AI Agents That Read and Write Google Workspace, Microsoft 365, and Zendesk

Integration tools are the write-side counterpart to data collectors: agents use them to act on external systems, not just search them. ZenSearch ships 85+ built-in integration tools across Google Workspace (18 tools — Gmail, Calendar, Drive, Docs, Sheets), Microsoft 365 (33+ tools — OneDrive, Outlook, Teams, SharePoint, Planner), Zendesk (14 tools), Airtable (8), and Notion (12). For anything not in the built-in set, agents can call user-defined webhooks or tools hosted on Model Context Protocol (MCP) servers.

This is the bit that turns RAG-with-citations into "agents that do things." A support agent finds the relevant ticket, drafts a response grounded in your internal docs, and updates the ticket. A sales agent pulls the account's activity history, writes a briefing, and sends it to the AE's calendar. The action is part of the deliverable.

Collectors vs Integration Tools

Two different jobs, two different codepaths:

• Collectors (services under services/collectors/) run on schedules, pull documents from source systems, parse them, and index them for hybrid search. They're read-only, bulk, and optimised for throughput.
• Integration tools (in pkg/agent/tools/) are invoked by an agent during a single run. They authenticate as the end user (usually via OAuth), operate at the object level (one email, one calendar event, one ticket), and can both read and write.

The distinction matters for authorisation: a collector ingests documents your admin gave it access to; an integration tool reads and writes as the user who triggered the agent run. That's why every tool call goes through the per-user OAuth token, not the connector's service credential.

Google Workspace (18 Tools)

Gmail read/write (list threads, read message, draft, send, apply label), Calendar (list events, create, update, delete, free-busy), Drive (search, read file content, list folder, upload), Docs (read doc, append, insert), Sheets (read range, update range, append row). Auth is OAuth 2.0 or — for tenancy-wide service accounts with domain-wide delegation — Google Service Account.

A typical agent flow: an automation reads the user's unread mail, groups by thread, drafts responses for the ones that match a "customer follow-up" trigger pattern, and writes each draft back to the corresponding thread for the user's review. Nothing is sent without the user clicking send.

Microsoft 365 (33+ Tools)

Broader surface: OneDrive, Outlook, Teams, SharePoint, Planner, plus Graph-powered tools for org chart, user profile lookup, and calendar free-busy. Auth is OAuth 2.0 through Azure AD, with scopes configured per tool family so a team that grants only Mail.Read doesn't accidentally get Calendars.ReadWrite.

The Teams tools in particular pair with ZenSearch's Teams assistant surface — an agent can post to a specific channel, reply in a thread, or schedule a meeting in the user's calendar, all with the same identity the user logged into ZenSearch with.

Zendesk, Airtable, Notion

Zendesk (14 tools): ticket CRUD, comment threads, help-center article search and create, user and organisation lookup. Agents built for L1 support triage use these heavily — find the relevant tickets, read the knowledge base, draft the response, update the ticket status.

Airtable (8 tools): base list, table read, record CRUD, comment management. Teams that use Airtable as a lightweight CRM or project tracker get an agent that can update the state of records based on other signals.

Notion (12 tools): page read and update, database query, comment management. Useful for agents that maintain running docs — a weekly KPI roll-up agent that updates the same Notion page every Monday.

Custom Webhook Tools

For everything not in the built-in set, ZenSearch supports user-defined webhook tools. Admins register an endpoint with its name, URL, HTTP method, auth type (none, API key, Basic, or custom headers), and a parameter schema. The tool becomes callable by agents in that team, with request timeout (0–300s, default 15s) and response size limits (65KB) enforced by the platform. Credentials are encrypted at rest in credential_secrets.

This is the escape hatch for internal systems — the agent can call your company's custom deploy endpoint, your internal ticketing system, your feature-flag API — without you forking the platform.

MCP Server Support

For third-party tool servers speaking Model Context Protocol, ZenSearch acts as an MCP client. Admins register a server (stdio, SSE, or HTTP) in mcp_servers, the platform negotiates the protocol and exposes whatever tools that server advertises to agents in that team. The MCP ecosystem covers everything from GitHub (Anthropic's reference server), to Linear, to Stripe, to internal servers your team runs — without ZenSearch having to ship each integration natively.

The Shape of What This Enables

An enterprise RAG platform that only reads is half a product. The moment an agent can find the context, draft the action, and take it — subject to approvals and audit logging — AI stops being a search tool and starts being a coworker who can actually finish a task. Integration tools are what make that shift possible.