Fig MCP Server

The Fig MCP (Model Context Protocol) Server enables AI tools such as GitHub Copilot CLI, Claude Desktop, Cursor, and other MCP-compatible clients to interact with your Fig configuration management system.

Experimental

This integration is experimental and may not work as expected. Create a ticket on GitHub if a problem is found.

Features

• Configuration Analysis: Ask AI to review settings across all clients, identify patterns, and suggest improvements
• Anomaly Detection: Let AI spot unusual values, inconsistencies, or misconfigured settings
• Event Log Queries: Query event logs and setting change timelines through natural language
• Session Monitoring: View active client run sessions with health, memory, and version info
• Setting History: Review the value change history of any setting
• Operational Management (opt-in): Update setting values, manage webhooks, lookup tables, users, and more
• Configurable Safety: Each operation category can be independently enabled or disabled — read-only by default

Architecture

Fig.Mcp is a standalone ASP.NET Core application that proxies requests to the Fig API over HTTP. It does not connect to the database directly. This design provides:

• Deployment independence — deploy and scale separately from the API
• Isolated attack surface — the MCP server runs in its own process
• Simple operation gating — disabled tools are simply not registered, so the AI never sees them
• Easy testing — mock HTTP responses without needing a full API bootstrap

┌──────────────┐     MCP Protocol      ┌──────────────┐    REST/HTTP     ┌──────────────┐
│   AI Tool    │ ◄──────────────────► │   Fig.Mcp    │ ◄──────────────► │   Fig.Api    │
│ (Copilot,    │   (stdio or HTTP)     │  MCP Server  │   (JSON + JWT)   │              │
│  Claude,     │                       │              │                   │              │
│  Cursor)     │                       └──────────────┘                   └──────────────┘
└──────────────┘

Transport Modes

Fig.Mcp supports two MCP transport modes:

Mode	Use Case	How It Works
stdio (default)	Local AI tools (Copilot CLI, Claude Desktop, Cursor)	AI tool launches fig-mcp as a subprocess
HTTP	Remote/server-side AI agents, hosted tool platforms	Runs as an HTTP server with Streamable HTTP + SSE

Configuration

All settings are configured via appsettings.json or environment variables.

Basic Settings

{
  "McpSettings": {
    "FigApiBaseUrl": "https://localhost:7281",
    "Username": "admin",
    "Password": "your-password",
    "Transport": "stdio"
  }
}

Setting	Description	Default
FigApiBaseUrl	URL of the Fig API	https://localhost:7281
Username	Fig user to authenticate as	admin
Password	Password for the Fig user	(empty)
Transport	stdio or http	stdio

Environment variable override format: McpSettings__FigApiBaseUrl, McpSettings__Password, etc.

Tool Gates

Each tool category can be independently enabled or disabled. By default, all reads are enabled and all writes are disabled.

{
  "McpSettings": {
    "ToolGates": {
      "ReadSettings": true,
      "WriteSettings": false,
      "ReadEvents": true,
      "ReadSessions": true,
      "ReadHistory": true,
      "ManageClients": false,
      "DeleteClients": false,
      "ManageUsers": false,
      "ManageWebHooks": false,
      "ManageLookupTables": false,
      "ManageScheduling": false,
      "ManageTimeMachine": false,
      "ExecuteCustomActions": false,
      "ImportExportData": false,
      "ManageConfiguration": false
    }
  }
}

Disabled tools are not registered with the MCP server — the AI never sees them and cannot invoke them.

Setup

GitHub Copilot CLI

Add to .github/copilot/mcp.json in your repository:

{
  "servers": {
    "fig": {
      "command": "dotnet",
      "args": ["run", "--project", "/path/to/src/mcp/Fig.Mcp"],
      "env": {
        "McpSettings__FigApiBaseUrl": "https://localhost:7281",
        "McpSettings__Username": "admin",
        "McpSettings__Password": "your-password"
      }
    }
  }
}

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "fig": {
      "command": "dotnet",
      "args": ["run", "--project", "/path/to/src/mcp/Fig.Mcp"],
      "env": {
        "McpSettings__FigApiBaseUrl": "https://localhost:7281",
        "McpSettings__Username": "admin",
        "McpSettings__Password": "your-password"
      }
    }
  }
}

HTTP Mode (Remote Agents)

Run the MCP server as an HTTP service:

McpSettings__Transport=http dotnet run --project src/mcp/Fig.Mcp -- --urls http://localhost:3001

Docker

docker run -e McpSettings__FigApiBaseUrl=http://fig-api:8080 \
           -e McpSettings__Username=admin \
           -e McpSettings__Password=your-password \
           -e McpSettings__Transport=http \
           -p 3001:8080 \
           mzbrau/fig-mcp:latest

Or using docker-compose (see docker-compose.yml in the repository root).

