MCP Server

Why this exists

Once your team starts using an AI assistant day to day — Claude, Cursor, an internal agent — the assistant still has to guess at what your backend does, because it has no way to see it. It can read code you paste in, but it can't check today's execution logs, trigger a test run, or query the database your workflow talks to.

MCP (Model Context Protocol) is the open standard that closes that gap: it defines how an AI assistant discovers and calls tools exposed by an external system. An MCP Server is Fimaflow speaking that protocol — so instead of describing your project to an assistant, you let it look and act directly, within limits you define.

How it works

Every project can expose an MCP Server. Once connected, the assistant sees a fixed set of tools — not your entire account, not other people's projects, and never raw credentials. Each tool call is logged the same way a manual action would be, so there is always a record of what the assistant actually did.

Built-in tools, available on every MCP Server without extra setup:

• • list_workflows — list workflows in the project, with their status.
• • get_workflow — inspect a workflow's nodes and edges.
• • run_workflow — trigger a run and return its result.
• • get_execution_logs — read the log trail of a past run.
• • query_database — run a read query through the project's Database node.

Setting it up

From your project's settings, open MCP and generate a connection. You get a URL and a token — add both to your assistant's MCP configuration exactly like you would for any other MCP server:

{
  "mcpServers": {
    "fimaflow": {
      "url": "https://mcp.fimaflow.com/p/<project-id>",
      "headers": { "Authorization": "Bearer <token>" }
    }
  }
}

Restart your assistant (or reload its MCP connections) and it should list the Fimaflow tools alongside whatever else it already has access to. See Integrations for the exact steps per assistant (Claude, Cursor, ChatGPT, JetBrains AI).

Example: debugging with an assistant

A teammate reports that Create Order failed overnight. Instead of opening the builder, they ask their assistant:

> "Why did the Create Order workflow fail last night?"

Assistant calls: list_workflows()
  → finds workflow "Create Order" (id: wf_8f21)

Assistant calls: get_execution_logs({ workflowId: "wf_8f21", limit: 1 })
  → 3. SQL   ✗  connection failed: DATABASE_URL is not set

"The SQL node failed because DATABASE_URL wasn't set — looks like
 the .ENV node's value was cleared. Want me to check when it changed?"

The assistant never saw your database credentials — only the error the SQL node already logs (masked, same as in the builder itself). See Debugging.

Custom MCP tools

Beyond the built-in tools, expose any workflow as a named tool of its own by adding an MCP Tool node in front of it — give it a name and a description written for the assistant, not for a human. A create_order tool, once exposed this way, lets an assistant place a real order as part of a larger task, the same way it would call any other tool it has access to.

Write the description for what the tool does and when to use it — that description is the only thing the assistant has to decide whether it's the right tool for the moment.

Permissions & safety

• • A connection is scoped to one project — an assistant with access to it can't see your other projects or workspaces.
• • query_database is read-only by default; a workflow exposed as a custom tool can write, since it's the same workflow you already built and trust.
• • Every tool call appears in the project's execution history, tagged as coming from MCP — indistinguishable from a manual run in every other way.
• • Revoke a connection at any time from the same MCP settings panel; already-issued tokens stop working immediately.

Limitations

• • Available on the Pro plan and above — see AI Credits for how usage is measured on your plan.
• • One MCP Server per project; there is no account-wide server that spans every project at once.
• • Tool calls run synchronously — a long-running workflow will make the assistant wait for the result, same as running it manually.