codex-tg: local Codex control plane

Local control plane for OpenAI Codex App Server, built in Go, with Telegram as the first production adapter.

codex-tg watches local Codex threads, keeps durable thread identity visible, routes operator input back to the right turn, and exposes high-signal controls such as Plan Mode prompts, Stop, Steer, Details, Tools file, and Get full log. The current adapter is Telegram; the v0.5 direction is a reusable local codex-control layer for future router agents, voice assistants, tray flows, and other private adapters.

Why codex-tg?

• Local Codex control plane for OpenAI Codex App Server.
• Control and observe Codex threads through adapters without exposing App Server to the internet.
• Use Telegram today as a high-signal notification, reply, approval, and Details surface.
• Reuse your existing Codex setup: skills, MCP servers, plugins, repo instructions, and local workflows.
• Thread-first routing keeps replies, tools, Plan Mode, Details, and Final cards attached to the right run.
• Built toward long-running local coding-agent orchestration and future router-agent workflows.

Current release: v0.5.0.

The demo flow is documented in docs/demo/telegram-plan-mode-demo.md.

Why It Matters

• Keep local Codex work observable and controllable without exposing Codex App Server to the internet.
• Give future router agents a stable local control surface for Codex threads, turns, approvals, events, and skills.
• Use Telegram as a low-friction fallback and notification surface on unreliable or constrained networks.
• Preserve local-first ownership: Codex sessions, workspaces, SQLite state, and tokens stay on your machine.

Remote Connections

Official Codex Remote Connections cover the broad mobile remote-control workflow for Codex. codex-tg is not trying to replace that feature. The project direction is a local control layer and adapter system: Telegram remains useful for operator notifications, replies, approvals, Details, and logs, while future adapters can consume the same Codex control core.

Demo Screenshots

1. User request and Plan Mode

2. Tool execution and output

3. Final answer and Details

Features