MCP Integrations: How A91I Connects to Everything

Most workflow platforms treat integrations as API wrappers. Call endpoint, pass parameters, get response. A91I is different — every integration is an MCP server, which means the AI in your workflows has genuine understanding of what each tool can do. Here's why that matters.

What Is MCP?

The Model Context Protocol (MCP) is an open standard developed by Anthropic for connecting AI models to external tools and data sources. Instead of describing an API endpoint as a string in a config file, MCP defines a structured protocol: each server exposes a set of tools with typed inputs, descriptions, and schemas.

The AI reads these tool definitions at runtime. It knows what Gmail's search_emails expects, what create_issue in Jira returns, and how to chain them together — without hardcoded logic.

The Traditional Integration Model vs MCP

In a traditional automation platform, you build a node for each API action. The node has a form: fill in the fields, map the outputs, move on. The AI has no access to this — it's outside the execution model.

With MCP, the integration is the tool definition. When an AI workflow node or agent executes, it receives the full list of available tools from all connected MCP servers. It decides which tool to call based on the task description, not a predefined flowchart. This is the difference between automation that follows rules and automation that understands context.

How A91I's MCP Servers Are Built

Every A91I integration is an MCP server — a TypeScript process that communicates over stdio with the A91I backend. Each server uses @modelcontextprotocol/sdk and our shared utilities library (@agenticflow/mcp-shared), which handles authentication, error handling, rate limiting, and retry logic consistently across all 34 integrations.

Each server exposes a set of tools with:

• →A name and description written for AI understanding (not just humans)
• →A Zod schema for input validation with field-level descriptions
• →Standardized error handling via MCPErrorHandler
• →OAuth token management with automatic refresh
• →Rate limiting to respect provider quotas

Tool Discovery and Dynamic Loading

A91I loads tools dynamically at runtime. When you open the workflow editor, the backend queries each connected MCP server for its tool list and transforms the schemas into the UI format. This means new tools appear automatically when servers are updated — no redeployment, no manual registration.

The same dynamic discovery applies to AI agents. When an agent runs, it receives a live list of available tools from all enabled integrations. If you add a new connection mid-session, the agent picks it up on the next invocation.

Credentials and Security

OAuth tokens and API keys are stored encrypted at rest using AES-256-GCM. When an MCP server needs credentials, it calls the A91I backend's internal credentials endpoint — the token is never embedded in the request or logged. HMAC-secured webhook endpoints ensure that incoming triggers can't be spoofed.

All MCP server communication happens over stdio on the same host — credentials never leave the server boundary over the network.

Current Integration Coverage

A91I currently ships with 34 MCP servers covering:

Custom Integrations

Don't see a tool you need? A91I supports two paths for custom integrations:

Custom HTTP Tools — define any REST endpoint as a tool directly in the UI. Name it, describe it, set the method and URL template, and it becomes available to your agents and AI workflow nodes immediately. No code required.

Custom MCP Servers — for power users, build a full MCP server using our open-source template and connect it to A91I. This is the right choice for complex integrations with multiple tools and custom authentication flows.

The MCP standard is growing rapidly. New servers from the community and from our team ship regularly — check the integrations page for the latest coverage.