Built-in nodes

Nodes return { success, output?, data?, error? }. output is a string that downstream templates can use. data carries structured values.

input

Exposes the run input. output is the input as JSON, and data is the input object. Templates can read the input directly as {{input.field}}, so this node is optional.

prompt

string

required

Text with {{path}} placeholders. The rendered text is the node’s output.

llm

One chat call through a named provider. There is no loop and no tool use.

string

required

Provider id. It must support chat.

string

Overrides the provider’s default model.

string

System prompt.

"summarize" | "classify" | "extract" | "rewrite" | "translate"

System prompt from a built-in preset. system wins if both are set.

string

Template for the user message. Defaults to the upstream outputs joined by blank lines.

number

Truncates the message before the call. Unset means no truncation.

decision

A typed decision. See decisions.

string

required

Provider id. It must support decide.

"choice" | "score" | "noul"

required

Decision type.

string

required

The question. Templated.

string

What to evaluate. Templated. Defaults to the whole run input as JSON.

string[] | string

Choices for choice, ordered levels (at least two) for score. A single {{path}} pulls a list from the run input.

number

Between 0 and 1. Gates choice decisions to none_of_these.

string

Overrides the provider’s default model.

http

A generic HTTP call. url, headers and body are templated. A response outside 200-299 fails the node.

string

required

Request URL.

"GET" | "POST" | "PUT" | "PATCH" | "DELETE"

default:"GET"

HTTP method.

Record<string, string>

Request headers.

unknown

Sent as JSON. Strings are sent as they are.

The result’s data is { status, body }, where body is parsed JSON when possible.

mcp

Calls a tool on another MCP server. See MCP.

string

required

Id of a server under mcpServers. Loading fails for an unknown id.

string

required

Tool name. Templated.

object

Tool arguments. Templated.

string[]

Tools this node may call. Required when tool is a template.

output is the tool’s text. data is its structured content, or the text parsed as JSON. A tool error fails the node.

output

Marks the flow result.

string

Rendered as the output. Defaults to the upstream outputs joined by blank lines.

data is the upstream node’s data, or a map by node id when several nodes are live. The run result exposes this node’s result as output.

Add your own node

import { z } from "zod";
import { defaultRegistry } from "@milfordai/core";

const registry = defaultRegistry().registerNode("upper", {
  configSchema: z.object({ from: z.string() }),
  async run(ctx, upstream) {
    return { success: true, output: (upstream[ctx.config.from]?.output ?? "").toUpperCase() };
  },
});

Get started

Concepts

Building blocks

Guides

Reference

API reference

input

prompt

llm

decision

http

mcp

output

Add your own node

​input

​prompt

​llm

​decision

​http

​mcp

​output

​Add your own node

input

prompt

llm

decision

http

mcp

output

Add your own node