tracking: BYO-AI distribution — shells, Connect surface, verifications, listings (#2698 follow-ups)

[os-zhuang]

Umbrella tracker for everything that makes the open edition's BYO-AI capability (#2698) reachable by real users and agents. Core protocol work is done and merged: OAuth 2.1 self-serve connect (#2709) + default-on MCP surface (#2712) — every deployment is now a connectable MCP endpoint out of the box. What remains is distribution: shells, product surfaces, verifications, and listings.

Principle (keep us honest): the MCP server is the product. Ecosystem integrations are thin shells over one single source — renderSkillMarkdown() in packages/mcp/src/skill.ts (ADR-0036 Amendment C: ONE generic skill, schema discovered live, no per-app artifacts). Anything that looks like a heavy per-ecosystem SDK is out of scope by design.

Phase 0 — prerequisite fix

• SKILL.md documents only 7 of 9 tools (fixed in fix(mcp): SKILL.md documents the business-action tools (#2714 Phase 0) #2715, merged): list_actions / run_action (business actions, feat(mcp): native business-action execution (list_actions / run_action) #2307) are missing from the generated skill's tool list — agents won't know they can run approvals/conversions directly. Small skill.ts PR; land before any shell repo copies the content.

Phase 1 — product surface + first shell (near term)

• feat(setup): 'Connect an agent' page — MCP URL, per-client connect cards, API keys, SKILL.md download objectui#2363 — Setup "Connect an agent" page: env MCP URL, per-client connect cards, API-key minting, SKILL.md download; includes the framework dependency GET /api/v1/mcp/skill (renderSkillMarkdown has no HTTP outlet today)
• feat(distribution): official Claude Code plugin — one plugin for every deployment #2713 — official Claude Code plugin (thin repo objectstack-ai/claude-plugin: plugin.json + .mcp.json with ${OBJECTSTACK_MCP_URL} + SKILL.md + /objectstack:connect; CI drift-guard against skill.ts)
• One-click install buttons on the Connect page + docs: "Add to Cursor" deeplink (base64 config) and "Install in VS Code" badge — links, not plugins (extend objectui#2363)
• ChatGPT connectors compatibility verification: our remote MCP + OAuth/DCR against the OpenAI client; fix or document any gaps; add the Connect card
• Codex CLI verification: remote HTTP MCP + OAuth maturity check; document config.toml snippet; fallbacks = x-api-key header or npx mcp-remote bridge. No shell repo — Codex has no marketplace mechanism; it reuses the same Agent Skills SKILL.md

Phase 2 — more shells + listings (after Phase 1 lands)

• Gemini CLI extension: thin repo objectstack-ai/gemini-extension (gemini-extension.json + GEMINI.md reusing the same skill content; same drift-guard)
• MCP registry / directory submissions (batch, metadata-only): official MCP registry (server.json, can live in this repo), mcp.so, Smithery, PulseMCP
• objectstack-ai/cloud#786 — Anthropic connector directory: decide the multi-tenant entry shape (account-routed gateway vs docs-first; RFC 8707 aud across hosts), then submit
• Agent-framework recipes in content/docs: ~10-line connect examples for LangChain/LangGraph, Vercel AI SDK, Microsoft agent framework — docs pages, NOT maintained SDK packages

Phase 3 — demand-gated

• n8n community node (n8n-nodes-objectstack, own repo + npm): only real development in the list — typed dropdowns from list_objects/describe_object beat the generic MCP client node for the low-code audience. Start when Phase 1 shows user signal
• Zapier / Make: watch — their native MCP support is evolving; likely free coverage

Repo topology (settled)

Item	Home
skill single source, /api/v1/mcp/skill, registry server.json, recipes	framework (existing)
Connect page + one-click buttons + client cards	objectui (existing)
Directory entry decision + submission	cloud (existing)
Claude Code plugin / Gemini extension	2 new thin repos (clone-to-install mechanics require it)
n8n node	1 new repo + npm, demand-gated

Refs #2698, #2709, #2712.