Available Tools

Read-Only Tools (enabled by default)

Tool	Gate	Description
ListClients	ReadSettings	List all registered clients with settings
GetClientDescriptions	ReadSettings	Lightweight client name + description list
GetEvents	ReadEvents	Query event logs by time range
GetEventCount	ReadEvents	Total event log count
GetClientTimeline	ReadEvents	Setting change timeline for a client
GetSettingHistory	ReadHistory	Value history for a specific setting
GetLastChanged	ReadHistory	Last-changed timestamps across all clients
GetRunSessions	ReadSessions	Active client sessions with health info
GetApiVersion	ReadSettings	Fig API version
GetApiStatus	ReadSettings	API health status
ListLookupTables	ReadSettings	All lookup tables
ListWebHooks	ReadSettings	All webhooks
ListWebHookClients	ReadSettings	Webhook clients
ListCheckPoints	ReadSettings	Time machine checkpoints
GetCheckPointData	ReadSettings	Checkpoint payload
ListDeferredChanges	ReadSettings	Scheduled setting changes
GetCustomActionStatus	ReadSettings	Custom action execution status
GetCustomActionHistory	ReadSettings	Custom action execution history
GetDeferredImports	ReadSettings	Pending deferred imports
GetClientRegistrationHistory	ReadSettings	Client registration history

Write Tools (disabled by default)

Tool	Gate	Description
UpdateSettingValues	WriteSettings	Update setting values for a client
ToggleLiveReload	WriteSettings	Enable/disable live reload for a session
RequestClientRestart	WriteSettings	Request a client restart
ChangeClientSecret	ManageClients	Rotate a client's secret
DeleteClient	DeleteClients	Delete a client
CreateLookupTable	ManageLookupTables	Create a lookup table
UpdateLookupTable	ManageLookupTables	Update a lookup table
DeleteLookupTable	ManageLookupTables	Delete a lookup table
CreateWebHook	ManageWebHooks	Create a webhook
UpdateWebHook	ManageWebHooks	Update a webhook
DeleteWebHook	ManageWebHooks	Delete a webhook
TestWebHookClient	ManageWebHooks	Test webhook connectivity
ApplyCheckPoint	ManageTimeMachine	Restore to a checkpoint
UpdateCheckPointNote	ManageTimeMachine	Add note to checkpoint
RescheduleDeferredChange	ManageScheduling	Reschedule a deferred change
DeleteDeferredChange	ManageScheduling	Cancel a scheduled change
ExecuteCustomAction	ExecuteCustomActions	Execute a custom action
ListUsers / GetUser / CreateUser / UpdateUser / DeleteUser	ManageUsers	User management
ExportData / ImportData / ExportValuesOnly / ImportValuesOnly	ImportExportData	Data import/export
GetConfiguration / UpdateConfiguration	ManageConfiguration	API configuration

Example AI Interactions

Once connected, you can ask your AI tool questions like:

• "List all registered Fig clients and summarize their settings"
• "Are there any settings with empty or default descriptions that should be documented?"
• "Show me the event log for the last 24 hours — are there any unusual patterns?"
• "Which clients have the most settings? Are any over-configured?"
• "Compare the settings between the production and staging instances of ServiceX"
• "Show me the value history for the DatabaseConnectionTimeout setting"
• "Which clients haven't reported a run session recently?"

Security Considerations

• The MCP server authenticates to the Fig API using a pre-configured user's credentials
• It inherits the permissions of that user — a ReadOnly user cannot perform write operations
• Tool gates provide an additional layer of protection at the MCP level
• For production, use environment variables or secret providers for credentials — never commit passwords
• In HTTP transport mode, consider network-level access controls (firewall, VPN, etc.)