MCP Servers Overview
How FlowTrux uses Model Context Protocol servers to connect AI agents with external tools
MCP (Model Context Protocol) is a standardized way for AI models to discover and use external tools. FlowTrux runs MCP servers as child processes that communicate over JSON-RPC via stdio, giving Agent nodes access to tools like Slack messaging, database queries, file operations, and more.
Architecture
Each MCP server runs as an isolated child process on the FlowTrux server. When an Agent node needs to call a tool:
- FlowTrux starts (or reuses) the MCP server process.
- The server advertises its available tools and their parameter schemas.
- The AI model decides which tool to call and with what parameters.
- FlowTrux sends the tool call to the server via JSON-RPC over stdin/stdout.
- The server executes the tool and returns the result.
- The result is passed back to the AI model.
Connection Pool
MCP server connections are pooled at the workspace level:
- Connections are reused across concurrent workflow executions within the same workspace.
- Idle connections are automatically cleaned up after 5 minutes.
- The pool holds a maximum of 50 entries.
- Each workflow execution gets its own
MCPManagerscoped to workspace-configured servers.
This means the first execution may take slightly longer to start (while connections are established), but subsequent executions reuse existing connections.
Built-in Servers
FlowTrux includes the following built-in MCP servers:
| Server | Description | Documentation |
|---|---|---|
| Slack | Send messages, manage channels, upload files | Slack |
| Telegram | Send messages and documents via Telegram Bot | Telegram |
| Google Workspace | Sheets, Gmail, Calendar, Drive | Google Workspace |
| Stocks | Financial market data (Polygon.io, Finnhub) | Stocks |
| Files | Organization file storage | Files |
| Filesystem | Local server disk file operations | Filesystem |
| SSH | Remote server command execution and file transfer | SSH |
| Twilio | SMS, voice calls, WhatsApp messaging | Twilio |
| YouTrack | JetBrains YouTrack issue tracking | YouTrack |
| PDF document generation (markdown, reports, tables) | PDF Generator | |
| Notion | Pages, databases, blocks, comments | Notion |
| GitHub | Issues, PRs, repos, Actions workflows | GitHub |
| Jira | Issues, projects, sprints, transitions | Jira |
| Wialon | Telematics: units, messages, geofences, reports, ACL | Wialon |
External Servers
You can add any MCP-compatible server. See External Servers for details.
Setup
1. Add an MCP Server
- Navigate to Settings > MCP Servers (requires OWNER role).
- Click Add Server.
- Choose a built-in server type or select External for a custom server.
- Enter the required configuration (API keys, tokens, etc.).
- Click Save.
All credentials are encrypted before being stored in the database.
2. Assign to a Workspace
MCP servers are not automatically available in all workspaces:
- Go to Settings > Workspaces.
- Click Edit on the target workspace.
- Under MCP Servers, select which servers this workspace can use.
- Click Save.
3. Use in a Workflow
Once assigned to a workspace, MCP tools become available in:
- Agent nodes -- Add tools in the Inspector panel so the AI can call them.
- Action nodes -- Select "MCP" as the action type to call a specific tool directly.
Security
- Each MCP server runs in its own child process, isolated from other servers.
- Credentials are injected as environment variables into the child process, never passed through the AI model.
- Server processes receive a 3-stage shutdown sequence (stdin close, SIGTERM, SIGKILL) to prevent zombie processes.
- File and organization data are scoped -- servers can only access resources belonging to their organization and workspace.
Plan Limits
Your organization's plan may limit the number of MCP server configurations. Check Settings > Plan & Usage to see your current limits.