Triage Inbox IA — Ton EA pour l'email

What's in this pack

This is the stack for someone whose inbox decides the day: an exec with 200 unread by 9am, a founder doing sales + hiring + investor updates from one address, a single AE running a real pipeline. Not a wishlist of every "AI email" product on Product Hunt. Every pick maps to one of the five jobs an executive assistant would do for you: triage, draft, follow up, schedule, and keep the deliverability clean.

The principle: the agent must hold the credential, not screenshot the screen. That's why the spine of this pack is an MCP server with Gmail and Calendar OAuth scope (pick 1), not a browser plugin that watches you type. With the MCP in place, the other eight picks become composable: a coaching agent to tighten tone, a sales automator for the outbound layer, a CLI mail tool for the second account (Feishu/Lark), a real scheduling endpoint, and a webhook layer that flags bounces before they cost you a sender reputation.

Install in this order (classify → draft → follow up → schedule → anti-spam)

Layer 1 — Classification & inbox ownership

1. Google Workspace MCP — Gmail/Calendar/Drive for Agents — the spine. OAuth into Gmail with messages.list, messages.modify, drafts.create scopes and into Calendar with events.insert. From here the agent can read what came in, label it (AI/Triage/Reply, AI/Triage/Archive, AI/Triage/FYI), and draft into the right thread. Everything below assumes you've done this step.

Layer 2 — Draft quality

2. Claude Code Agent: Communication Excellence Coach — tone calibrator. Once the MCP has drafted a reply, this agent rewrites it to match your voice: "warm but quick," "firm decline without burning the bridge," "executive bullet response with no preamble." Pair as a second pass; never the first.
3. Claude Code Agent: Sales Automator — when the inbound is a lead, not a peer. Cold and warm outbound replies, follow-up sequences, case-study insertions. Use this for the outbound side of the inbox so the Communication Coach can stay focused on internal voice.

Layer 3 — Multi-account bridge

4. Lark CLI Skill: Mail — Email Management — for the second mailbox. Most execs have a Gmail and a Lark/Feishu (or Outlook) account. This Lark CLI gives you draft/send/reply/forward primitives from the agent for the China-side mailbox without spinning up a second OAuth dance.

Layer 4 — Calendar & scheduling

5. Cal.com — Open-Source Scheduling Infrastructure — the public booking endpoint. When the agent replies "happy to meet, here's my link," Cal.com is the link. Self-hostable, webhooks for every booking, conflict checking against the Google Calendar the MCP already owns.
6. Tempo MCP — Calendar and Time Tracking for Agents — the working-hours brain. When the agent goes to propose three slots, Tempo MCP answers "these are the ones that don't break focus blocks or overlap a deep-work timer." Without it the agent proposes 8am Mondays.

Layer 5 — Outbound deliverability & follow-up

7. Composio — 250+ Tool Integrations for AI Agents — the fallback rail. When you need Outlook (no first-class MCP yet), Slack DM as a follow-up channel, or HubSpot/Salesforce write-back from an inbound reply, Composio is the bridge. Don't make it the spine — make it the long tail.
8. Resend Webhooks — Email Delivery Events for AI Agents — the bounce/spam radar. Every transactional reply the agent sends through a custom domain fires delivered/bounced/complained. Route to a mailops channel so a bounced VIP doesn't sit silent for three days.
9. Resend Audiences — Manage Email Lists from AI Agents — the suppression layer. Unsubscribes, hard bounces, and "please stop emailing me" replies feed back into a suppression list the agent reads before it drafts the next outbound. This is the difference between an agent and a spam cannon.

How they fit together

     Inbound mail
          │
          ▼
   Google Workspace MCP (1)
   (read + label + draft)
          │
   ┌──────┼──────┐
   ▼      ▼      ▼
  Reply  Forward  Archive
   │
   ├── internal voice ──► Communication Coach (2)
   └── outbound lead  ──► Sales Automator (3)
          │
          ▼
   (Lark / Feishu side?)
          │
          ▼
   Lark CLI Mail (4)
          │
          ▼
   Needs a meeting?
          │
   ┌──────┴──────┐
   ▼             ▼
  Cal.com (5)   Tempo MCP (6)
  public link   focus-aware slots
          │
          ▼
   Composio (7) — Outlook / Slack / CRM long tail
          │
          ▼
   Resend Webhooks (8) → bounce/spam alerts
          │
          ▼
   Resend Audiences (9) — suppression list
          │
          ▼
   (next outbound checks here first)

The non-negotiable join is the MCP holds the Gmail and Calendar credential, the agent operates through it, and every send writes back to a thread. Lose the MCP and you're back to copy-pasting from a chat window into a browser tab.

Tradeoffs you'll hit

• Privacy — the agent reads every email — once you grant Gmail read scope to the MCP, anything in the inbox is in the context window the next time the agent runs. Pricing leaks, M&A threads, HR letters. Mitigation: scope the MCP to a child label like AI/Triage and a Gmail filter that puts only sender-allowlisted mail there. Don't grant blanket Inbox read on day one.
• Vendor lock on the scheduling link — Cal.com is self-hostable, Calendly is not. If you put calendly.com/you into 300 outbound emails this quarter, switching to Cal.com later means 300 dead links. Stand up Cal.com first, even if you fall back to Calendly during onboarding.
• Model tone drift — the Communication Coach is great until the model gets updated and "warm but quick" turns into 6-bullet preambles. Save your last-known-good prompt to git and diff against the model's new output monthly.
• Auto-reply risk — letting the agent send without review on the first day is how an outage notice gets a chipper "thanks for sharing!" reply. Run for 30 days in draft-only mode (the agent drafts into Gmail Drafts, you press send). After that, whitelist categories case-by-case — never globally.
• Outlook is a second-class citizen here — Composio bridges it, but there is no first-class Outlook MCP in this pack. If you're on Exchange/Outlook only, the spine flips: Composio becomes pick 1 and Google Workspace MCP drops out. The other seven still apply.

Common pitfalls

• The agent auto-replies to a human who hasn't read your inbox — "thanks for confirming, see you at 3pm" lands in the wrong thread because the agent matched on [Re: tomorrow]. Fix: every send requires either (a) an explicit from: agent.allowlist label or (b) human draft-then-send. The mistake is one model run away.
• Misclassification deletes an investor email — if AI/Triage/Archive is a Gmail filter that also marks read, you lose visibility. Use a separate label and a weekly review of AI/Triage/Archive for the first month. Auto-archive only after you've validated 200 decisions.
• GDPR / CCPA on the read-and-summarize path — summarizing a customer email through a US-hosted LLM is a data transfer. EU-residency teams need either an EU model endpoint or a PII redaction step before the prompt. Bake the redaction in early; retrofitting after a sub-processor audit is painful.
• Multi-account bridge is leaky — Lark CLI for the Feishu mailbox, Google Workspace MCP for Gmail, Composio for Outlook — three credential paths means three rotation cadences. Document which scope sits where; the alternative is the agent silently failing on the account you forgot you provisioned.
• No bounce monitoring = silent deliverability collapse — the day you ship a misconfigured DKIM, every outbound email starts going to spam. Without Resend Webhooks routing to a chat channel, you'll find out two weeks later when an investor mentions they "didn't see your update." Wire alerts on day one.