codex-reasonix-bridge
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@codex-reasonix-bridgestart a Reasonix task to update dependencies and ask me before applying changes"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
codex-reasonix-bridge
Codex → Reasonix Studio 的持久协作桥,由 codex-dsh-bridge 0.2.0 派生。
Codex 规划与验收;Reasonix 执行;桥保存任务、会话、审批、结果及交付文件指纹。
默认 serve 连接共享后台:多个 Codex 聊天共用一个状态所有者,避免 MCP 启动时争抢写锁。
completed 仅表示执行轮次结束。
默认连接安装包里的 reasonix-studio-host,通过受认证的 loopback HTTP + SSE 工作。
桥启动自己管理的后端实例,沿用 Studio 的 Reasonix home、登录及模型配置;不会接管 GUI 的活跃聊天。
安装包无需包含 reasonix CLI。Node.js ≥24,零外部依赖。
共享后台支持 Linux/macOS Unix socket 与 Windows 命名管道;参见平台配置示例。
三平台离线 CI 使用假后端;真实 Windows/macOS Studio 仍待验收。
node scripts/configure-bridge.mjs --desktop-root '/opt/Reasonix Studio'
node src/cli.mjs preflight
node scripts/probe.mjs
node src/cli.mjs serve配置脚本可直接读取本地 Reasonix 提供商与模型菜单;--list-providers 只读列出目录。
此辅助读取使用 Python ≥3.11;不可用时退回缓存或手工输入,桥运行时仍只需 Node.js。
默认 provider/model/reasoningEffort 为 null,沿用 Studio 设置。工具审批默认 ask,使用
reasonix_approvals / reasonix_approve;不依赖手工工具表。原生提问与文本澄清通过 reasonix_reply 续跑。
Studio 的 effort HTTP 接口会改写全局配置,桥仅核对或保持当前档位;改档请在 Studio 内进行。
支持任务幂等、工作区排队、会话复用与恢复、未知结果人工裁定、显式重试、取消和有界退出。
默认 studioRuntimeReuse=false:每次后续提示从同一持久会话重建 runtime,兼容不支持
previous_response_id 的模型服务。确认服务支持链式响应后可设置为 true。失败请求不会自动重放。
状态目录与原 DSH 桥分开;不要把 DSH 的状态或配置直接复制过来。
可选 backend=acp 使用独立 Reasonix CLI,但尚未完成真实 CLI 验收。
npm run check
npm test
npm run test:unit
npm run smoke以上检查离线且不调用模型。probe 启动真实 Studio 并创建临时工作区会话,不发模型提示。
真实模型验收需显式运行 node scripts/smoke.mjs --allow-model --config bridge.config.json。
克隆仓库后运行配置脚本,再按使用说明注册 Codex MCP。 本机配置、会话状态、原始验收报告和配置备份不纳入公开源码。
GitHub Actions 在每次 push、PR 或手动运行时执行离线检查并生成源码包。
推送与项目版本一致的 v* 标签后,检查通过才创建草稿 Release,附源码包与 SHA-256 校验文件;审核后在 GitHub 发布。
不运行真实模型测试,不发布 npm。发布步骤见 CI 与发布。
This server cannot be deployed
Maintenance
Related MCP Connectors
Reliable async execution for agent tool calls: schema gating, retries, idempotency, audit trail.
Execution control plane for agents: capabilities, durable jobs, budgets, receipts, and BYOK.
Cross-agent artifact workspace with provenance across Claude Code, Codex, Cursor, LangGraph.
Hosted runtime for persistent agent teams, durable workflows, memory, schedules, and goals.
Related MCP Servers
- AlicenseNot gradedqualityBmaintenanceEnables durable bidirectional handoffs between any MCP client and OpenAI Codex Desktop tasks, with persistent callbacks, acknowledgements, and session-scoped state across restarts.1Apache 2.0
- FlicenseBqualityBmaintenanceEnables local, stdio-based task dispatch from Codex to DeepSeek Harness with independent review, bounded rework, persistent queues, and completion tracking.15-
- AlicenseAqualityBmaintenanceBridges MCP clients such as Claude Code and Cursor to independent Codex app-server sessions, letting them start, poll, steer, review, compact, cancel, and answer prompts for asynchronous Codex tasks over local stdio or a remote authenticated WebSocket. Runs read-only by default, with workspace-write sandboxing only when a user explicitly enables it, and reports job status, model listing, and progress back to the client.101MIT
- AlicenseAqualityAmaintenanceEnables a Chief of Staff client to find Codex tasks, create and submit project work with exact prompts, track progress via durable receipts, and continue or manage the same conversation locally.1077 npm3MIT