Skip to content

AI agents (MCP)

Kontext is an MCP server (Model Context Protocol): an AI agent like Claude works with the mails, appointments, contacts and tasks of the person who connected it – with exactly their rights, shared mailboxes and “read only” included.

What an agent can do

  • Look things up: “What happened last with Hoffmann?” – contact, latest mails from all mailboxes, appointments before and after, open tasks; search mails with the language of the search field (von:, in:info@, ist:ungelesen …), read conversations, list appointments and tasks.
  • Create and change, if you allow it: drafts (also as a reply with a quote), appointments with an invitation to colleagues, contacts, tasks.
  • Never outside on its own: there is no deleting and no mail to anyone outside the company. A reply waits in “Drafts” until someone sends it in Kontext. To an appointment, the agent only invites colleagues: they get the invitation and the appointment in their calendar as if from Kontext; it refuses an address from outside, which you invite yourself in Kontext.
  • Log: every change is listed with the connection under “Recent changes”; at most 100 changes per hour.

Connecting

  • Claude (app, desktop, claude.ai): add https://kontext.firma.de/mcp as a connector and sign in in the browser (OAuth 2.1); Kontext asks what the agent may do, and for the password.
  • Claude Code, scripts and other clients: create an access key under Settings → AI agents and pass it as Authorization: Bearer kontext_….

Every connection has a session of its own and can be disconnected on its own; when an employee is blocked, their agents are disconnected too. There is nothing to set up on the server (Installation → AI agents). The tools in detail: Architecture → Agents.

Assistant in the chat

It also works without an agent of your own: at the top of the chat is the Assistant. You ask it in words (“What happened lately with Keller Elektro?”, “Write Ms Keller that we’ll come on Tuesday”); it uses the same tools as a connected agent, with the rights of the person asking, and on request creates drafts, appointments, contacts and tasks; what it created is listed under its answer. Like an agent, it sends nothing outside the company and deletes nothing: you send drafts yourself, and it invites only colleagues to appointments.

  • Setting up: the Firmenverwaltung enters a Claude API key (console.anthropic.com) under Settings → Chat and picks the model (Installation → AI agents). Without a key there is no assistant.
  • Cost: Anthropic bills by use; a question with a few lookups usually costs a few cents with the default model (Claude Sonnet 5.5).
  • Data: the question and what the assistant looks up for it (mail text, appointments, contacts) go to Anthropic; this is said under the input field. Kontext does not store the conversation; it stays in the open tab until you reload it, sign out or choose “New conversation”.