Workchats MCP server

The Workchats MCP server lets an AI assistant work in Workchats as you. MCP is the Model Context Protocol, the open standard that Claude, ChatGPT, Cursor, VS Code and other assistants use to call tools. After you connect an assistant, you can ask it to:

• catch you up on unread messages and @mentions
• search your groups and channels, and your direct messages when you ask
• read, create and update tasks, and turn a message into a task
• list your meetings, summarise a meeting and list its action items
• find files, and ask the company Filing Cabinet a question
• write a reply for you to approve

The assistant is not a bot. It sees only what you see in Workchats, and it can do only what you can do.

​

Addresses

Paste the MCP URL of your environment into your assistant. You need nothing else: no API key and no App registration. Add /readonly to the URL, for example https://mcp.workchats.com/mcp/readonly, to connect an assistant that can only read. See Read-only URL and toolsets.

​

Before you start

• Your company must turn it on. AI assistants are off in every company until a company admin turns them on. See For company admins.
• You must be an employee. Guests can’t connect. Your account must be active.
• One connection, one company. If your email belongs to more than one company, you pick one company when you approve. To use a second company, add the server again in your assistant and pick the other company.

​

How it works

1. You paste the MCP URL into your assistant.
2. The assistant opens Workchats in your browser. You sign in, pick a company and approve.
3. The assistant calls Workchats tools for you. Each call checks your access again.
4. To send a message, the assistant first writes a draft. You see the draft in the assistant and in your Workchats message box. The assistant sends it only after you say yes.

​

What an assistant can and cannot do

It can:

• read what you can read: your direct messages, groups, channels, tasks, meetings, calls, files, events, polls, announcements and notifications
• create tasks, task comments, subtasks, meetings, polls, direct messages and group chats as you
• react to, pin and mark messages read, answer meeting invitations, and set your status
• draft a message, and send that draft after you approve it

It cannot:

• send a message in one step. Every message starts as a draft. See Drafts and the “via” label.
• edit or delete a message
• use admin power. A connection always acts as a regular employee, also when you are a company admin. It sees no contact details that people hide from regular employees, and it has no admin tools.
• choose a company in a tool call. The company always comes from the connection.

​

Privacy and safety

• Your company sees the activity, not the content. The company’s activity log records each tool call: who, which assistant, which tool, which item, the result and the time. It never stores message text, tool arguments or tool results.
• Text from other people is marked. Tool results wrap text that other people wrote, such as messages, task titles and file names, in an <untrusted_content> element. The server tells the assistant to treat that text as data and never to follow instructions in it. This makes prompt injection harder. It does not make it impossible, so read a draft before you approve it. The one exception is the short title of a search or fetch result, which ChatGPT shows as a label. See Search and fetch.
• Direct messages stay behind their own tool. General search never returns a direct message. The assistant must call search_dms, or search_messages with type set to dm, so you can see when it reads them.
• You can disconnect at any time. A disconnected assistant’s next request fails. Your company admin can also disconnect any assistant. See Revocation for every way a connection ends.

​

Next