Skip to main content
Glama
README.md
# codex-bridge

**Codex jobs for Claude Code that survive crashes, closed windows and reboots, with Codex image generation built in.** A dependency-free MCP server built Windows-first: every call is a job saved on disk, so you can close Claude, lose power, or kill a process, and pick the work up again with `codex_resume`.

Roles: **Claude Code implements. Codex plans and builds images.** Optionally (opt-in, one folder), Codex can also write code while Claude reviews: see [Let Codex write code](#let-codex-write-code-opt-in). And an optional pane beside the chat lists your jobs live and shows your Codex usage and when it resets: see [Codex jobs pane](#codex-jobs-pane).

![Recorded session: Claude Code starts codex-bridge and asks Codex for a plan](docs/demo.svg)

<sub>The server calls in this recording are real (Codex at `effort: low`, idle time trimmed); the `claude mcp add` line is scripted. The raw recording is [docs/demo.cast](docs/demo.cast); play it with `asciinema play docs/demo.cast`.</sub>

## Quickstart

You need [Node.js](https://nodejs.org) 20 or newer and the [Codex CLI](https://github.com/openai/codex) installed and logged in.

```bash
claude mcp add codex-bridge -- npx -y @xallindrome-labs/codex-bridge
```

Restart Claude Code, then ask it something like *"Get a plan from Codex for adding rate limiting, in the background, and keep going"*. Claude will call `codex_plan` with `background: true` and collect the answer later.

Other MCP clients: run `npx -y @xallindrome-labs/codex-bridge` as a stdio server.

> **Platform note:** developed and tested on Windows 11, where real Codex jobs have been run. CI checks on Linux, macOS and Windows (Node 20 and 22) that the server starts, speaks MCP and works when installed from the npm tarball, and on Linux and macOS that jobs run, finish and survive the server being killed, using a fake `codex`. CI has no real Codex CLI, so real Codex jobs on macOS and Linux are untested. Reports and fixes are welcome.

| Tool | What it does | Sandbox | Default model / effort |
|---|---|---|---|
| `codex_plan` | Codex writes implementation plans, design reviews and second opinions. It inspects the project in `cwd`. | `read-only` (always) | Your Codex config defaults |
| `codex_image` | Codex generates or edits images. The bridge copies each one into `output_dir` as soon as it's made. | `read-only` (always) | **gpt-6-luna / high** |
| `codex_code` | **Off by default.** Codex implements a change by editing files in one folder you choose. See [below](#let-codex-write-code-opt-in). | `workspace-write`, that folder only, no network | Your Codex config defaults |
| `codex_result` | Gets a job's result or progress (`wait_seconds` up to 240), or lists recent jobs. | | |
| `codex_resume` | Continues an interrupted, timed-out or failed job from where it stopped. | | |
| `codex_cancel` | Stops a running job. | | |

`codex_plan`, `codex_image` and `codex_code` accept:
- `effort`: `low`, `medium`, `high`, `xhigh`, `max` or `ultra` (Luna supports up to `max`).
- `model`: overrides the default model for that call.
- `background: true`: returns a `job_id` immediately so Claude can keep working.
- `session_id`: continues an earlier Codex thread, for example "make the icon darker".

`codex_image` also takes `count`, the number of images wanted. Set it for batches: it enables `k/count` progress and resumes that ask for exactly the missing images.

## Let Codex write code (opt-in)

By default Codex can only read. If you want it to implement things while Claude reviews, name the one folder it may edit:

```bash
claude mcp add codex-bridge -e CODEX_BRIDGE_WRITE_ROOT=/path/to/your/project -- npx -y @xallindrome-labs/codex-bridge
```

Restart Claude Code and a fourth tool, `codex_code`, appears. Ask for it in plain words: *"Have Codex implement the plan in the background, then review its diff."*

- **One folder.** `cwd` must be `CODEX_BRIDGE_WRITE_ROOT` or inside it. Anything else is refused, checked on real paths so `..` and symlinks don't escape. A drive root or your home folder is refused as too broad.
- **No network, no commits.** Codex runs with network access off and is told to leave changes uncommitted so you can see them in `git diff`. The no-commit rule is a prompt, not a sandbox rule.
- **Claude reviews.** Every `codex_code` result ends with a reminder to read the diff before accepting it. A failed or cancelled run says Codex may have left partial edits.
- **Same job machinery.** It survives restarts, `background: true` and `codex_resume` work the same way, and a resume re-checks the folder rule.
- **Temp files.** In our test on Windows 11, writes outside the folder and network access were blocked, but the system temp folder stayed writable. Treat temp as writable. `CODEX_BRIDGE_WRITE_ALLOW_TMP=1` stops asking Codex to block it.
- **Don't edit the same files yourself** while a `codex_code` job runs.

Unset the variable and restart, and the tool is gone; `codex_plan` and `codex_image` are always read-only.

