Unified AI System 0.4.0: one-command Codex MCP gateway with local prompt enhancement

[happy520ai]

I wanted a way to exercise an AI gateway from Codex before configuring a paid model provider or placing credentials in the runtime.

I built that path into Unified AI System, an Apache-2.0, terminal-first self-hosted gateway. Version 0.4.0 now adds deterministic, provider-free prompt enhancement and a complete Codex plugin bundle.

The smallest direct MCP path is still one anonymous container command:

codex mcp add unified-ai-system -- docker run --rm -i ghcr.io/happy520ai/unified-ai-system/mcp-server:0.4.0

Restart Codex, then use /mcp to inspect nine tools:

• gateway_health, gateway_readiness, gateway_prompt_enhance, gateway_chat
• knowledge_readiness
• workflow_health, workflow_actions
• workforce_health, workforce_agents

The new gateway_prompt_enhance tool turns a rough English or Chinese request into a structured prompt locally. It preserves the original request, reports the detected language and selected profile, and makes no provider call. Applying that enhanced prompt to chat remains explicit; ordinary chat behavior is unchanged.

A useful first Codex task is:

Use gateway_prompt_enhance to turn "build a small API for my team" into a coding prompt. Show the preserved original request, detected language, enhanced prompt, and proof that no provider was called.

For the complete Codex bundle, the repository is also a plugin marketplace source:

codex plugin marketplace add happy520ai/unified-ai-system --ref master

After restarting the ChatGPT desktop app, open Plugins from ChatGPT Work or Codex, choose the Unified AI System marketplace, and install the plugin. The bundle adds the focused gateway skill and starts the credential-free MCP runtime in an isolated container.

The default path is intentionally constrained:

• The MCP process starts and owns an isolated managed gateway.
• Readiness verifies the deterministic fake provider and confirms real-provider calls are disabled.
• Chat rechecks that boundary on every call and fails closed if real execution may be active.
• The managed gateway and container are removed when the MCP session ends.
• The Codex plugin pins the reviewed 0.4.0 multi-platform image digest, disables container networking, and drops Linux capabilities.

Container-level verification still matters here. An early image launched through a package-manager wrapper whose lifecycle banner reached stdout before JSON-RPC and broke the stdio stream; source tests did not expose it. The current verifier exercises the published image's real default command through initialization, tool discovery, every documented tool, fake execution, and process cleanup.

Current evidence and guides:

• Repository: https://github.com/happy520ai/unified-ai-system
• v0.4.0 release: https://github.com/happy520ai/unified-ai-system/releases/tag/v0.4.0
• Official MCP Registry: https://registry.modelcontextprotocol.io/v0.1/servers/io.github.happy520ai%2Funified-ai-system/versions/0.4.0
• 60-second Codex guide: https://github.com/happy520ai/unified-ai-system/blob/master/docs/codex-mcp-quickstart.md
• Prompt-enhancement behavior: https://github.com/happy520ai/unified-ai-system/blob/master/docs/prompt-enhancement.md
• Codex plugin manifest: https://github.com/happy520ai/unified-ai-system/blob/master/.codex-plugin/plugin.json
• Current CI: https://github.com/happy520ai/unified-ai-system/actions/runs/30755010041
• Container build and publish: https://github.com/happy520ai/unified-ai-system/actions/runs/30755010044
• HOL Plugin Scanner: https://github.com/happy520ai/unified-ai-system/actions/runs/30755010056

This is an unofficial public engineering preview and is not affiliated with or endorsed by OpenAI. Real-provider access remains a separate explicit configuration and authorization path; this is not a production-readiness, L5-autonomy, or AGI claim.

For people using Codex with operational MCP servers, I would value feedback on two questions:

1. Should deterministic prompt enhancement remain a dedicated tool, or would an optional chat field feel more natural?
2. Which readiness evidence should a gateway expose before a user can opt into real-provider execution?

If the direction is useful, testing, feedback, or a Star would help other Codex users discover it.