Zapier MCP — Sample Page (agent triggers real app actions)

Goal: a sample page where rabbit's agent, connected to Zapier MCP, executes real SaaS actions (send Gmail, create a Notion page, post a Slack message) from a natural-language instruction — no custom per-app integration code.

Source: jay's Zapier MCP one-liner summary, pasted in session 2026-07-17 (no URL provided — this doc is the canonical copy). Positioning: the consumer counterpart of LLM Day 39 (building an MCP server yourself) — start by consuming someone else's MCP.

1. What Zapier MCP is

• Zapier exposes its 8,000+ app integrations as MCP tools. An LLM connected to the server can trigger those actions as ordinary tool calls.
• Setup is config, not code: create a Zapier MCP endpoint (per-account URL + auth), pick which actions are allowed, connect the client.
• For jay concretely: a post-processing pipeline like morning report done → summary to Notion → Gmail notification to self becomes assembly, not integration work.

2. Sample page design (/etc/zapier or a new ETC subpage)

• UI: a minimal panel — connection status → allowed-tools list (fetched from the MCP server) → a prompt box ("send me an email saying hi") → execution log showing the tool call + arguments + result.
• Wiring (server-side, key never in the browser):
  • lib/ai.ts already abstracts providers; add an MCP client path. Two options:
    1. Anthropic API MCP connector — pass the Zapier MCP URL in the API request (mcp_servers), let Claude call tools remotely. Least code; ties the demo to the Anthropic provider.
    2. Generic MCP client in the route (@modelcontextprotocol/sdk) — /api/zapier connects, lists tools, forwards the model's tool calls. Provider-agnostic, more code; reuses thinking from §5 (KB-over-MCP), where rabbit is the MCP server — here it's the client.
• Env: ZAPIER_MCP_URL (+ auth token) as server secrets, same handling as the AI-chat keys in design §2 (encrypted/env, never client-side).
• Safety rail: allowlist only harmless actions in the Zapier console for the demo (e.g. "Gmail: send email to self", "Notion: create page in a sandbox DB") — an agent-triggerable Gmail is a footgun if left broad.

3. Concrete demo scenarios (pick 1–2)

1. Email-to-self: prompt → agent calls Zapier's Gmail send tool → mail arrives at linked0@gmail.com. Smallest end-to-end proof.
2. Notion note: "summarize this market snapshot to Notion" → agent pulls /api/indices data → creates a Notion page. Shows chaining a rabbit-internal source with an external action.
3. (stretch) Morning-report pipeline: wire the summary+notify flow from the source note — report text → Notion page → Gmail notification, both via one agent turn.

Suggested first cut: scenario 1 with wiring option 1 (Anthropic MCP connector) — smallest possible loop, ~0.5–1d including the Zapier-side setup.

Status

Backlog / to do — not yet scheduled. Natural sequencing: after the AI Chat KB-via-MCP+RAG work (see ai-chat.md) so the MCP client/server concepts land together, but it's independent enough to pull forward as a quick win.