP-capable tool call Codex as a tool. That path is gone.

1. Aug 24, 2026 Deprecated. OpenAI's changelog: “codex mcp-server is now deprecated. Use the Codex app server instead.”
2. Sep 5, 2026 Removed. Codex CLI 0.154.0 dropped the entry point (openai/codex PR #42993). Any assistant, orchestrator, or editor that still calls Codex as an MCP tool now fails with CONNECTION_CLOSED on every session.
3. The replacement The Codex App Server. A bespoke JSON-RPC protocol. OpenAI's own documentation says it “isn't an MCP server or a drop-in replacement” and is “experimental and not supported for production workloads.”

OpenAI's migration guidance covers two cases.

Safe on an experimental upstream

OpenAI calls the app server experimental and unsupported for production workloads. The bridge runs on it anyway — with a stricter operating posture than the command it replaces.

A lost start reply is uncertainty, not permission to resubmit. Replay the same request and the bridge returns the existing task. This is durable duplicate prevention across the failures the bridge can see, and it is described that way, not as a stronger guarantee.

When to use the app server directly

The bridge deliberately exposes a small part of the native API. Each task creates one new thread and one turn. If you are building a full Codex client, go to the app server — it is the only place you will find:

We would rather lose a user to the right tool than keep one with a false promise.

“Why not just codex exec?” For a one-shot script, do. For an orchestrator with timeouts, task identity, and cancellation, exec holds one tool call open for the whole run and gives you nothing to replay or cancel safely.

Get started

GitHub hosts the source, not a running service. You clone the bridge, pin a reviewed commit, install its locked dependencies, write a private profile, and register the local executable with your MCP client. Unix commands are shown; the platform guide covers Windows PowerShell.

1. Clone and select your platform.

   # macOS, Linux, native Windows 11 or WSL2
   git clone https://github.com/Aetheris-Platform/codex-mcp-app-server-bridge.git
   cd codex-mcp-app-server-bridge
   BRIDGE="$PWD"
   python3 tools/integrate.py

2. Install the locked dependencies. Setup is the only time anything is downloaded; launch never invokes uv, pip, or a build.

   UV_CACHE_DIR="$BRIDGE/.local/uv-cache" uv sync --locked --no-dev \
     --no-install-project --no-python-downloads

3. Create a private profile. Copy the example, replace every placeholder, and keep the real file outside version control with mode 0600. Record the exact executable's --version output and SHA-256; the bridge verifies both before dispatch. Do not put credentials in it.

   {
     "runtime_path": "/absolute/private/bin/codex",
     "runtime_version": "EXACT OUTPUT OF codex --version",
     "runtime_sha256": "REPLACE_WITH_64_HEX_SHA256_OF_EXECUTABLE",
     "socket_path": "/absolute/private/run/app.sock",
     "allowed_worktree_roots": ["/absolute/workspaces"],
     "allowed_models": ["EXPLICIT_MODEL_ID"],
     "allowed_sandboxes": ["read-only"],
     "data_root": "/absolute/private/state"
   }

4. Register it with your MCP client. Adapt the enclosing format to your client; the paths must be absolute.

   {
     "mcpServers": {
       "codex-bridge": {
         "command": "/absolute/path/codex-mcp-app-server-bridge/.venv/bin/python",
         "args": [
           "-m",
           "codex_app_server_adapter",
           "--profile",
           "/absolute/private/profile.json"
         ],
         "env": {
           "PYTHONPATH": "/absolute/path/codex-mcp-app-server-bridge/src",
           "CODEX_HOME": "/absolute/private/codex-home"
         }
       }
     }
   }

5. Use it. Call codex_start, retain the task ID, then codex_status, codex_result and codex_cancel as needed. Before activating a host, run the client-neutral integration test procedure in the repo.