CodexBridge

CodexBridge

中文版本

CodexBridge wraps the official Codex CLI/SDK and exposes it as an OpenAI-compatible /v1/chat/completions endpoint. It also ships with a tiny CLI so you can talk to Codex locally. Any OpenAI-style client (OpenWebUI, Cherry Studio, curl, etc.) can treat CodexBridge as a drop-in model.

Highlights

• OpenAI-compatible API – /v1/chat/completions (sync + SSE) plus /v1/models.
• Persistent sessions – Use session_id / conversation_id / thread_id / user to keep Codex context; omit them for ephemeral runs.
• Multimodal input – Accepts image_url and local_image blocks (HTTP(S), file://, data: URIs, local paths) and automatically converts them to Codex local_image.
• Structured output – Maps OpenAI response_format / output_schema to Codex outputSchema so the agent must return JSON that matches your schema.
• Configurable sandbox – Environment variables control filesystem access, working directory, networking, web search, and approval policy.
• Self-hosted storage – Codex sessions live in ~/.codex/sessions; bridge mappings stay in .codex_threads.json.

Requirements

• Node.js 18+
• Codex CLI installed and authenticated on the same machine
• npm (or adapt scripts for pnpm / yarn)

Installation

git clone https://github.com/begonia599/CodexBridge
cd codexbridge
npm install
cp .env.example .env
cp .env .env.local   # optional custom config

Edit .env / .env.local before exposing the service (API key, sandbox mode, working directory, etc.).

CLI chat

npm run codex:chat

• Type natural language commands; Codex replies inline.
• /reset starts a new Codex thread.
• /exit leaves the REPL.

Thread IDs are cached in .codex_threads.json, so you can close the CLI and continue later.

HTTP bridge

npm run codex:server

• Default port: 8080 (PORT overrides)
• Health check: GET /health
• Session map: .codex_threads.json (delete to reset bridge-managed sessions)

Basic request

curl http://localhost:8080/v1/chat/completions \
  -H "content-type: application/json" \
  -H "authorization: Bearer 123321" \
  -d '{"model":"gpt-5.5:medium","session_id":"demo","messages":[{"role":"user","content":"ls"}]}'

Session IDs & persistence

• In production/front-end setups enable CODEX_REQUIRE_SESSION_ID=true so every request must include a session identifier.
• IDs can be supplied in the JSON body (session_id / conversation_id / thread_id / user) or via headers (x-session-id, session-id, x-conversation-id, x-thread-id, x-user-id).
• IDs must be non-empty strings (UUID, user ID, chat ID, etc.). Avoid numeric-only or nested objects.
• If all identifiers are omitted and CODEX_REQUIRE_SESSION_ID is false, CodexBridge flattens the entire messages history into [ROLE] blocks and sends it as a one-off prompt.

Streaming

curl -N http://localhost:8080/v1/chat/completions \
  -H "content-type: application/json" \
  -H "authorization: Bearer 123321" \
  -d '{"model":"gpt-5.5:high","session_id":"stream","stream":true,"messages":[{"role":"user","content":"Explain how to run npm init step by step."}]}'

Response is SSE (data: {...}) ending with data: [DONE].

Multimodal input

The bridge understands OpenAI-style content arrays. Examples:

• {"type":"image_url","image_url":{"url":"https://example.com/demo.png"}}
• {"type":"local_image","path":"./images/demo.png"}

Remote URLs and data URIs are downloaded to a temp folder, converted to Codex local_image, then cleaned up after the turn.

Structured JSON output

CodexBridge forwards OpenAI response_format to Codex outputSchema:

{
  "model": "gpt-5.5",
  "session_id": "lint",
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "lint_report",
      "schema": {
        "type": "object",
        "properties": {
          "summary": { "type": "string" },
          "status": { "type": "string", "enum": ["ok", "action_required"] }
        },
        "required": ["summary", "status"],
        "additionalProperties": false
      }
    }
  },
  "messages": [
    { "role": "user", "content": "Check src/ for lint problems and return JSON following the schema." }
  ]
}

type: "json_schema" must include a schema. Missing or invalid schemas trigger HTTP 400.