## Codex jobs pane

An optional add-on that opens a pane beside the Claude Code chat listing your Codex jobs live: running, done, interrupted, with age, request and `images k/n` for batches. It also shows your Codex usage: the 5-hour and weekly windows, percent used, when each resets, and how many images were generated in each (Codex has one shared meter, so images use the same limits). It also has a form for starting a job yourself: type your ask, pick plan, code or image, an effort and a model, and press Run (or Enter). No Claude turn is spent. It only reads the job files on disk, so it shows jobs from any session, including ones that kept running after you closed Claude. It also puts `codex: N running` in the status line and shows a toast when a job you watched finishes.

At the prompt of a terminal Claude Code session:

```
/plugin install codex-jobs --marketplace xallindrome-labs/codex-bridge
```

In the desktop app, where `/plugin` does not run: click **+** next to the prompt box, then **Plugins**, **Add marketplace** with `xallindrome-labs/codex-bridge`, then **Add plugin**. Without a marketplace, for one session: `claude --plugin-dir "$(npx -y @xallindrome-labs/codex-bridge --pane-path)"`.

Then type `/codex-jobs`. It also opens by itself the first time a job is running in a session.

The form uses `codex-bridge run`, which you can also call yourself: it starts the same durable job the MCP tools make and prints `{"job_id": ...}`.

```bash
echo "Review the auth flow" | npx -y @xallindrome-labs/codex-bridge run --type plan --cwd . --effort high --prompt-stdin
```

For `code`, the form passes the folder you typed as that one job's write folder, because starting it from your own form is your explicit choice. A `code` job asks first: the first Run names the folder Codex will edit, and the second Run starts it (changing anything in the form takes the question back). The form needs a keyboard, so the phone surface shows the jobs and usage but no form. Details and how it works are in [plugins/codex-jobs](plugins/codex-jobs/README.md). It needs a Claude Code build that supports mods; it is tested against the engine's own test kit, not every surface's paint.

## How it compares

OpenAI's [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) is the official way to use Codex from Claude Code, and it's good at what it does: slash commands for review and rescue. codex-bridge is an MCP server instead, with a different focus: **jobs that don't get lost**, and **images**.

Several open issues in the plugin's tracker are about jobs getting lost, and codex-bridge is designed to avoid each of them. These come from reading the designs, not from a head-to-head test:

| Reported in codex-plugin-cc | How codex-bridge handles it |
|---|---|
| Background job shows "running" forever after its worker dies ([#704](https://github.com/openai/codex-plugin-cc/issues/704), [#699](https://github.com/openai/codex-plugin-cc/issues/699)) | Status is worked out from files on disk plus a heartbeat and a process check. A dead runner shows as `interrupted`, never "running", and `codex_resume` continues it. |
| Windows `cancel` silently fails ([#708](https://github.com/openai/codex-plugin-cc/issues/708)) | Cancel writes a marker file; the runner that owns the Codex process sees it within about 2 seconds and stops the process tree by calling `taskkill.exe` directly, with no shell in between. |
| Windows broker processes leak ([#718](https://github.com/openai/codex-plugin-cc/issues/718), [#782](https://github.com/openai/codex-plugin-cc/issues/782)) | There is no shared broker. Each attempt is one runner process that exits when Codex does. |
| One session ending kills other sessions' jobs ([#671](https://github.com/openai/codex-plugin-cc/issues/671)) | Runners are started outside Claude's process tree (via WMI on Windows), so closing a session can't take them down. |

What the plugin does that this doesn't: slash commands, and its review-gate and rescue workflows. And this one is newer, with far fewer users. If the plugin works for you, keep using it. If you've lost jobs, or want Codex to generate images that survive a crash, try this.

## Jobs survive restarts

Every call is a job saved under `~/.codex-bridge/jobs/<job_id>/`, including calls that aren't run in the background.

- **Claude closes or crashes:** the job keeps running. `server.mjs` starts `runner.mjs` through WMI (`Win32_Process.Create`), so the job isn't part of Claude Code's process tree. Closing Claude can't kill it, and any later Claude session can collect it with `codex_result`.
- **The job itself dies** (reboot, power loss, Codex crash, or killed process): the job is marked `interrupted`. The next Claude session is told about it when it starts. `codex_resume` continues it:
  1. It copies out any images generated just before the interruption.
  2. It resumes the same Codex thread: "N are already done, generate only the remaining ones."
  3. If no Codex thread had started yet, it reruns the original request.

## How images are collected

Codex never writes image files. Its image tool stores every generated image in `~/.codex/generated_images/<thread>/`. The runner checks that folder every 2 seconds and copies new images into `output_dir` as `<filename>.png`, `<filename>-2.png`, and so on, in the order they were made. `copied.json` in the job folder records what has been copied, so an image is never copied twice across attempts.

This gives three benefits:
- Image jobs run in a read-only sandbox.
- A crash loses at most the one image being generated.
- Resume counts exactly what's done.

