imageCritical Manufacturing

CM GPT plugin

Brings CM GPT — the Critical Manufacturing MES documentation assistant — into both Claude Code / Cowork and ChatGPT / Codex, from the same plugin folder. Bundles the connector to the CM GPT MCP server together with the search-and-answer instructions that were previously carried as a LibreChat agent system prompt.

This works on both platforms because Claude Code and OpenAI's Codex/ChatGPT plugin system now share the open Agent Plugins manifest format. Concretely:

• skills/ask/ is read as-is by both platforms — one skill file, no duplication.
• plugin.json (root) + mcp.json (root) are the portable Agent Plugins manifests, used by Codex/ChatGPT.
• .claude-plugin/plugin.json + .mcp.json are Claude Code's own manifest format. OpenAI's docs say it also accepts these as a legacy/compatibility fallback, but this repo keeps both pairs explicit rather than relying on that fallback.

Components

Setup — Claude Code / Cowork

Uses .mcp.json. The connector uses OAuth. On first use in a session, the user is prompted to sign in; the plugin is pre-configured with:

• Client ID: CMGPT
• Scope: IAM-CMGPT-Docs
• Transport: streamable HTTP

No environment variables or secrets are stored in the plugin.

appendOfflineAccess is set to true so the sign-in requests a refresh token and users are not re-prompted every session. If the CM identity provider does not issue offline_access for this client, remove that line from .mcp.json.

Setup — ChatGPT / Codex

Uses the root plugin.json + mcp.json (portable Agent Plugins manifests). Per OpenAI's MCP authentication docs, OAuth for a plugin's MCP server is negotiated by the server itself (standard MCP OAuth discovery, e.g. Client ID Metadata Documents or dynamic client registration) when a user connects — it's not declared in mcp.json, so no client ID/secret lives in this manifest. The CM GPT MCP server already handles OAuth this way for the Claude connector above; ChatGPT/Codex users hitting the same server should get an OAuth prompt the same way.

To register it locally for testing (repo-scoped): this repo's .agents/plugins/marketplace.json already lists this plugin. In the ChatGPT desktop app or Codex CLI:

codex plugin marketplace add ./cmgpt-plugin # or point at this repo's git URL once hosted

Then enable it in .codex/config.toml:

[plugins."cm-gpt@critical-manufacturing"]
enabled = true

Alternatively, in ChatGPT: Settings → Security and login → turn on Developer mode → ChatGPT Plugins → add the MCP server URL directly (https://criticalmanufacturing.ai/docs/mcp) for a personal plugin without going through the marketplace file.

Unverified: this hasn't been installed end-to-end in a live ChatGPT/Codex environment yet — the manifest shapes above follow OpenAI's published examples exactly, but actually connecting and confirming the OAuth handshake against criticalmanufacturing.ai still needs to be tested.

Usage

The skill triggers on its own for any Critical Manufacturing or MES question — no slash command needed. Users can also invoke it explicitly with /ask (or /cm-gpt:ask if another plugin's skill also claims /ask).

Examples that trigger it:

• "How do I configure a material dispatch rule?"
• "Does MES support OPC UA?"
• "What changed in the Data Dictionary in 11.1?"
• "Ask CM GPT about the Collaboration Hub"

Asking for a document, deck, or spreadsheet pulls in references/deliverables.md, which layers the citation and version-labelling rules onto the standard output-format skills.

Behaviour notes

The skill is a faithful port of the tested LibreChat system prompt. Two things were adapted for Cowork:

1. Tool visibility. Cowork shows tool calls in the interface, so the original "do not reveal tool names" instruction cannot hold and was replaced with a rule against pasting the skill contents or narrating internal reasoning. Nothing about search behaviour changed.
2. Deliverables. Cowork can produce Word, PowerPoint, Excel and PDF files, which LibreChat could not. The deliverables guidance is additive — deleting references/deliverables.md and the final section of SKILL.md returns the plugin to pure chat Q&A.

Everything else — the source table, tone, scope boundaries, the three-step userQueryToEmbed construction with its banned-word check, the 5-search hard limit, and the citation format — is unchanged from the tested prompt.

Rollout (admin)

End users cannot add remote MCP servers themselves, so this plugin must reach them through an admin-provisioned channel. Three options:

1. Git-hosted plugin marketplace (recommended)

This repo already is that marketplace — see the root .claude-plugin/marketplace.json, which lists this plugin with a git-subdir source pointing at plugins/cm-gpt. Add the repo to the allowedPluginMarketplaces managed-configuration key:

[{
  "source": "github",
  "repo": "criticalmanufacturing/agent-plugins",
  "ref": "<full 40-character commit SHA>",
  "credentialKind": "userGit",
  "installationPreference": "auto_install"
}]

auto_install and required both demand a full commit SHA — a branch or tag name is refused. Roll out an update by committing the change and pointing ref at the new SHA.

2. HTTPS-hosted marketplace

Serve marketplace.json and a zipped copy of this plugin from the same origin on an internal web server, using "source": "archive" with a sha256 per archive. Use this when devices don't have git. A manifest served from your own inference-gateway or bootstrap-server origin can mark plugins auto_install inside the manifest itself, so publishing a new version needs no configuration change.

3. org-plugins/ directory via MDM

Push this directory to each managed device at /Library/Application Support/Claude/org-plugins/cm-gpt (macOS) or C:\Program Files\Claude\org-plugins\cm-gpt (Windows), with a version.json alongside .claude-plugin/. Bump the version string to trigger a re-sync. Use only when devices can reach neither git nor an HTTPS file host.

Installation preference

Add to .claude-plugin/plugin.json to install without each user opting in:

"installationPreference": "required"

required reinstalls on every sign-in and hides Uninstall; auto_install lets users remove it; omitting it (current state) leaves the plugin available to install manually. Worth leaving as-is for a pilot group, then switching once the behaviour is validated.

Tool policy

Plugin-delivered MCP servers don't carry toolPolicy in .mcp.json. Both CM GPT tools are read-only, so auto-approving them removes a confirmation prompt per search. Set it via orgPluginSettings in managed configuration, keyed on the server name cm-gpt:

{ "cm-gpt": { "toolPolicy": { "search_documentation": "allow", "get_adjacent_chunks": "allow" } } }

Versioning

Current version: 0.1.1. Bump the version field in both .claude-plugin/plugin.json and the portable plugin.json on every change to the skill or connector config, so installed copies can be told apart during rollout on either platform.