Suite de Productivité MCP — Connecte l'IA à tes outils de travail

What's in this pack

This is the suite for the person whose week is half Notion docs, a quarter Linear tickets, and the rest scattered across Slack threads, GitHub PRs, Gmail, and a calendar that lies to them. The goal is not "plug in everything." The goal is to get Claude/Codex/Cursor to read and write the tools you actually live in, one tool at a time, without giving the agent more permission than you'd give an intern on day one.

Every pick here is an MCP server (Model Context Protocol) that exposes tool calls — search a Notion page, create a Linear ticket, draft a Slack reply, read a Gmail thread, comment on a GitHub PR. Most are official servers from the vendor. A few are multi-tool aggregators (Le Chat Connectors, Airweave, Composio) for when you want one MCP that reaches a dozen SaaS apps. Aggregators are useful but they're also a bigger blast radius — they go last in this pack on purpose.

Install order — pick one SaaS, then the next

1. Notion MCP — start here if your team's source of truth is Notion. Search pages, read content, append blocks, manage database rows. Read-only token first; you can always upgrade scopes later.
2. Linear MCP — if you triage tickets daily, this is the highest-leverage pick. Create issues from a Slack thread, comment on issues from a PR, move statuses without leaving the editor.
3. Slack MCP — read channels you're already in, post as yourself, fetch threads for context. Treat the bot scope like a loaded gun: start in DM-only, add channels deliberately.
4. GitHub MCP — repository management, PR review, issue triage. If you're an engineer this is non-negotiable; if you're a PM, the read-only flavor still pays for itself by Friday.
5. MCP Atlassian — Jira + Confluence in one server. Substitute for Asana if your team is Atlassian-shaped (most enterprises are). Scope to one project before you scope to a workspace.
6. Google Workspace MCP — Gmail, Calendar, Drive in one. Read calendar for free; sending email needs an explicit scope upgrade. Don't grant gmail.send until you've watched it draft for a week.
7. Tempo MCP — Calendar + time tracking. The piece Google Workspace MCP can't do: log hours, query worklogs, build weekly summaries. Skip if you don't track time.
8. Le Chat Connectors — Mistral's bundle: Slack + Notion + Drive in a single MCP. Pick this instead of installing 3 separate servers if you don't need the deeper per-tool capabilities.
9. Airweave — context retrieval across 50+ integrations. Use as a unified search/read layer; pair with the official write-capable servers above. Heavier setup, bigger payoff.
10. Composio — 250+ tool integrations including the long tail (HubSpot, Intercom, Salesforce, Trello). The escape hatch when a vendor has no official MCP.

Auth tips — get this right or skip the pack

• OAuth > raw API tokens whenever offered. OAuth tokens are scoped, revocable, auditable. Raw API tokens with admin:write are how blast radii happen.
• Notion: create an internal integration, share it to one parent page first. Database access is per-page; the integration sees only what you explicitly share.
• Linear: personal API key works, but OAuth app with team-scoped permissions is safer for any agent that touches production tickets.
• Slack: bot token (xoxb-) for posting, user token (xoxp-) only if you need to read DMs. Never paste either into a public repo — Slack auto-revokes tokens that hit GitHub, but agents reading config files can still leak them upstream.
• GitHub: prefer fine-grained Personal Access Tokens scoped to specific repos. Avoid classic PATs with repo scope unless you actually need to write to every repo you can see.
• Google Workspace: separate scopes for read-vs-write. Calendar read is harmless; Gmail send is not. Most aggregators ask for the kitchen sink — say no, then add scopes one at a time.
• Atlassian: API token + email pair for Cloud, OAuth for Server. Project-scoped users are a 5-minute admin task that prevents most of the "agent moved 400 tickets" disasters.

Scope safety — a checklist before each install

1. Read-only first. Almost every MCP server has a read-only flavor. Wire it in, prove the agent's call patterns match what you expected, then upgrade to write.
2. One workspace, one project. Don't scope a new MCP to your entire org. Pick the lowest-stakes project you have and scope to it.
3. Watch the first 5 calls. Most MCP clients (Claude Code, Cursor, Codex) log tool calls. Tail them while the agent works for the first hour.
4. Rotate tokens on a schedule. 90 days is plenty. Calendar reminder, not a TODO.
5. Audit who has what. Notion integrations, GitHub apps, Slack bots — each platform has an admin page that lists every connected token. Visit it monthly.
6. Kill switches. Know how to revoke each integration before you install it. "How do I yank this token at 2am" should have an answer.

Workflows this pack unlocks

• Weekly review without tab-switching — Notion + Linear + Calendar in one agent context. Ask: "What shipped, what slipped, what's blocking next week?" and let it pull from the three sources directly.
• PR-to-ticket loop — GitHub MCP reads the diff, Linear MCP creates a follow-up ticket with the PR linked, Slack MCP notifies the right channel. Three tool calls, no human glue.
• Customer thread digest — Slack + Gmail + Notion. Agent reads a customer escalation thread, drafts a Notion postmortem stub, leaves the Gmail reply in drafts for you to send.
• Standup replacement — Calendar + Linear + Slack. Agent reads yesterday's closed tickets, today's calendar, and posts a Slack standup that's actually accurate.

Common pitfalls

• Granting admin:* scopes "just to test" — you'll forget. Start scoped; expand only when needed.
• Installing 5 MCP servers in one afternoon — overwhelms the agent's tool-call budget and makes failures hard to diagnose. One per day, verify, then the next.
• Forgetting that OAuth tokens expire — refresh flow needs to work, or the agent silently degrades to "I don't have access to that tool" answers a month later.
• Aggregators that re-surface every scope as their own — Airweave/Composio/Le Chat each ask for combined scopes. Read what they're actually requesting; it's often more than the sum of the individual integrations.
• Letting the agent loose on a shared Slack channel — set a dedicated #agent-output channel for the first month. Promote to your team channels only after the failure mode is well understood.
• No revocation drill — if you can't revoke all 10 integrations in 5 minutes, you don't have a suite, you have a liability.