MCP — the agent's tools
Updated 2026-09-16
The agent's tools arrive over MCP (Model Context Protocol). They are the same for every engine: a new tool on the platform is seen by any agent without touching its machine.
Where they come from
How they are named
Inside the model each tool carries its MCP server's prefix: ghosty__doc_read, ghosty__form_create, ghosty__reminder_create. If you tell the agent "use doc_read" it finds it anyway: the prefix belongs to the transport, not the name.
Families in Teams: chat_*, doc_*, form_*, reminder_*, email_send, memory_*, brand_*, prospect_*.
Scope per channel
Scope is decided by the turn, not the agent:
- DM or private channel: all the workspace's tools and the invoker's connectors.
- Public channel (WhatsApp): no personal connectors and no destructive actions; the agent can escalate to a person.
- Editor over ACP: the machine's tools plus the workspace's if the agent is activated in one.
An agent can only do what the person who invoked it could do.
Permissions
Some actions ask for confirmation (deleting, sending to third parties). The client receives a permission event with options and answers; if nobody answers within ten minutes it is denied. See .
Bring your own MCP servers
With the you give your agent its own MCP servers (http or stdio). The list is replaced whole and the agent restarts:
Only on engines with their own machine (Ghosty · Lite and Goose). A stdio server needs the binary to exist on the machine (Node and Python are there).