mcphub-operator

name	mcphub-operator
description	Operate MCPHub groups, endpoints, compression, and CLI from repo registry. Use for hub preflight, group picking, tunnel vs local exposure. NOT harness sync.
license	MIT
metadata	{"author":"wyattowalsh","version":"1.0.0"}
user-invocable	true
argument-hint	[preflight|groups|compression|tunnel|cli|doctor]

MCPHub Operator

Repo-managed MCPHub control plane: workflow groups, bearer endpoints, optional compression, ChatGPT tunnel policy. Smart Routing stays off in tracked config.

NOT for: editing the agent-stack sync script, live skills-sync apply installs, or PostgreSQL/pgvector smart-routing setup (use maintainer OpenSpec + explicit user request).

Dispatch

$ARGUMENTS	Action
(empty)	Show group picker summary, preflight command, and reference index
preflight / doctor	Run bash scripts/preflight.sh from skill dir; stop on fail
groups / picker	Load references/group-picker.md
compression	Load references/compression-opt-in.md
tunnel	Explain ChatGPT tunnel group-only policy + .env.mcphub tunnel vars
cli	Load references/cli-cheatsheet.md

Operator Contract

preflight

1. From repo root (or pass --cwd to bundled doctor), run:

bash skills/mcphub-operator/scripts/preflight.sh

2. Stop when JSON ok is false or any check with status: fail touches settings parity, smart routing, or client profiles.
3. Treat warn as advisory (hub not running, bearer unset locally).

groups

1. Default local harnesses attach harness only, with per-server endpoints available but disabled by default.
2. Treat harness as the smallest high-signal set for context-bloat control: Brave Search, DuckDuckGo Search, DDGS (metasearch/news), Context7, DeepWiki, llms.txt catalog (llms-txt-explorer list_websites only), Fetch, and package metadata. Use check_website only via opt-in workflow groups. Prefer DDGS snippets; avoid fetch_full_content unless the user needs full pages.
3. Use daily for routine opt-in expansion; use coding, research, review, or release for scoped agentic dev flows.
4. Escalate to capability groups (web-search, web-read, docs, repo, browser, media, notebooks, design, productivity, accounts, references) when the task needs that surface.
5. Use reasoning only for hard problems; keep reasoning-lab experimental.
6. Attach account-backed, live-browser (Chrome DevTools and Scrapling headless scrape sessions), heavy, or experimental groups only with explicit user intent.
7. Never attach tunnel to local harness configs unless testing the remote consumer shape.

compression

1. Tracked mcp_settings.json keeps toolResultCompression.enabled: false.
2. Opt-in locally via .env.mcphub only after user approves; see compression reference.
3. Disable compression when debugging structured tool JSON or investigating truncation.

tunnel

1. ChatGPT remote MCP uses tunnel group only (https://mcp.w4w.dev/mcp/tunnel when tunnel enabled).
2. Do not widen ChatGPT to the global /mcp route or per-server endpoints without explicit user request.
3. Tunnel credentials stay in .env.mcphub / ~/.cloudflared/ — never commit.

cli

1. Prefer repo recipes: just mcphub-generate, just mcphub-generate-check, just mcphub-validate, just mcphub-doctor, just mcphub-smoke.
2. Use upstream MCPHub CLI for live hub inspection when the process is running; see CLI reference.

Maintainer Loop

just mcphub-generate
just mcphub-generate-check
just mcphub-validate
uv run python skills/mcphub-operator/scripts/check.py

Regenerate settings after editing config/mcp-registry.json groups or servers.

References

• Group picker
• CLI cheat sheet
• Compression opt-in
• Public maintainer doc: docs/ai-tools/mcphub.md