MCP security, permissions and limits

How access works when you connect an AI assistant to Herodesk: roles, consent, audit log, rate limits, SSO and what MCP can't do.

Written by Anders Eiler
Last updated 2026-11-13
On this page

The Herodesk MCP server is built so that an AI assistant can never do more in Herodesk than the person who connected it. This article explains how access works, what is logged, which limits apply and what you should know before you connect.

Access is personal

  • One user, one workspace. Every connection belongs to the user who signed in, on one workspace. To use several workspaces, connect once per workspace.
  • Same permissions as in the app. Your assistant gets your role (agent, manager or admin) and your team inbox access. If an admin changes your role or teams, the change applies to your assistant on its next request.
  • No shared keys. There is no workspace-wide API key for MCP. Your password is only entered on Herodesk's own sign-in page; the assistant only receives a token.
  • You approve each connection. The consent screen shows which assistant is asking, for which Herodesk user, and whether it wants to only read or also make changes. Self-registered clients are marked as Unverified third-party app – only continue if you started the connection yourself.

What each role can do

RoleThrough MCP
AgentRead and work on conversations in their inboxes, contacts, companies, quick replies (their own) and help center content. Agents can see the list of AI agents and their topics, but not change them
ManagerEverything an agent can, plus tags, SLA policies and schedules, AI agents (setup, where they answer, knowledge, skills, topics, instructions, playground tests, goals and fine-tuning), deleting contacts and deleting help center content
AdminEverything a manager can, plus users, teams, webhooks, the audit log and the deletion log

Lite seats can read, add internal notes, tag and assign, but can't send replies to customers or start new conversations.

What is logged

  • Changes made through MCP are recorded in the audit log under your name, the same as changes you make in the app.
  • Conversation activity (status changes, assignments, tags) is shown in the conversation's log as done by you.
  • In the webhook delivery log, events caused by an AI assistant have the source MCP.
  • Testing an AI agent as a live chat test customer is recorded in the audit log, because the agent then looks up that customer's real orders.
  • Connecting and revoking AI clients is logged too.

Revoking access

Each user can see and revoke their own connected assistants under Profile → Security → Connected AI clients. Revoking takes effect on the next request.

Access is also removed automatically when:

  • you change or reset your password,
  • an admin turns on enforced single sign-on (SSO),
  • your user is deactivated or your workspace is closed,
  • the connection hasn't been used for 30 days.

There is no admin overview of colleagues' connections. To cut off a user completely, deactivate the user or enforce SSO.

Single sign-on and two-factor authentication

  • Enforced SSO: when your workspace enforces SSO, MCP is not available, except for users an admin has made exempt from SSO.
  • Required two-factor authentication: users who haven't set up two-factor authentication after the grace period must do so before they can connect.

Rate limits

To protect your workspace and your customers, requests are limited per user. Reading counts less than changing something.

PlanApproximate limit per user per minute
Basic and free trialAbout 30 lookups or 10 changes
PlusAbout 300 lookups or 100 changes
ProAbout 600 lookups or 200 changes

Messages to customers (replies and new conversations) also count against your workspace's hourly send limit, shared with the Herodesk API. It grows with your number of paid seats, so a runaway assistant can't flood your customers. If you hit a limit, the assistant is told how long to wait.

Testing an AI agent counts as two changes, and your whole workspace can run 60 AI agent tests per hour through MCP and the API together. Tests in the Playground inside Herodesk don't count.

Your data and the AI provider

When you ask your assistant something, the Herodesk data it reads – conversations, customer details, attachments – is sent to the AI provider you use (for example Anthropic, OpenAI or xAI) and handled under that provider's terms. Check that this fits your company's data policy before you connect, and only connect assistants you trust.

Tip: when your assistant reads attachments, it receives a download link. Treat AI chats that contain customer data like any other place you store customer data.

Good to know

  • Replies are sent immediately, as plain text, in your name, and can't be undone. A reply also reopens the conversation.
  • AI agent changes apply right away. Instructions, topic modes and channel settings change how the AI agent answers your customers from the next conversation. On a channel set to Auto, the agent replies without a human checking.
  • Not available through MCP: snoozing, marking as spam, merging or deleting conversations, and creating inboxes, channels, rules or smart views.
  • Search: quick search matches conversation subjects, contact names and e-mails, and company names – not the text inside messages. Contact search by e-mail needs the exact address.
  • Long conversations: your assistant reads the newest 50 messages of a conversation.
  • Attachments: your assistant can look at up to 5 images at a time. Other files are available as download links.
  • Sending files and Cc: up to 25 files and 25 MB per message, and Cc/Bcc on e-mail only. See Send e-mails with attachments and Cc through MCP.