It also avoids a real problem: when Luna was allowed to save files itself, it re-typed each image as base64 through dozens of shell commands. That was slow, and the saved file no longer matched the original.
- **Codex stalls:** the runner stops a run after 10 minutes with no output. The job is marked `timed_out`, and it can be resumed.

Each attempt keeps its own `prompt.txt`, `events.jsonl` (the Codex event stream), `stderr.txt`, heartbeat and `result.json`. Finished jobs are deleted after 14 days.

## How it works

Each attempt runs `codex exec --json` (or `codex exec resume <thread>`) with the prompt on stdin. No shell is involved.

Every call runs read-only, except `codex_code` when you have opted in (it runs `workspace-write` in your one folder, with network off). New sessions pass `-s read-only` plus `-c sandbox_mode="read-only"`; resumed sessions pass `-c sandbox_mode="read-only"` (`codex exec resume` takes no `-s`). Your Codex config may default to `danger-full-access`, but that default never applies to bridge calls. The `plan` smoke test checks that a write attempted from a resumed session fails.

| File | Role |
|---|---|
| `server.mjs` | The MCP server. It creates jobs, launches runners, works out job status from the files on disk, and handles resume and cancel. |
| `runner.mjs` | Runs one attempt: starts Codex, records the heartbeat, watches for a stall or cancellation, collects images and writes `result.json`. |
| `lib.mjs` | Shared helpers: finding the Codex binary, building arguments, parsing events, handling images. |

## Install options

```bash
# for all your projects
claude mcp add -s user codex-bridge -- npx -y @xallindrome-labs/codex-bridge

# or from a local checkout
claude mcp add codex-bridge -- node /path/to/codex-bridge/server.mjs
```

Remove it with `claude mcp remove codex-bridge` (add `-s user` if you installed it that way).

## Configuration (env vars)

| Var | Default |
|---|---|
| `CODEX_BIN` | Auto-detected. It prefers a `codex.exe` next to `codex-code-mode-host.exe`, which is currently `~/.codex/plugins/.plugin-appserver/codex.exe`. |
| `CODEX_BRIDGE_HOME` | `~/.codex-bridge` |
| `CODEX_BRIDGE_IDLE_TIMEOUT_MS` | `600000` (stops a run after 10 minutes with no Codex output) |
| `CODEX_BRIDGE_PLAN_MODEL` / `CODEX_BRIDGE_PLAN_EFFORT` | Your Codex config defaults |
| `CODEX_BRIDGE_IMAGE_MODEL` / `CODEX_BRIDGE_IMAGE_EFFORT` | `gpt-6-luna` / `high` |
| `CODEX_BRIDGE_WRITE_ROOT` | Unset: `codex_code` is off. Set to one folder to let `codex_code` edit files there. |
| `CODEX_BRIDGE_WRITE_ALLOW_TMP` | Unset. `1` stops asking Codex to block temp-folder writes (not enforced on Windows anyway). |
| `CODEX_BRIDGE_CODE_MODEL` / `CODEX_BRIDGE_CODE_EFFORT` | Your Codex config defaults |

## Test

```bash
node smoke-test.mjs <scratch-dir> [plan|image|jobs|durable|all]
```

The smoke test drives the server over stdio like an MCP client and needs a logged-in Codex CLI; it uses a little of your Codex quota. The `durable` part uses `taskkill`, so it is Windows-only for now.

- `plan`: runs a plan, then tries a write from a resumed session, which must fail.
- `image`: generates one image.
- `jobs`: tests background jobs, cancellation and the `effort` check.
- `durable`:
  1. Kills the server while a job is running, then checks that a new server collects the result.
  2. Hard-kills an image batch partway through, then checks that a new server reports it as interrupted and that `codex_resume` finishes it.

## Speed

Each call has a fixed cost of about 4–5 seconds, and changing settings barely moves it. The rest of the time is Codex thinking and generating, which takes seconds to minutes. To spend less time waiting, use `background: true`, a lower `effort` for quick questions, and include relevant code in `context` so Codex reads less of the project.

## Notes

- This uses Codex's own login and quota, with no separate API key. A job left running after Claude closes keeps using that quota until it finishes.
- Bridge calls run headlessly, so they don't show up as conversations in the Codex Desktop sidebar.
- Don't send secrets in prompts. Prompts go to OpenAI.

## Issues and contributing

Bug reports and ideas are welcome on the [issue tracker](https://github.com/xallindrome-labs/codex-bridge/issues). Pull requests are welcome too, but reviewed slowly; see [CONTRIBUTING.md](CONTRIBUTING.md). Please report security problems privately, as described in [SECURITY.md](SECURITY.md).

## License

[MIT](LICENSE)

---

**Built by [Xallindrome](https://xallindrome.co)**: we build automation and AI workflows for small businesses. If you'd like something like this for your business, [book a diagnostic](https://xallindrome.co).