MCP (Model Context Protocol)
Turn engini workflows and connections into MCP tools for Claude, ChatGPT, Cursor and any MCP client, scoped per server, with every call recorded.
Book a demoYour workflows, as tools for any AI agent.
engini exposes any connection or workflow as an MCP server. Claude, ChatGPT, Cursor or any MCP client calls the tool, engini runs it against your systems of record, and a structured answer goes back. The model decides when to call; the workflow decides how the work is done, step by step, the same way every time.
Cursor, your code
audit log
connections as tools
Build a tool from a workflow. Three building blocks.
In Workflow Studio, a workflow becomes an MCP tool with a trigger and a response. Describe the inputs once, and every MCP client sees a typed tool.
Building blocks
Send a structured reply back to the AI tool that called the workflow.
Run another MCP workflow from a step, with the inputs you map.
Triggers
Starts the workflow when an AI tool calls it, with the inputs you defined, plus who called and from which conversation.
You name the tool, describe what it does and define its inputs, or generate them from a sample JSON or XML payload. The description is what the model reads when it decides to call the tool.
Connect your AI client. One address, signed in through engini.
Every MCP server in engini has its own address. Add it to your client once; the client signs in through engini on first use.
Claude, ChatGPT, Cursor
Create an MCP server in engini with your MCP connection and only the tools you choose. Add its address to your client and sign in through engini. There is no key to copy.
MCP client setup →Python and TypeScript
Hand your agent a toolset of MCP tools. The SDK gives OpenAI and Anthropic models the tools in the format they expect and runs the ones they pick.
SDK guide →For coding agents
The engini CLI connects MCP and runs its tools from a terminal, with machine-readable output, so an AI coding agent can drive it without getting stuck.
CLI guide →Any language
Every MCP tool is an endpoint. Call it with exact inputs from a scheduled job or a sync, with no model in the loop.
API reference →.cursor/mcp.json. Claude Code: claude mcp add.// .cursor/mcp.json: Cursor, connected to your systems through engini
{
"mcpServers": {
"engini": {
"url": "<your engini MCP server address>"
}
}
}
Create the MCP server in engini with your systems' connection and the tools this agent may use, then paste its address. The client signs in through engini on first use. Claude, VS Code and other clients take the same address: see the MCP client setup guide.
Each request can map to one workflow tool that does several steps across systems, in a fixed order, and returns one answer.
What teams build as MCP tools. The model picks when, the workflow does how.
One tool for a whole process
A create-order tool checks the credit limit in NetSuite, creates the order in SAP and posts to Slack. The agent calls one tool; the steps always run in the same order.
Answers from several systems
A balance tool reads the ERP, the bank file share and the CRM, and returns one structured answer.
A legacy procedure as a tool
Wrap a stored procedure in IBM DB2 or SQL Server in a workflow, and agents call it like any other tool, through the On-Prem Agent.
Know who asked
Every call carries the engini user who made it and the conversation it came from, so the workflow can act on it and the record shows it.
Every integration runs both ways, so any of these flows can also run in reverse. These are examples, not the limit: any flow between MCP and another system engini connects is available.
You decide what each agent may do in MCP. engini enforces it on every call.
Only the tools you allow
A toolset lists the MCP tools an agent may call and the connection each one uses. Give one agent read tools only and another the write it needs. engini refuses any call outside the toolset.
Scoped when you set them up
An MCP server exposes only the connections and tools chosen when it is created, and engini shows who uses each server.
Whose permissions apply
Connect a person's own MCP account and engini acts as that person, with their MCP permissions. Use a shared integration account and it acts as that account. Every run is recorded under the user who started it.
Kept out of prompts
Your agent never needs your credentials: engini stores them and adds them when it runs each call, so keys stay out of prompts and tool inputs.
Every call on record
Every call lands in History and the audit log, which you can export or stream to your SIEM.
Test before production
Workspaces keep teams apart, and development, staging and production environments keep a test agent away from live data.
Connect MCP in three steps. Then decide what each agent may do.
Build the workflow
Start with the MCP Tool Trigger, describe the tool and its inputs, do the work in the steps, and end with MCP Tool Response.
Expose it
Create an MCP server in engini with the workflows and connections this agent may use.
Connect the client
Add the server's address to Claude, ChatGPT, Cursor or another client, and sign in through engini.
MCP and engini. Straight answers.
What is the difference between an MCP server for a connection and an MCP workflow?
A connection exposed over MCP gives the agent that system's tools, and the model decides how to combine them. A workflow exposed as a tool does several steps in a fixed order, across systems, and returns one answer. Use both: tools for questions, workflows for processes.
Which AI clients can use it?
Claude, ChatGPT, Cursor, VS Code and other MCP clients. The client signs in through engini with OAuth, so there is no key to copy.
Can I limit what an agent sees?
Yes. Each MCP server exposes only the connections and tools chosen when it is set up. Give a narrow agent its own server.
Does the agent see my credentials?
No. Your agent never needs your credentials: engini stores them and adds them when it runs each call, so keys stay out of prompts and tool inputs.
Is every call recorded?
Yes. Every call lands in History and the audit log, and engini shows who uses each MCP server.