Skip to main content

MCP Tool

The MCP tool connects your flow to an MCP server (Model Context Protocol) and turns the actions it exposes into tools your agent can use. MCP is an open standard, so a single MCP server can bring in a whole bundle of capabilities — file access, database queries, third-party APIs — without a dedicated Flowera integration for each one.

Use it inside an Agent node so the agent can call the server's actions during its reasoning.

Adding it to a flow​

In the Tool node's tool picker, choose Custom MCP, paste the server config, then pick which of the server's actions to enable.

Inputs​

ParameterDescriptionRequired
MCP Server ConfigThe server connection, as JSON. For a remote server, provide its URL (and any headers).Yes
Available ActionsThe actions the server exposes. Click refresh to load them, then enable the ones you want.Yes

Remote (URL) config — the recommended form:

{
  "url": "https://api.example.com/endpoint/sse",
  "headers": {
    "Authorization": "Bearer {{$vars.myApiKey}}"
  }
}
Keep secrets in variables

You can reference custom variables inside the config with {{$vars.name}} — put API keys there instead of hard-coding them into the config.

Connecting and picking actions​

  1. Paste the server config.
  2. Click the refresh icon on Available Actions — Flowera connects to the server and lists everything it offers.
  3. Enable only the actions your flow needs.

If the list comes back empty, check the URL and any auth headers in the config, then refresh again.

Outputs​

Each enabled action becomes a callable tool. When the agent calls one, the tool returns the server's result and the flow continues to the next node. It's a single-output tool.

Example​

Give an agent database lookups via MCP

  1. Add an Agent node and, in its tools, add Custom MCP.
  2. Paste the config pointing at your MCP server's URL, with an Authorization header sourced from a custom variable.
  3. Refresh and enable just the read actions (e.g. query, list_tables).

The agent can now answer data questions by calling the MCP server, without you wiring up each query by hand.

Tips​

  • Prefer remote (URL) servers. A hosted MCP endpoint is the most reliable and portable option.
  • Enable only what you need. Turning on every action makes the tool list noisy and slower for the agent to reason over.
  • Store credentials in variables. Use {{$vars.name}} in headers so keys aren't embedded in the config text.
  • Refresh after config changes. Changing the URL or auth requires a refresh to reload the action list.