Networking & search

• CODEX_NETWORK_ACCESS=true allows Codex to run networked commands (curl, git clone, API calls).
• CODEX_WEB_SEARCH=true lets Codex use its built-in web search tool.

Both default to false.

Source deployment

cp .env.example .env
npm run codex:server

Keep the process running via pm2, systemd, forever, or any other supervisor. If you need HTTPS, front the service with Nginx/Caddy pointing to localhost:8080. You can run npm run codex:chat alongside the server for CLI interaction.

Dashboard prototype

Visit http://<host>:8080/dashboard to see the placeholder dashboard (served from public/dashboard.html). The page currently loads mock content; once token parsing is wired up you can populate it with real account metadata and session stats.

One-command setup for Linux

curl -fsSL https://raw.githubusercontent.com/begonia599/CodexBridge/master/scripts/install.sh | bash

The script:

1. Installs git, Node.js, Docker, and Docker Compose if missing.
2. Clones/pulls the repo to ~/codexbridge (override via CODEXBRIDGE_DIR env var).
3. Copies .env.example → .env (edit afterward).
4. Runs docker compose up -d --build.

Requires sudo privileges to install system packages and Docker.

Docker

Docker Compose

docker compose up -d
docker compose logs -f codexbridge

Compose mounts Codex data under ./codex-data (auto-created). Delete that directory to reset Codex threads. Use docker-compose instead if you rely on the legacy command.

Docker CLI

docker build -t codexbridge .

docker run --rm -p 8080:8080 \
  --env-file .env \
  -v "%cd%":/workspace \
  -v "%LOCALAPPDATA%\Codex":/root/.codex \
  -w /workspace \
  codexbridge

Environment variables

Variable	Default	Description
PORT	8080	HTTP port
CODEX_MODEL	gpt-5.5	Default model
CODEX_REASONING	medium	Default reasoning effort (minimal / low / medium / high / xhigh)
CODEX_BRIDGE_API_KEY	123321	API key for Authorization: Bearer / x-api-key
CODEX_SKIP_GIT_CHECK	true	Skip Codex “trusted Git repo” requirement
CODEX_SANDBOX_MODE	read-only	read-only / workspace-write / danger-full-access
CODEX_WORKDIR	empty	Force Codex threads to run inside this directory
CODEX_NETWORK_ACCESS	false	Allow networked commands
CODEX_WEB_SEARCH	false	Allow built-in web search
CODEX_APPROVAL_POLICY	never	never / on-request / on-failure / untrusted
CODEX_LOG_REQUESTS	false	Log incoming payloads
CODEX_REQUIRE_SESSION_ID	false	Require session identifiers (true recommended for production)
CODEX_JSON_LIMIT	10mb	express.json() body limit
CODEX_DYNAMIC_MODELS	true	Fetch the live model list from the Codex backend for /v1/models; falls back to a static list if unreachable
CODEX_MODELS_TTL_MS	300000	Cache TTL (ms) for the dynamic model list
CODEX_MODELS_ENDPOINT	Codex backend	Override the model-listing endpoint URL
CODEX_CLIENT_VERSION	SDK version	client_version sent to the model-listing endpoint

Troubleshooting

• 413 PayloadTooLargeError – Increase CODEX_JSON_LIMIT when sending base64 images or large bodies.
• “Invalid or missing API key” – Provide the correct Authorization: Bearer <CODEX_BRIDGE_API_KEY> or x-api-key.
• Codex refuses to run (Git check) – Disable CODEX_SKIP_GIT_CHECK only if the working directory is a trusted repo.
• Need a clean slate – Stop the server and delete .codex_threads.json and/or ~/.codex/sessions.

License

Non-commercial use only. See LICENSE for the full terms. Commercial usage requires prior written permission from begonia.

Support & contact

• Repository: https://github.com/begonia599/CodexBridge
• Email: [email protected]

Feel free to reach out if you run into problems, want new features, or plan to fork and rework CodexBridge. Contributions and derivative projects are welcome.