Appearance
MCP Tools
GhostAI can connect to external services through MCP (Model Context Protocol) servers, giving the AI access to your organization's tools directly in chat. When your admin connects a service — such as a meeting notes tool, CRM, or project tracker — its capabilities appear as toggleable tools in the chat interface.
Architecture
Server Scopes
MCP servers exist at three levels, forming a hierarchy that mirrors GhostAI's multi-tenant model:
| Scope | Who manages it | Visible to |
|---|---|---|
| Cloud | Platform admin | All tenants and users across the platform |
| Tenant (org) | Org admin | All users within the organization |
| User | Individual user | Only the connecting user |
Cloud-level servers are ideal for platform-wide tools (e.g., a shared knowledge base). Tenant-level servers cover organization-wide services (e.g., a company CRM). User-level servers let individuals connect personal tools (e.g., their own Granola, GitHub, or Notion account) with private credentials.
Tool Naming Convention
Tools are namespaced to avoid collisions across scopes:
mcp__{scope}.{server_name}.{tool_name}Examples:
mcp__cloud.KnowledgeBase.search_docs— platform-wide toolmcp__tenant.Granola.get_meetings— org-level toolmcp__user.MyGitHub.list_repos— personal tool
When the AI resolves available tools for a chat, it merges all three scopes. If the same server name exists at multiple scopes, the most specific scope wins (user > tenant > cloud).
Personal MCP Servers
Users manage their own servers from Settings > MCP Tools. Personal servers support the same connection modes (URL and Command) and authentication methods (bearer token, OAuth) as org-level servers. Key differences:
- Credentials are private — org admins cannot view or export personal tokens. The admin panel shows only aggregate connection counts per user.
- Policy floor — personal servers inherit the organization's minimum PII scan depth. Users cannot configure a weaker policy than their org requires.
- Thread sharing — when a thread is shared with another user, tool call results from personal servers are redacted. The recipient sees that a tool was called but not the response data.
- Force-disconnect — org admins can force-disconnect a user's personal server for incident response without accessing credentials.
How It Works
When an MCP server is connected and enabled, its tools automatically appear in the chat toolbar as a single pill button labeled with the server name (e.g., Granola, Linear, Notion). The AI knows about the available tools and will use them when your question relates to what they can do.
Example: Meeting Notes
If your organization connects a meeting-notes service like Granola:
"What was my last meeting about?" "Show me the transcript from yesterday's standup" "List all my meetings this week"
The AI calls the appropriate tool (e.g., get_meetings, get_meeting_transcript), retrieves the data, and presents the results in the chat.
Example: Project Management
If a project tracker is connected:
"What are my open tasks?" "Show me the status of the Q2 roadmap project"
Toggling MCP Tools
Each connected MCP server appears as a pill button in the toolbar below the chat input. Toggle it on or off like any other tool:
- Blue pill with dot — the server's tools are active and available to the AI.
- Gray pill — the server's tools are disabled for this session.
Toggling a server pill enables or disables all of its tools at once. Your preference is saved and restored in future sessions.
What Happens When a Tool Runs
When the AI uses an MCP tool, you'll see:
- Activity indicator — shows which server and tool is being called (e.g., "Using Granola: get_meetings").
- AI response — the AI processes the tool's output and responds in natural language.
- MCP Protection badge — if PII was detected in the tool call or response, a small badge appears on the message showing how many items were scanned and protected.
Privacy and Security
All MCP tool calls are protected by GhostAI's DataIntel pipeline:
- Inbound scanning — arguments sent to MCP tools are scanned for PII and sensitive data. Detected items are tokenized before leaving the platform.
- Outbound scanning — results returned by MCP tools are scanned and any PII is redacted before being sent to the AI model.
- Audit logging — every MCP tool call is logged with full details for compliance review.
- Admin controls — administrators can enable or disable individual tools per server, set risk-level overrides, and review all MCP activity in the admin panel.
For Administrators
To connect MCP servers and manage tools, see the MCP Gateway admin guide. Key admin capabilities:
Connecting and Editing Servers
Add MCP servers via the admin panel (Settings > MCP Connections). Both URL-based and command-based servers are supported. Once connected, you can edit a server's endpoint, data sensitivity level, and authentication token without removing and re-adding it — click the server row to open the detail panel and use the edit option.
Per-Tool Enable/Disable
Each connected server exposes a set of tools. In the server detail panel, toggle switches let you enable or disable individual tools. Disabled tools are hidden from the AI — they won't appear in the chat toolbar or be available for tool calls. This is useful for restricting which capabilities a service exposes to your users.
DataIntel Scanning Profiles
When a server's tools are first discovered, GhostAI auto-creates a DataIntel scanning profile for each tool with an auto-classified risk level based on the tool's name and data flow. Admins can override the risk level per tool from the DataIntel Scanning Profiles card to increase or decrease scan depth.
Default PII Scanning
All MCP tool calls are scanned for PII even if no explicit security policy has been configured. By default, GhostAI applies an observe-only policy (standard scan depth, warn action) so that PII detections are logged without blocking the tool call. Admins can upgrade this to active enforcement by creating a per-tool or per-server policy in the Tool Security Center.
Supply Chain Monitoring
The Tool Security Center's Supply Chain & Shadow tab provides continuous monitoring of all connected MCP servers, including transport type, TLS verification, authentication status (bearer tokens and OIDC/OAuth are both recognized), schema fingerprinting for drift detection, and an aggregate risk score. See Tool Security Center for details.
Audit Trail
Every MCP tool call is logged with full details for compliance review, including the specific PII types detected (e.g., PERSON, EMAIL, SSN). Blocked tool calls are also recorded with action=blocked. Export the audit trail as CSV from the MCP Connections admin panel or via the /api/mcp/audit endpoint.